Acoustic deep learning for defect detection in aluminium wheel rims Authors: Arkadiusz Rychlik Published: 2026 ABSTRACT Reliable detection of structural defects in aluminium wheel rims remains challenging in routine workshop diagnostics, where inspection is primarily based on visual and organoleptic assessment supported by operator experience. This study proposes RimNet, a deep-learning-based diagnostic framework for automated condition assessment of aluminium wheel rims using their acoustic response to a single impulse excitation. The method combines short-time Fourier transform representations with a convolutional neural network and incorporates domain filtering and hard-negative mining to improve the separability of defect-related conditions. The dataset comprised 930 labelled examples representing a broad spectrum of real-world rim states, including geometric deviations, material defects, and serviceable conditions. RimNet achieved strong classification performance, reaching 96.9% accuracy at the file level, while no wheel-level misclassifications were observed under the adopted evaluation protocol. Analysis of uncertainty measures—including prediction margins, entropy, Brier score, and energy-based indicators—revealed clear relationships between prediction stability and defect type. Misclassified cases were associated with reduced classification margins and increased uncertainty, particularly in samples influenced by geometric variability. The RimNet Predictor application enables real-time diagnostics and provides uncertainty estimates for individual rims. Because the proposed approach relies solely on microphone-based acquisition and does not require modifications to wheel balancing equipment, it can be readily integrated into workshop and service environments. These results demonstrate the potential of acoustic deep-learning-based diagnostics for structural health monitoring of lightweight mechanical components. Introduction Industrial motivation and problem background Wheel rims are critical structural components affecting vehicle safety and performance. Their condition must be regularly assessed to ensure reliable operation. To guarantee safe operation, the rim must satisfy strict geometric and strength requirements. Two sides of the rim are distinguished: the inner (IN) and outer (OUT). Correspondingly, axial runout (A), describing lateral deviations of the rim bead, and radial runout (R), describing deviations in the radial direction, are commonly used to quantify geometric accuracy. Assessment of wheel rim condition is a mandatory part of vehicle inspections in Europe and worldwide. However, legal regulations specify only general safety requirements and do not define explicit numerical limits for radial or axial runout. In Poland, the legal basis is provided by national regulations1, while in Germany the applicable framework is defined by §29 StVZO and Annex VIII2. These regulations require the absence of cracks, severe deformation, significant corrosion, and improper mounting, but they do not specify quantitative tolerance values. ECE regulation 1243 defines homologation requirements for passenger-car wheels and rims, focusing primarily on durability and fatigue resistance rather than geometric runout evaluation. More detailed geometric recommendations are provided by the European Tyre and Rim Technical Organisation (ETRTO)4, which defines compatibility and mounting dimensions for rims and tires. Based on these guidelines, typical runout tolerances of approximately 1.0–1.5 mm can be inferred for passenger-car rims. Manufacturer documentation provides similar limits, typically within the range of 1.0–1.5 mm for radial and lateral runout5,6, with some light-alloy rims requiring stricter tolerances of about 0.8 mm5,7. In contrast, steel rims may tolerate deviations of up to approximately 2.0 mm8. In practical diagnostics, rim evaluation remains largely qualitative because the rim is usually assessed together with the mounted tire. Inspectors rely on visual inspection, observed wheel runout during rotation, and the overall condition of the tire and suspension components. As a result, only severe deformation cases can typically be detected during standard vehicle inspection procedures. Consequently, current diagnostic practice suffers from limited sensitivity, subjectivity, and strong dependence on examiner experience. These limitations indicate a clear technological gap in rim condition assessment. Closing this gap requires diagnostic approaches that provide objective, quantitative, and repeatable evaluation of rim structural integrity. Acoustic and vibration-based methods offer a promising solution. Impulse excitation of a rim generates characteristic acoustic signals that reflect both its geometric and material condition. Such signals can be acquired in a non-destructive and repeatable manner and subsequently used as input data for machine learning models. This enables the development of automated diagnostic systems capable of detecting subtle structural defects. In the present work, these signals form the basis of a multi-stage diagnostic system employing deep neural networks for reliable classification of rim condition. Section "Methods" introduces the experimental dataset and presents the RimNet diagnostic pipeline used for automated analysis of acoustic impulse responses. State of the art in wheel rim inspection Inspection of wheel rim condition is an important element of vehicle diagnostics, as rim defects may significantly affect operational safety and durability of components subjected to dynamic loads. Existing research in this field includes modal and vibration-based diagnostics, acoustic approaches, imaging-based inspection techniques, and artificial intelligence–supported methods. Modal and vibration-based methods Vibration-based diagnostics of wheel rims has been investigated in several studies. Rychlik9 proposed a non-invasive method based on analysing natural-frequency spectra obtained during forced vibration excitation. Variations in modal parameters such as frequencies, amplitudes, and damping were linked to structural deformation, stiffness loss, cracking, and geometric deviations including radial and lateral runout. Further developments were presented by Borecki et al.10, who combined vibration-spectrum analysis with artificial neural networks and integrated the diagnostic module with a commercial wheel balancer. This system enabled automatic classification of rims into operational categories such as “new”, “serviceable”, and “unserviceable”. Subsequent work by Borecki, Rychlik, and co-authors8 incorporated damping-factor evaluation and dispersion metrics derived from experimental modal analysis, demonstrating high sensitivity to fatigue-related degradation of steel rims. In addition, Rychlik et al.11 applied self-organising maps (SOM) to analyse vibration-derived FFT features, enabling detection of subtle structural changes that may remain undetected in classical modal analysis. Acoustic methods Acoustic methods constitute an alternative diagnostic approach, particularly for lightweight aluminium rims. Rychlik12 demonstrated that acoustic parameters such as reverberation time, sound absorption coefficient, and acoustic energy can be used as diagnostic indicators of rim condition. Experimental data collected during fatigue testing were used to train a deep neural network capable of classifying rims originating from real service conditions. Imaging-based methods: X-ray inspection and image analysis Imaging-based inspection methods are mainly applied during the manufacturing stage of cast aluminium rims. Wu and Huang13 developed a system combining X-ray imaging with you only look once (YOLO)-based object detection to identify casting defects such as porosity, inclusions, and material discontinuities. Such approaches enable highly accurate detection of internal defects but require specialised equipment. Design-supporting methods: optimisation and modal analysis Research on rim dynamics has also been supported by computational modelling and optimisation. Topaloğlu et al.14 proposed surrogate models and neural networks for predicting structural parameters of rims, including mass, modal frequencies, and displacement fields, enabling optimisation of rim design with respect to weight and dynamic performance. Table 1 summarises the main categories of wheel rim inspection methods and highlights their diagnostic capabilities and practical limitations. Traditional geometric measurements and visual inspection allow only detection of surface defects and geometric deviations. Modal and vibration-based approaches offer higher sensitivity to structural changes, while acoustic methods provide relatively simple and rapid measurements. Imaging-based techniques enable detection of internal defects, whereas artificial intelligence–based approaches allow integration of multimodal diagnostic information and automated classification of rim condition.Table 1Comparison of contemporary wheel rim inspection methods.Method categoryRange of detectable damage/parametersAdvantagesLimitationsTypical applicationExample publicationsGeometric measurements and visual inspectionGeometric deformation; impact damage; visible cracks; corrosion; rim profile deviationVery fast and inexpensive; widely available in workshops; no specialised equipmentCannot detect internal or fatigue-related defects; low sensitivity to micro-cracks; subjective assessmentService diagnostics; preliminary inspection1,2,4Modal and vibration-based methodsStiffness loss; plastic deformation; weld defects; fatigue damage; shifts in natural frequencies and dampingHigh sensitivity to structural changes; can be automated; compatible with balancing systemsRequire controlled excitation and sensors; sensitivity to boundary conditionsOperational diagnostics; fatigue testing9,10,15Acoustic methodsFatigue-related degradation; structural integrity changes; resonance-related parameters; spectral pattern variationsSimple measurement chain; non-contact sensing; high sensitivity to fatigue effectsSensitive to environmental noise; limited defect localisation; dependence on excitation conditionsLaboratory testing; rapid classification; condition monitoring12,16Imaging-based methods (X-ray + image analysis)Internal casting defects: pores, inclusions, cracksDirect insight into internal structure; high detection accuracyHigh equipment cost; radiological safety constraints; limited applicability in service conditionsProduction quality control13Design-supporting methods (ML + FEA)Rim mass; modal frequencies; displacement fields; structural optimisation parametersSupports structural optimisation and design; predictive capabilitiesNot directly applicable to fault diagnosis; requires detailed modelsRim design and optimisation14,17–19Neural networks / AI in rim diagnosticsClassification based on modal, acoustic and image-derived features; detection of subtle and early-stage defectsHigh classification accuracy; ability to detect subtle defects; automation potential; scalable to real-world dataRequire representative training datasets; limited interpretability; risk of dataset biasAutomatic rim classification; decision support systems10–12,20,21 Table 1Comparison of contemporary wheel rim inspection methods.Method categoryRange of detectable damage/parametersAdvantagesLimitationsTypical applicationExample publicationsGeometric measurements and visual inspectionGeometric deformation; impact damage; visible cracks; corrosion; rim profile deviationVery fast and inexpensive; widely available in workshops; no specialised equipmentCannot detect internal or fatigue-related defects; low sensitivity to micro-cracks; subjective assessmentService diagnostics; preliminary inspection1,2,4Modal and vibration-based methodsStiffness loss; plastic deformation; weld defects; fatigue damage; shifts in natural frequencies and dampingHigh sensitivity to structural changes; can be automated; compatible with balancing systemsRequire controlled excitation and sensors; sensitivity to boundary conditionsOperational diagnostics; fatigue testing9,10,15Acoustic methodsFatigue-related degradation; structural integrity changes; resonance-related parameters; spectral pattern variationsSimple measurement chain; non-contact sensing; high sensitivity to fatigue effectsSensitive to environmental noise; limited defect localisation; dependence on excitation conditionsLaboratory testing; rapid classification; condition monitoring12,16Imaging-based methods (X-ray + image analysis)Internal casting defects: pores, inclusions, cracksDirect insight into internal structure; high detection accuracyHigh equipment cost; radiological safety constraints; limited applicability in service conditionsProduction quality control13Design-supporting methods (ML + FEA)Rim mass; modal frequencies; displacement fields; structural optimisation parametersSupports structural optimisation and design; predictive capabilitiesNot directly applicable to fault diagnosis; requires detailed modelsRim design and optimisation14,17–19Neural networks / AI in rim diagnosticsClassification based on modal, acoustic and image-derived features; detection of subtle and early-stage defectsHigh classification accuracy; ability to detect subtle defects; automation potential; scalable to real-world dataRequire representative training datasets; limited interpretability; risk of dataset biasAutomatic rim classification; decision support systems10–12,20,21 Comparison of contemporary wheel rim inspection methods. In contrast to traditional inspection methods such as visual assessment and geometric measurements, which rely on observable features and operator experience, the proposed RimNet framework provides an objective and data-driven approach to rim condition evaluation. Compared to conventional vibration-based and acoustic methods, RimNet enables automated feature extraction and classification, reducing the dependence on expert interpretation and handcrafted signal features. Furthermore, unlike imaging-based techniques such as X-ray inspection, which require specialised equipment and controlled environments, the proposed method operates using a simple measurement setup based on impulse excitation and a microphone, making it suitable for practical workshop conditions. The proposed RimNet framework extends existing acoustic and AI-based approaches by integrating domain-aware learning, uncertainty analysis, and multi-level decision fusion. This combination of simplicity, sensitivity, and automation positions RimNet as a practical solution for real-world diagnostic applications. Overall, existing studies demonstrate substantial progress in wheel rim diagnostics; however, no universal solution currently enables fast, non-invasive, and fully automated assessment of wheel rim condition under typical workshop conditions. This limitation motivates the development of advanced diagnostic systems such as RimNet. Deep learning in vibration and acoustic signal analysis Deep learning methods have recently become widely applied in technical diagnostics and condition monitoring of mechanical systems. The increasing availability of sensing technologies, including accelerometers and microphones, combined with growing computational power, has enabled the use of neural networks for automatic extraction of diagnostically relevant features from vibration and acoustic signals. These signals constitute natural responses of mechanical structures to damage, material degradation, and fatigue processes. Deep learning in vibration-based diagnostics Vibration signals have long been used in the diagnostics of rotating machinery and thin-walled structures. Classical approaches relied on manually engineered features such as Fast Fourier Transform (FFT) spectra, envelopes, cepstrum coefficients, or statistical descriptors. In contrast, deep learning models enable end-to-end analysis directly from raw signals or time–frequency representations. Contemporary studies indicate that convolutional neural networks (CNN), long short-term memory (LSTM) and gated recurrent unit (GRU) networks, and hybrid architectures dominate modern vibration diagnostics and demonstrate high effectiveness in detecting weak faults in noisy environments22. Another important research direction is transfer learning and domain adaptation, which allow effective model training when only limited or imbalanced datasets are available. For example, Zhao et al. 23 demonstrated that transfer learning significantly improves diagnostic performance in small-data scenarios typical of industrial applications. Autoencoder-based approaches have also been proposed for anomaly detection without requiring labelled fault data. Pawlik et al. 24 introduced the AE-FIT autoencoder-based method, in which an autoencoder trained on healthy data generated interpretable spectra enabling the identification of typical gearbox faults. Deep learning in acoustic signal analysis Acoustic sensing represents an attractive alternative when the use of contact sensors is difficult or impractical. Microphones enable simple and non-invasive measurements while still capturing diagnostically relevant information about structural condition. Safaei et al. 25 applied the continuous wavelet transform to convert acoustic signals into time–frequency representations classified using CNNs, achieving high diagnostic accuracy. Similar results were reported by Yurdakul and Taşdemir 26, who used DenseNet applied to mel-spectrograms and demonstrated high robustness to noise. In addition to time–frequency representations, cepstral features derived from speech processing have also been successfully applied in structural health monitoring. In particular, mel-frequency cepstral coefficients (MFCCs) have been shown to effectively capture damage-induced changes in structural responses 16. Furthermore, enhanced cepstral analysis techniques based on nonlinearly spaced filterbanks have been investigated to improve sensitivity to structural dynamic behaviour and damage detection 27. Data augmentation and generative models are increasingly used to improve model generalisation under varying acoustic conditions. For example, Cycle-GAN-based approaches have been proposed to enhance training datasets in acoustic diagnostics28. In addition, acoustic sensing has been applied in mobile and vehicle-based diagnostics, where Jagatheesaperumal et al. 29 demonstrated classification of road surface types based on wheel-area acoustic signals. Acoustic emission techniques are also widely used for detecting micro-cracks and structural damage30–33. In contrast to classical AE-based approaches, the present work uses broadband acoustic signatures generated by impulsive excitation, which enables diagnostics using only a microphone. Decision fusion and reliability of deep learning diagnostics In many industrial diagnostic applications, multiple measurements are collected from a single object at different locations or time instants. Such scenarios naturally lead to modelling within the multi-instance learning (MIL) framework. Tseng et al. 34 demonstrated that MIL improves classification performance in acoustic event analysis when combined with late fusion of CNN-extracted features. Decision fusion techniques further improve robustness of diagnostic systems. Lin et al. 35 applied evidence-based fusion methods to ultrasonic diagnostics, demonstrating improved detection of subtle structural damage. Probability-based fusion, in which object-level decisions are derived from aggregated instance-level predictions, is particularly effective in handling signal variability and measurement noise. Another important aspect of practical diagnostic systems is classifier calibration. Deep learning models often produce overconfident predictions, which may reduce reliability in safety-critical applications. Calibration techniques such as temperature scaling and expected calibration error analysis improve agreement between predicted probabilities and the true class distribution36. Improving model robustness and reliability under varying operating conditions therefore remains a key research direction in diagnostic signal analysis. Overall, the literature indicates that deep learning has become a leading tool in vibration and acoustic diagnostics. Major research trends include the transition from manual feature engineering to end-to-end learning, the development of transfer learning and domain adaptation techniques, the use of multimodal data fusion, and increasing attention to reliability and robustness of diagnostic models under real operating conditions. Novelty and contributions of this work The main contributions of this work are: A complete acoustic diagnostic pipeline (RimNet) for aluminium wheel rim inspection based on impulse responses.A dataset of 930 labelled rim recordings representing real workshop conditions.A deep convolutional classifier operating on STFT time–frequency representations of acoustic signals.Integration of domain filtering and hard-negative mining mechanisms that improve classification robustness.An uncertainty analysis framework including entropy, classification margins, and Brier score.Demonstration of practical applicability through the RimNet Predictor tool enabling real-time wheel rim diagnostics. A complete acoustic diagnostic pipeline (RimNet) for aluminium wheel rim inspection based on impulse responses. A complete acoustic diagnostic pipeline (RimNet) for aluminium wheel rim inspection based on impulse responses. A dataset of 930 labelled rim recordings representing real workshop conditions. A dataset of 930 labelled rim recordings representing real workshop conditions. A deep convolutional classifier operating on STFT time–frequency representations of acoustic signals. A deep convolutional classifier operating on STFT time–frequency representations of acoustic signals. Integration of domain filtering and hard-negative mining mechanisms that improve classification robustness. Integration of domain filtering and hard-negative mining mechanisms that improve classification robustness. An uncertainty analysis framework including entropy, classification margins, and Brier score. An uncertainty analysis framework including entropy, classification margins, and Brier score. Demonstration of practical applicability through the RimNet Predictor tool enabling real-time wheel rim diagnostics. Demonstration of practical applicability through the RimNet Predictor tool enabling real-time wheel rim diagnostics. Methods Materials The objects of this study were aluminium automotive wheel rims collected during seasonal tyre replacement at a specialised tyre service station. Data acquisition was conducted over seven consecutive tyre-replacement seasons (summer and winter) between 2022 and 2025. This approach enabled the collection of a diverse population of rims with respect to age, operational mileage, and degree of material wear. In total, 930 acoustic recordings of individual wheel rims were obtained. The rims originated from multiple manufacturers and covered a size range of 13–19 inches. The general characteristics of the investigated rim population are summarised in Table 2.Table 2General characteristics of the investigated wheel rims.ParameterIn-service (used) wheel rimsNumber of samples930Rim size range13–19 inchesRim ageNew to 18 yearsOperational mileage0–255,000 km Table 2General characteristics of the investigated wheel rims.ParameterIn-service (used) wheel rimsNumber of samples930Rim size range13–19 inchesRim ageNew to 18 yearsOperational mileage0–255,000 km General characteristics of the investigated wheel rims. Because all measurements were conducted at the same service location, some rims were recorded more than once during subsequent tyre-replacement seasons. This occurred when customers returned to the same workshop for routine service or tyre storage. Such repeated measurements reflect realistic service conditions and allow the stability of the proposed diagnostic method to be assessed under quasi-operational scenarios. Acoustic diagnostic features (ADF) method Acoustic signal acquisition was performed using a proprietary experimental setup referred to as ADF (acoustic diagnostic features). The ADF method enables simultaneous identification of acoustic features and geometric parameters of wheel rims, allowing a comprehensive assessment of their technical condition. The same setup was used for all measurements in this study. A general view and a schematic diagram of the experimental stand are shown in Fig. 1.Fig. 1General view and schematic diagram of the ADF diagnostic setup for identifying acoustic and geometric features of a wheel rim: (1) tested wheel rim, (2) shaft with a self-centering chuck, (3) RG-50 microphone, (4) vibration exciter, (5) sensor head for axial and radial runout measurement. General view and schematic diagram of the ADF diagnostic setup for identifying acoustic and geometric features of a wheel rim: (1) tested wheel rim, (2) shaft with a self-centering chuck, (3) RG-50 microphone, (4) vibration exciter, (5) sensor head for axial and radial runout measurement. The tested rim (1) was mounted on a vertical shaft equipped with a self-centering chuck (2), ensuring axial alignment with the measurement system and stable gravitational clamping against the hub surface. This configuration minimized positioning errors and ensured repeatable measurement conditions. Excitation was introduced using an impulsive gravitational exciter (4), which impacted the lower rim flange in a tangential direction relative to the wheel-disc plane. The direction of excitation remained perpendicular to the microphone axis, enabling acquisition of the axial acoustic response of the structure. During each measurement the rim remained stationary. The acoustic response was recorded using a free-field ROGA RG-50 microphone (3) positioned perpendicularly to the mounting shaft axis. The microphone captured sound-pressure waveforms corresponding to the resonant response of the rim. After acoustic acquisition, bead runout was measured using a non-contact laser displacement sensor (Keyence LV-7080) (5). Measurements were performed in two planes—axial (A) and radial (R)—for both the inner (IN) and outer (OUT) rim sides. Maximum runout values were determined relative to the centring bore with an accuracy of ± 0.01 mm. The ROGA RG-50 microphone provides a flat amplitude response within 20 Hz–4 kHz (± 1.0 dB) and 4 kHz–20 kHz (± 1.5 dB). Signals were sampled at 8 kHz, providing full coverage of the frequency range up to 4 kHz. Since rim resonance frequencies did not exceed 500 Hz, this sampling rate significantly exceeded the requirements of the Shannon criterion. The measurement chain was interfaced with an NI USB-6210 data acquisition card (National Instruments). Prior to each measurement series, the system was calibrated using an acoustic calibrator generating 94 dB SPL at 1 kHz. The exciter delivered an impact energy of 1 ± 0.1 J, determined by the hammer mass \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$m_m = 0.425$$\end{document} kg and free-fall height \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H_m = 0.240$$\end{document} m. Signal acquisition, processing, visualisation, and archiving were implemented in MATLAB. Dedicated scripts automatically detected the acoustic response and stored the waveform in .wav format together with metadata encoded in the filename (see Section "Acoustic diagnostic features (ADF) method"). ADF measurement procedure Acoustic and geometric measurements were performed according to a predefined experimental protocol. Typically, four rims originating from a single vehicle were examined during seasonal tyre replacement. First, basic rim information was recorded, including vehicle make and model, rim production year, operational mileage, and wheel position (FL/FR/RL/RR). This information was stored as metadata associated with each measurement file. Each rim then underwent visual inspection to detect cracks, repair traces, welding marks, material loss, or visible deformation. Rims without visible defects were labelled VC (visual compliance), whereas rims exhibiting surface defects or cracks were labelled VD (visual defect). Next, the rim was mounted in the ADF self-centering chuck and geometric runout measurements were performed. Axial (A) and radial (R) runout were measured for both the inner (IN) and outer (OUT) sides of the rim bead relative to the centring bore. Rims were classified as GC (geometric compliance) for runout values <1.25 mm and GD (geometric defect) for values ≥ 1.25 mm. The runout values were stored in the metadata in the order A_IN, R_IN, A_OUT, R_OUT. Acoustic measurements were then performed by impulsive excitation of the rim and recording of the resulting acoustic response using the RG-50 microphone. Each waveform was saved in .wav format together with metadata encoded in the filename using the following structure: vehicle_brand_model_year_mileage_VC/VD_GC/GD_AIN_RIN_AOUT_ROUT_FL/FR/RL/RR. For statistical analysis, each rim was measured five times. After each measurement the rim was removed from the chuck, remounted, and re-excited. This procedure enabled evaluation of measurement repeatability and mounting-related variability. For routine diagnostic measurements the protocol was simplified: the acoustic response was recorded twice without removing the rim from the chuck, reflecting real workshop conditions during seasonal tyre replacement. During acquisition, a small number of rims exhibited clearly atypical acoustic responses. Such cases occurred in 10 out of 930 measurements (approximately 1.0%). These recordings were labelled as X (Ambiguity String) and stored in the dataset but excluded from neural-network training due to the lack of unambiguous material or geometric interpretation. A detailed analysis of these cases is presented in Section "Representative acoustic signatures". Signal preprocessing The acoustic signals acquired using the ADF setup were processed using a standardized preprocessing pipeline designed to ensure comparability between measurements and to isolate the most representative resonant responses. The algorithm was implemented in MATLAB and applied uniformly to all recordings. The general structure of the preprocessing procedure is illustrated in Fig. 2.Fig. 2Block diagram of the acoustic preprocessing procedure applied to wheel-rim signals prior to RimNet classification. Block diagram of the acoustic preprocessing procedure applied to wheel-rim signals prior to RimNet classification. After loading each signal, detrending was performed by subtracting the mean value, followed by normalization with respect to the maximum amplitude. This step removed the DC component and ensured a consistent amplitude scale for further analysis. To identify individual resonant responses, the signal envelope was computed using the Hilbert transform and smoothed with a 60 ms moving window. Local maxima were detected using adaptive thresholds and a minimum inter-peak distance criterion. For each maximum, the corresponding high-energy segment containing a complete resonant response was extracted, producing a set of non-overlapping signal fragments. Typically, four consecutive impact responses were recorded during a single measurement session. A segmentation procedure selected up to four of the longest and most stable fragments, which were treated as technically independent signal segments. During RimNet training, only one representative segment per rim—the longest and least distorted fragment—was used as the input instance. The remaining segments were used exclusively for auxiliary analyses, including repeatability assessment, spectral comparison, and data-quality verification. All selected segments were cropped or zero-padded to a fixed duration of 2.5 s and resampled to 8192 Hz, ensuring a uniform input representation for the neural network while preserving the dominant acoustic modes of the rim. The processed segments were stored as .wav files with metadata encoded in the filename. In the final RimNet v3s-X dataset, a single acoustic segment per rim was used for training and evaluation, whereas additional repetitions served only for statistical analysis of measurement repeatability. Proposed method: RimNet diagnostic pipeline General concept of the diagnostic system The RimNet diagnostic system performs a two-stage assessment of wheel-rim condition using impact-excited acoustic signals. At the level of a single recording (diagnostic instance), a convolutional neural network (CNN) estimates the probability of a material defect. These predictions are subsequently aggregated at the rim level using weighted probabilistic fusion (soft fusion), producing a stable final decision regarding the material condition of the rim. The overall structure of the proposed diagnostic system is illustrated in Fig. 3.Fig. 3Conceptual overview of the RimNet diagnostic pipeline. Acoustic signals generated by impulse excitation are transformed into time–frequency representations and analysed using a convolutional neural network (CNN) to estimate file-level defect probability. These predictions are aggregated at the rim level using probabilistic fusion, and combined with geometric assessment to produce the final diagnostic decision. Conceptual overview of the RimNet diagnostic pipeline. Acoustic signals generated by impulse excitation are transformed into time–frequency representations and analysed using a convolutional neural network (CNN) to estimate file-level defect probability. These predictions are aggregated at the rim level using probabilistic fusion, and combined with geometric assessment to produce the final diagnostic decision. The acoustic diagnosis is combined with an independently determined geometric assessment (GC/GD), resulting in a three-state output: serviceable rim (OK), rim with a material defect (BAD), or rim with a geometric defect (GD). To improve robustness, two additional mechanisms were introduced during training: (i) domain filtering, which excludes geometrically non-compliant recordings (class \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {GD}$$\end{document}) and acoustically ambiguous samples (class X), and (ii) hard-negative mining, which selectively incorporates spectrally atypical but geometrically compliant rims into the negative class. From a diagnostic perspective, it is important to note that the CNN model is trained primarily to detect material-related defects (VD), which produce strong and localized changes in the acoustic response. In contrast, geometric defects (GD), such as runout or deformation, affect the global shape of the rim and typically result in more subtle and distributed spectral variations. As a consequence, GD cases are excluded from supervised training and can be interpreted as out-of-distribution or borderline samples with respect to the learned decision boundary. In the operational setting, however, all non-compliant cases, including GD, are treated as defective, and their presence is resolved through the integration of geometric assessment and uncertainty analysis. This distinction resolves the apparent discrepancy between the training objective and the broader diagnostic interpretation used in practice. Input representation of the acoustic signal Each diagnostic instance corresponds to a single acoustic response obtained after impact excitation and preprocessing (signal duration 2.5 s, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_{\textrm{s}} = 8192$$\end{document} Hz). Time-domain signals are converted into time–frequency representations in the form of log-mel spectrograms used as CNN inputs. Spectrograms were generated using a Hanning window of 1024 samples with 75% overlap (hop size 256), a 1024-point FFT, and 64 mel frequency bands covering the informative frequency range of the rim response. The resulting power spectra were transformed to the logarithmic scale and globally normalised by mean subtraction and standard deviation scaling. Each diagnostic instance is represented as a single-channel image (frequency × time × 1) and passed directly to the network. RimNet network architecture The core of the proposed method is a two-dimensional convolutional neural network designed for log-mel spectrogram analysis (Fig. 4).Fig. 4Schematic architecture of the RimNet network. Schematic architecture of the RimNet network. The network consists of three convolution–normalisation blocks followed by max-pooling layers, enabling extraction of progressively more complex time–frequency features while maintaining a compact number of parameters. This design reduces overfitting and improves generalisation across rims of different sizes and manufacturers. The final fully connected layer, regularised with dropout, outputs the confidence estimate of a material defect. Training procedure and hard-negative mining Domain filtering Before training the classifier, recordings dominated by geometric deviations or acoustically ambiguous responses were removed. Rims belonging to the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {GD}$$\end{document} class and X-class recordings exhibit spectral variability primarily related to geometric distortion rather than material integrity. Removing these samples reduces intra-class variance and improves separation between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {OK}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {BAD}$$\end{document} rims. Hard-negative mining Within the compliant subset, some rims exhibit atypical acoustic responses that partially resemble defective cases. These hard negatives are identified during training as samples either misclassified as positive or assigned high defect probability. They are then reintroduced into the negative class to sharpen the decision boundary and reduce false-positive detections. Training configuration The RimNet model was trained in MATLAB using the trainNetwork function with the Adam optimiser. The main hyperparameters were:MiniBatchSize = 64MaxEpochs = 75InitialLearnRate = \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1\times 10^{-3}$$\end{document}Shuffle = every epochValidationData = dedicated validation setValidationFrequency ≈ 1 epochValidationPatience = 5 (early stopping)Binary classification was trained using categorical cross-entropy loss. Class weights were applied in the classificationLayer to compensate for class imbalance between defect and non-defect rims. MiniBatchSize = 64 MiniBatchSize = 64 MaxEpochs = 75 MaxEpochs = 75 InitialLearnRate = \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1\times 10^{-3}$$\end{document} InitialLearnRate = \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1\times 10^{-3}$$\end{document} Shuffle = every epoch Shuffle = every epoch ValidationData = dedicated validation set ValidationData = dedicated validation set ValidationFrequency ≈ 1 epoch ValidationFrequency ≈ 1 epoch ValidationPatience = 5 (early stopping) ValidationPatience = 5 (early stopping) Integration of DF and HNM Domain filtering was applied during dataset preparation by excluding all \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {GD}$$\end{document} and X-class samples. After initial training, hard-negative mining identified negative samples with high predicted defect probability and reintroduced them into the training set. Increasing their contribution in subsequent training iterations improved robustness and reduced confusion between compliant rims and true material defects. Probability fusion and decision thresholds The RimNet framework operates at two decision levels: the file level and the wheel level. For each acoustic recording, the convolutional neural network produces the probability of a material defect, denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\textrm{file}} = p(\text {BAD})$$\end{document}. In this study, Softmax outputs are interpreted as confidence estimates rather than calibrated statistical probabilities, since no explicit probability calibration procedure was applied. When multiple recordings of the same rim are available, the individual predictions are aggregated using probabilistic fusion. The final rim-level probability is computed as the arithmetic mean of file-level probabilities:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\textrm{wheel}} = \frac{1}{N}\sum _{i=1}^{N} p_{\textrm{file},i},$$\end{document}where N denotes the number of acoustic recordings associated with the same rim. The final classification is obtained by comparing the aggregated probability with a calibrated decision threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau _{\textrm{wheel}}$$\end{document}. If \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\textrm{wheel}} \ge \tau _{\textrm{wheel}}$$\end{document}, the rim is classified as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {BAD}$$\end{document}; otherwise it is classified as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {OK}$$\end{document}. The threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau _{\textrm{wheel}}$$\end{document} was selected on the validation set to maximise the F1 score of the defect class. This calibration step improves robustness to variability of individual impact recordings and stabilises the final rim-level decision. In the final RimNet v3s-X configuration used in this study, a single representative acoustic recording per rim was used for classification. Nevertheless, the proposed probabilistic fusion framework allows straightforward extension to multi-instance diagnostic scenarios involving multiple impacts per rim. Model export and deployment pipeline After calibration of the decision thresholds, the trained RimNet model is exported to a binary file (RimNet.mat) using MATLAB serialization. This separates the computationally intensive training stage from lightweight inference during diagnostics. Operational use requires only:loading the trained model,computing the log-mel spectrogram of a newly recorded acoustic signal,estimating the file-level defect probability via a forward pass through the CNN,computing the wheel-level probability,assigning the final three-state decision.Inference for a single recording requires only a few milliseconds on a standard central processing unit (CPU), enabling real-time operation in workshop environments. loading the trained model, loading the trained model, computing the log-mel spectrogram of a newly recorded acoustic signal, computing the log-mel spectrogram of a newly recorded acoustic signal, estimating the file-level defect probability via a forward pass through the CNN, estimating the file-level defect probability via a forward pass through the CNN, computing the wheel-level probability, computing the wheel-level probability, assigning the final three-state decision. assigning the final three-state decision. For practical deployment, a standalone inference tool termed RimNet Predictor was developed. The module loads the exported model, processes new acoustic recordings using the same preprocessing pipeline as during training, and outputs the predicted rim condition together with associated probabilities. The system also computes additional uncertainty measures (logit, margin, entropy, Brier score, and normalised uncertainty)37–42, which support interpretation of predictions and identification of uncertain diagnostic cases. Results and discussion Evaluation protocol The performance of the proposed RimNet framework was evaluated at multiple levels to reflect both the training setup and the intended operational use. At the file level, the classifier was evaluated on a held-out test set consisting of individual acoustic recordings. Each recording corresponds to a single impact response acquired from a given rim. The dataset was split strictly at the rim level using unique WheelID identifiers to prevent information leakage between training, validation, and test sets. In addition to file-level evaluation, aggregated predictions were analysed at the wheel level. In this setting, multiple recordings associated with the same rim are combined using probabilistic fusion to obtain a more stable and reliable diagnostic decision. This reflects the practical scenario in which multiple measurements may be available for a single wheel during inspection. Furthermore, a full-dataset inference mode was considered to assess the behaviour of the trained model across the entire dataset, including samples excluded during supervised training (e.g., GD and X cases). This analysis provides insight into the generalisation capability of the model beyond the strictly defined training domain. These evaluation modes are complementary and are reported separately throughout the manuscript to ensure clarity and consistency of interpretation. These distinctions are essential, as the reported performance metrics correspond to different evaluation settings and are not directly comparable. Dataset characteristics The dataset used for training and validation of the RimNet system consists of 930 aluminium alloy wheel rims collected during seasonal tyre replacement campaigns conducted between 2022 and 2025. The analysed rims represent a wide range of operational conditions, including rim sizes from 13 to 19 inches, rim age ranging from new to 18 years, and operational mileage up to 255,000 km. Such diversity enables reliable evaluation of the diagnostic system under realistic workshop conditions. The dataset reflects realistic workshop conditions rather than controlled laboratory experiments. The distribution of diagnostic categories is presented in Table 3. The classification is based on visual inspection (VC/VD), geometric assessment (GC/GD), and identification of acoustically ambiguous cases (X). The dataset structure reflects the natural composition of rims encountered in service practice, where geometrically and visually compliant rims dominate but various defect scenarios are also present.Table 3Distribution of diagnostic categories in the rim dataset.Diagnostic categoryNumber of rimsVD (visual defect)366GD (geometric defect)157VD + GD (combined defects)84VC + GC (fully compliant rims)397X (acoustically ambiguous cases)10Total930 Table 3Distribution of diagnostic categories in the rim dataset.Diagnostic categoryNumber of rimsVD (visual defect)366GD (geometric defect)157VD + GD (combined defects)84VC + GC (fully compliant rims)397X (acoustically ambiguous cases)10Total930 Distribution of diagnostic categories in the rim dataset. Approximately 42% of the rims exhibit visual defects (VD), 17% show significant geometric deformation (GD), and 9% contain both visual and geometric defects. Fully compliant rims (VC and GC) constitute approximately 43% of the dataset, while acoustically ambiguous cases (X) account for about 1% and were excluded from the model training process. The presence of both serviceable rims and multiple defect scenarios provides a realistic diagnostic dataset and enables reliable validation of the RimNet system under practical workshop conditions. Data structure and terminology To ensure clarity, the following terminology is used throughout the manuscript:Recording—a single acoustic signal acquired from an impulse excitation applied to a rim.Segment—a selected time fragment of a recording used as input to the classifier after preprocessing.Rim—a physical wheel rim corresponding to a unique WheelID in the dataset.Wheel-level prediction—an aggregated diagnostic decision obtained by combining multiple recordings associated with the same rim.The dataset contains multiple recordings for selected rims, reflecting repeated measurements under varying conditions. However, for supervised training and evaluation, a single representative segment per rim was used to ensure consistency and to avoid over-representation of specific rims. Recording—a single acoustic signal acquired from an impulse excitation applied to a rim. Recording—a single acoustic signal acquired from an impulse excitation applied to a rim. Segment—a selected time fragment of a recording used as input to the classifier after preprocessing. Segment—a selected time fragment of a recording used as input to the classifier after preprocessing. Rim—a physical wheel rim corresponding to a unique WheelID in the dataset. Rim—a physical wheel rim corresponding to a unique WheelID in the dataset. Wheel-level prediction—an aggregated diagnostic decision obtained by combining multiple recordings associated with the same rim. Wheel-level prediction—an aggregated diagnostic decision obtained by combining multiple recordings associated with the same rim. The dataset split was performed strictly at the rim level using unique WheelID identifiers. This guarantees that no information from the same physical rim appears across training, validation, and test subsets, thereby preventing data leakage. This structured definition of data levels ensures a consistent interpretation of results across different evaluation settings. The dataset exhibits moderate class imbalance, which is addressed by reporting class-specific metrics, particularly for the BAD class. Representative acoustic signatures Representative examples of acoustic responses are shown in Fig. 5, illustrating characteristic differences between serviceable (OK) rims and rims with material defects (BAD).Fig. 5Example spectrograms of (a) a serviceable (OK) rim and (b) a rim with a material defect (BAD), corresponding to a typical passenger vehicle rim geometry (e.g., 7Jx17H2). The defective rim exhibits a more irregular and diffuse spectral energy distribution, with reduced stability of dominant frequency bands compared to the OK case. Example spectrograms of (a) a serviceable (OK) rim and (b) a rim with a material defect (BAD), corresponding to a typical passenger vehicle rim geometry (e.g., 7Jx17H2). The defective rim exhibits a more irregular and diffuse spectral energy distribution, with reduced stability of dominant frequency bands compared to the OK case. These differences are consistent with the expected impact of structural discontinuities on the modal response of the rim, resulting in a more complex and less structured spectral pattern for defective rims. This behaviour supports the discriminative capability of the proposed CNN-based diagnostic approach. Repeatability of the acoustic signal for a single rim To verify that acoustic signals obtained using the ADF (Acoustic Diagnostic Features) method are repeatable and not dominated by random variations caused by rim mounting or manual impulse excitation, a repeatability study was conducted for a reference rim. Four independent acoustic recordings (r1–r4) were acquired. After each measurement, the rim was removed from the fixture, remounted, and excited again using a new mechanical impulse. The resulting signals were analysed both qualitatively and quantitatively, including comparisons of signal energy, Hilbert envelopes, and frequency-domain characteristics. Basic signal quality metrics After time alignment (delay compensation and resampling to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_{\textrm{s}} = 8192$$\end{document} Hz), basic signal quality metrics were calculated. The analysed recordings exhibited inter-recording delays of 0–3 ms, RMS values in the range 0.1026–0.1498, and an energy coefficient of variation of approximately \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$CV \approx 16\%$$\end{document}. The peak amplitude remained constant (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {PEAK} = 0.999$$\end{document}), while spectral centroid locations were nearly identical (4095.8–4095.9 Hz). The observed RMS variability results primarily from differences in manual excitation energy, whereas the modal structure of the acoustic response remains stable. Similarity of time-domain signals and envelopes Similarity between recordings was evaluated using Pearson correlation coefficients. For the raw time-domain signals, correlations ranged from r = 0.80 to 0.91 (mean r = 0.855). For the Hilbert envelopes, correlations were significantly higher, ranging from r = 0.956 to 0.987 (mean r = 0.966). These results indicate a highly repeatable modal acoustic response of the rim and confirm that variations in rim positioning or excitation force do not significantly affect the underlying resonant structure. Frequency-domain spectral agreement Spectral similarity was further evaluated using cosine similarity, a common metric in time–frequency signal analysis. The cosine similarity matrix of the spectra for the analysed recordings is shown in Fig. 6.Fig. 6Cosine similarity matrix of the spectra for four independent acoustic recordings (r1–r4) of the same rim. Cosine similarity matrix of the spectra for four independent acoustic recordings (r1–r4) of the same rim. Cosine similarity values range from 0.946 to 0.967, with a mean value of 0.955. Such high similarity indicates that the dominant resonant frequencies and their relative amplitudes remain practically unchanged between measurements. In modal diagnostics, values above 0.90 are generally interpreted as strong evidence of spectral stability43. Conclusions from the repeatability analysis The results confirm that the acoustic response of a wheel rim remains highly repeatable even under quasi-workshop measurement conditions involving manual excitation and repeated mounting. Both envelope correlation and spectral similarity demonstrate the stability of the rim’s modal response, which constitutes a key prerequisite for the ADF measurement approach and the RimNet diagnostic pipeline. These findings justify the use of normalised time–frequency representations as network inputs and provide a methodological basis for further analysis of the full dataset consisting of 930 recordings. Analysis of X-type cases (ambiguity) Cases marked as X (ambiguity class) represent a specific group of wheel rims for which the acoustic response differs audibly from other rims originating from the same vehicle, despite the absence of clear visual or geometric indicators of damage. In the analysed dataset, 10 such cases were identified among 930 rims (approximately 1.0%). These rims were characterised by a reduced resonant tone, poorly defined modal response, and atypical decay of acoustic energy. To verify whether X-type rims constitute an acoustically distinct category, a comparative analysis was performed for one such rim and the remaining three rims from the same vehicle (X–FL, FR, RL, RR). The results were compared with a reference set of nominal rims obtained from another vehicle. The corresponding spectral similarity matrices are shown in Fig. 7.Fig. 7Cosine similarity matrices of the spectra for (left) a reference rim set and (right) an X-type rim. Cosine similarity matrices of the spectra for (left) a reference rim set and (right) an X-type rim. For the nominal rims, very high and homogeneous spectral similarity values are observed (cosine similarity ≈ 0.95), indicating stable modal behaviour across rims of the same vehicle. In contrast, the X-type rim shows significantly lower similarity relative to the remaining rims. Spectral and envelope similarity metrics Similarity analysis was performed using three complementary metrics: correlation of the time-domain signals, correlation of Hilbert envelopes, and cosine similarity of amplitude spectra. For nominal rims, spectral similarity typically exceeds 0.94–0.97 (Section "Frequency-domain spectral agreement"). In the analysed X-type case, considerably lower values were obtained: spectral similarity between non-X rims ranged from 0.726 to 0.735, while similarity between the X-type rim and the remaining rims was only 0.635–0.660. Similar discrepancies were observed for envelope correlation (0.594–0.852) and time-domain signal correlation (0.181–0.359), indicating a substantial difference in modal structure. Modal and diagnostic interpretation According to established modal-analysis criteria43, spectral similarity values below approximately 0.80 typically indicate significant structural changes such as stiffness loss, boundary-condition variations, material damage, or increased damping. The obtained results therefore confirm that X-type rims exhibit pronounced acoustic anomalies that distinguish them clearly from the remaining rims of the same vehicle. Justification for excluding the X class from RimNet training The presented analysis justifies excluding X-type samples from the RimNet training dataset. Their strongly deviating spectral characteristics could introduce noise into the feature space and prevent reliable assignment to either the “material defect” or “no defect” class. Furthermore, the lack of additional diagnostic evidence (e.g., X-ray inspection or Original Equipment Manufacturer (OEM) service history) makes it impossible to assign a reliable ground-truth label. Consequently, X-type rims are analysed only as case-study examples and are excluded from the supervised training process. This prevents blurring of the decision boundary between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {OK}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {BAD}$$\end{document} classes and improves stability of the trained classifier. Classification performance of RimNet File-level classification results The performance of the RimNet classifier was evaluated at the level of individual recordings using a test dataset consisting of 184 audio files, including 72 recordings of rims with material defects and 112 recordings without defects. The dataset was split at the rim level (by WheelID) to prevent information leakage between recordings originating from the same rim. The training procedure also incorporated domain filtering (exclusion of GD and X cases) and the hard-negative mining strategy. The training set initially contained 264 recordings and was extended with 45 hard-negative examples from the GC-only class, resulting in a final dataset of 309 training recordings, 93 validation recordings, and 184 test recordings (see Section "Classification performance of RimNet"). The confusion matrix for the argmax-based file-level classification is shown in Table 4.Table 4Confusion matrix of the RimNet classifier at the file level for the test dataset (184 recordings).Pred OKPred BADTrue OK8923True BAD072 Table 4Confusion matrix of the RimNet classifier at the file level for the test dataset (184 recordings).Pred OKPred BADTrue OK8923True BAD072 Confusion matrix of the RimNet classifier at the file level for the test dataset (184 recordings). A comparison with classical machine-learning baselines is presented in Table 5.Table 5Comparison of RimNet with classical machine-learning baselines on the test dataset (file-level classification).ModelAccuracyPrecision (BAD)Recall (BAD)F1 score (BAD)SVM (RBF kernel)85.3%72.7%100.0%84.2%Random forest78.8%65.1%98.6%78.5%RimNet (proposed)87.5%75.8%100.0%86.2% Table 5Comparison of RimNet with classical machine-learning baselines on the test dataset (file-level classification).ModelAccuracyPrecision (BAD)Recall (BAD)F1 score (BAD)SVM (RBF kernel)85.3%72.7%100.0%84.2%Random forest78.8%65.1%98.6%78.5%RimNet (proposed)87.5%75.8%100.0%86.2% Comparison of RimNet with classical machine-learning baselines on the test dataset (file-level classification). While the SVM and random forest models achieve competitive performance, the proposed RimNet framework provides improved recall for the defect (BAD) class and a more favourable balance between sensitivity and precision. In particular, RimNet achieves a recall of 1.00 in the analysed dataset (i.e., no missed defects in the analysed dataset), which is critical in safety-related diagnostic applications. The baseline models were trained and evaluated on the same data partition as RimNet. These results confirm the advantage of deep-learning-based feature extraction over classical approaches relying on fixed feature representations. The overall classification accuracy reached 87.5%. From a diagnostic perspective, the key result concerns the detection of material defects. The classifier achieved 72 true positives, 23 false alarms, and no missed defects (FN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {BAD}} = 0$$\end{document}). For the BAD class, RimNet achieves high precision while maintaining a recall of 1.00 in the analysed dataset, as summarised in Table 5. The absence of false negatives indicates that no defective rim was misclassified as non-defective. Consequently, the classifier operates in a deliberately over-sensitive regime at the file level, prioritising defect detection at the expense of a moderate number of false alarms. The decision threshold was set to the default value \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau _{\textrm{file}} = 0.50$$\end{document}, while the wheel-level threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau _{\textrm{wheel}}$$\end{document} was calibrated on the validation set. ROC–AUC analysis Receiver operating characteristic (ROC) analysis yielded an area under the curve (AUC) of 0.982, indicating strong separability between the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {OK}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {BAD}$$\end{document} classes across a wide range of operating thresholds. Distribution of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\textrm{file}}$$\end{document} The distribution of file-level defect probabilities \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\textrm{file}}$$\end{document} shows clear separation between the classes: the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {OK}$$\end{document} population is concentrated near \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\textrm{file}} \approx 0$$\end{document}, whereas the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {BAD}$$\end{document} class is shifted towards \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\textrm{file}} \approx 1$$\end{document}. Only a narrow overlap region is observed, corresponding mainly to hard-negative and borderline cases. The decision threshold was set to the default value \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau _{\textrm{file}} = 0.50$$\end{document}, as calibrated Softmax outputs from the validation set were not available. The reported results are consistent across the analysed dataset and indicate stable classifier behaviour under the adopted evaluation protocol. Wheel-level classification results (probabilistic fusion) The second stage of the RimNet pipeline aggregates file-level predictions into a single diagnostic score at the wheel level. This probabilistic fusion mechanism reduces variability associated with impulsive acoustic measurements and produces a stable probability of defect presence, denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\textrm{wheel}}$$\end{document}. For the analysed dataset, no wheel-level misclassifications were observed under the adopted evaluation protocol, correctly identifying all N = 47 wheels (29 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {OK}$$\end{document} and 18 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {BAD}$$\end{document}). The distribution of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\textrm{wheel}}$$\end{document} is shown in Fig. 8. Serviceable rims form a compact cluster near \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\textrm{wheel}} \approx 0$$\end{document}, while defective rims occupy the region close to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\textrm{wheel}} \approx 1$$\end{document}, with no overlap between the classes. The calibrated decision threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau _{\textrm{wheel}}$$\end{document} lies within the gap between these distributions, ensuring robust separation between defective and non-defective rims. The corresponding ROC analysis yielded an AUC of 1.00 for the analysed wheel-level dataset, indicating excellent class separability under the adopted evaluation protocol. These results demonstrate that probabilistic fusion effectively converts noisy per-impact predictions into a reliable wheel-level decision. The two-stage RimNet architecture—CNN-based file-level inference followed by calibrated probabilistic fusion—therefore provides a robust framework for practical rim diagnostics where multiple acoustic excitations are available.Fig. 8Distribution of the wheel-level defect probability \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\textrm{wheel}}$$\end{document} for serviceable (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {OK}$$\end{document}) and defective (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {BAD}$$\end{document}) rims obtained using the probabilistic-fusion classifier. The vertical line denotes the calibrated threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau _{\textrm{wheel}}$$\end{document}, separating the two distributions without visible overlap under the adopted evaluation protocol. Distribution of the wheel-level defect probability \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\textrm{wheel}}$$\end{document} for serviceable (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {OK}$$\end{document}) and defective (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {BAD}$$\end{document}) rims obtained using the probabilistic-fusion classifier. The vertical line denotes the calibrated threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau _{\textrm{wheel}}$$\end{document}, separating the two distributions without visible overlap under the adopted evaluation protocol. Impact of domain filtering and hard-negative mining The performance of the RimNet classifier is strongly influenced by the structure of the training dataset. Two components of the proposed pipeline—domain filtering (DF) and hard-negative mining (HNM)—play a key role in stabilising the feature space and improving separability between the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {OK}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {BAD}$$\end{document} classes. Their influence is further analysed in the ablation study presented in Section "Ablation study of RimNet components". Domain filtering Domain filtering removes recordings associated with pronounced geometric deviations (GD) and acoustically ambiguous cases (X-type). As discussed in Sections "Representative acoustic signatures" and "Repeatability of the acoustic signal for a single rim", such recordings introduce spectral variability unrelated to material defects and may destabilise the training process. Applying DF results in a more compact representation of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {OK}$$\end{document} class and reduces false alarms at the file level. In particular, the average values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\textrm{file}}$$\end{document} for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {OK}$$\end{document} rims shift towards zero while their variance decreases. This leads to more consistent predictions at the wheel level, where fewer unstable file-level outputs influence the fused probability \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\textrm{wheel}}$$\end{document}. Importantly, domain filtering does not remove any recordings corresponding to material-defect cases; it only eliminates variability caused by geometric deviations or measurement artefacts. Hard-negative mining Hard-negative mining complements domain filtering by explicitly introducing the most challenging \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {OK}$$\end{document} recordings during training. These include rims with atypical acoustic responses or irregular excitation conditions that could otherwise be misclassified as defective. By repeatedly exposing the network to such difficult negative examples, HNM encourages the model to learn sharper decision boundaries. As shown in the ablation study, this mechanism reduces high-probability outliers within the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {OK}$$\end{document} class and improves precision and F1-score for the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {BAD}$$\end{document} class. Effect of DF and HNM on network performance The combined use of domain filtering and hard-negative mining significantly improves prediction stability. Without these mechanisms, the model exhibits larger fluctuations of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\textrm{file}}$$\end{document} for the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {OK}$$\end{document} class and occasional false alarms. After applying DF and HNM, the class distributions become clearly separated and the calibrated threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau _{\textrm{wheel}}$$\end{document} lies within a distinct gap between the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {OK}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {BAD}$$\end{document} classes (Fig. 8). This improvement is also reflected in ROC analysis: domain filtering removes acoustically noisy samples, while hard-negative mining increases the separation margin between classes. The resulting wheel-level AUC of 1.00 confirms the effectiveness of both mechanisms. Significance for practical diagnostics Domain filtering and hard-negative mining encode practical knowledge derived from real-world rim inspection. Domain filtering prevents geometric deviations from being interpreted as material defects, whereas HNM improves recognition of rims with unusual yet acceptable acoustic responses. Together, these mechanisms increase the robustness and interpretability of RimNet under industrial conditions, where variability between rims and measurement conditions is unavoidable. RimNet Predictor and uncertainty analysis of classification After training the RimNet classifier in the binary configuration (OK/BAD), a lightweight diagnostic application termed RimNet Predictor was developed to enable inference under conditions similar to real workshop operation. The application loads the trained model (rimnet_3st_X_binary_*.mat), performs preprocessing identical to the training stage (ADF segmentation, normalisation, feature extraction), and treats each .wav recording as an unknown case. For each recording the system computes class probabilities\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Score}_{\text {OK}}, \quad \text {Score}_{\text {BAD}}, \quad \text {Score}_{\text {OK}} + \text {Score}_{\text {BAD}} = 1,$$\end{document}and assigns the final label \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {PredLabel} \in \{\text {OK},\text {BAD}\}$$\end{document} using the argmax rule. Additionally, several uncertainty indicators are calculated, including prediction entropy, normalised margin, Brier score, and the corresponding logit value z. To evaluate the behaviour of the system in inference mode, the predictor was executed on the complete dataset of 930 ADF recordings. Predicted labels were compared with reference labels recovered from file names, and the associated uncertainty measures were subsequently analysed. Classification performance in inference mode For the complete dataset (N = 930 recordings), RimNet Predictor produced 898 correct classifications and 32 incorrect decisions, corresponding to an overall accuracy of approximately 96.6%. The confusion statistics are summarised as follows:True BAD: 435 correctly classified as BAD and 14 misclassified as OK,True OK: 463 correctly classified as OK and 18 misclassified as BAD.For the defect class, the resulting performance metrics are:recall\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {BAD}} \approx 96.9\%$$\end{document},precision\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {BAD}} \approx 96.0\%$$\end{document},\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {F1}_{\text {BAD}} \approx 0.96$$\end{document}.In this inference scenario, the true BAD category includes both material defects (VD) and geometric defects (GD), as all non-compliant states are treated as belonging to the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {BAD}$$\end{document} class. True BAD: 435 correctly classified as BAD and 14 misclassified as OK, True BAD: 435 correctly classified as BAD and 14 misclassified as OK, True OK: 463 correctly classified as OK and 18 misclassified as BAD. True OK: 463 correctly classified as OK and 18 misclassified as BAD. recall\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {BAD}} \approx 96.9\%$$\end{document}, recall\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {BAD}} \approx 96.9\%$$\end{document}, precision\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {BAD}} \approx 96.0\%$$\end{document}, precision\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {BAD}} \approx 96.0\%$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {F1}_{\text {BAD}} \approx 0.96$$\end{document}. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {F1}_{\text {BAD}} \approx 0.96$$\end{document}. Uncertainty measures: general overview Global uncertainty statistics are summarised in Table 6. The uncertainty analysis is based on several complementary measures describing prediction confidence and decision stability. Prediction entropy is defined as:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H(p) = - \sum _{i} p_i \log p_i$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_i$$\end{document} denotes the predicted class probability. Higher entropy values indicate greater uncertainty of the classifier. The decision margin is computed as:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$m = p_{\max } - p_{\text {second}}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\max }$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {second}}$$\end{document} denote the two highest class probabilities. Lower margin values correspond to predictions closer to the decision boundary. The Brier score is defined as:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$BS = \frac{1}{N}\sum _{i=1}^{N}(p_i - y_i)^2$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_i$$\end{document} is the predicted probability and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_i$$\end{document} is the reference label. Lower Brier scores indicate better agreement between predictions and observed labels. The adopted uncertainty metrics follow standard definitions commonly used in uncertainty-aware machine-learning analysis 44–46. The uncertainty metrics were implemented using MATLAB R2024b routines and custom analysis scripts. Additionally, the logit value z, corresponding to the pre-Softmax activation of the classifier, was analysed as an auxiliary confidence indicator. Correct predictions are characterised by very low entropy and uncertainty (0.022 and 0.007 respectively), whereas misclassified samples exhibit substantially higher values (0.431 and 0.126). At the same time, the normalised decision margin decreases from 0.989 to 0.873 for incorrect predictions, indicating that classification errors occur primarily near the decision boundary.Table 6Uncertainty measures for correct and incorrect decisions of the RimNet Predictor.MeasureCorrect decisions(N = 898)Incorrect decisions(N = 32)Entropy0.0220.431Uncertainty0.0070.126Normalised margin0.9890.873LogitHigh, stableReducedBrier scoreVery lowElevated Table 6Uncertainty measures for correct and incorrect decisions of the RimNet Predictor.MeasureCorrect decisions(N = 898)Incorrect decisions(N = 32)Entropy0.0220.431Uncertainty0.0070.126Normalised margin0.9890.873LogitHigh, stableReducedBrier scoreVery lowElevated Uncertainty measures for correct and incorrect decisions of the RimNet Predictor. The strong contrast between these groups indicates that incorrect predictions are typically associated with increased entropy, higher uncertainty, and reduced decision margins. Domain-wise uncertainty analysis (VC, VD, GD) Although the RimNet Predictor operates in binary mode (OK/BAD), additional semantic information can be recovered from file metadata, allowing three groups to be analysed: serviceable rims (VC), material defects (VD), and geometric defects (GD). The corresponding mean prediction and uncertainty measures for these groups are summarised in Table 7.Table 7Mean prediction and uncertainty measures for the three semantic groups.GroupNErrorsScore_BAD (mean)Uncertainty (mean)Entropy (mean)Brier score (mean)NormMargin (mean)OK (VC)48118 (3.7%)0.0470.0130.048low0.97BAD_MAT (VD)3760 (0%)0.99990.00010.0011≈ 00.9998BAD_GEO (GD)7314 (19.2%)0.800.0320.111elevated0.94 Table 7Mean prediction and uncertainty measures for the three semantic groups.GroupNErrorsScore_BAD (mean)Uncertainty (mean)Entropy (mean)Brier score (mean)NormMargin (mean)OK (VC)48118 (3.7%)0.0470.0130.048low0.97BAD_MAT (VD)3760 (0%)0.99990.00010.0011≈ 00.9998BAD_GEO (GD)7314 (19.2%)0.800.0320.111elevated0.94 Mean prediction and uncertainty measures for the three semantic groups. Material defects are detected with very high confidence and no observed classification errors within the analysed dataset. Serviceable rims also exhibit low entropy and high decision margins. In contrast, geometric defects show increased uncertainty and a higher error rate, reflecting their weaker acoustic manifestation. Interpretation of misclassification cases All 32 misclassifications originate from the GD domain. These cases correspond to rims with geometric disturbances whose acoustic spectra partially resemble those of geometrically compliant rims. Consequently, they occupy a transitional region near the decision boundary between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {OK}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {BAD}$$\end{document} classes. The associated uncertainty indicators confirm this interpretation: entropy and uncertainty are substantially higher for misclassified samples, while the decision margin is significantly reduced. Significance for practical diagnostics The obtained results demonstrate that RimNet not only achieves high classification accuracy but also provides interpretable uncertainty information. Material defects are detected with very high confidence, whereas borderline cases—primarily associated with geometric deviations—are accompanied by increased entropy and reduced decision margins. This behaviour is desirable for practical applications, as the system can identify cases requiring additional inspection or repeated measurement, thereby supporting reliable diagnostic decisions in workshop environments. Case studies This section presents four representative case studies illustrating the behaviour of the RimNet Predictor under practical diagnostic conditions: (i) a fully serviceable rim (OK), (ii) a rim with a material defect (BAD/VD), (iii) a challenging borderline case associated with geometric deviation, and (iv) a complete wheel set with heterogeneous technical condition. Each case includes the model prediction, selected uncertainty measures, and a short diagnostic interpretation. All case studies are based on real acoustic measurements acquired under workshop conditions. No additional impact repetitions were performed specifically for model evaluation. Case A: serviceable rim Case A represents a fully serviceable rim (VC + GC); see Table 8. The acoustic response exhibits a stable modal structure without local energy anomalies or spectral deviations typical of defective rims.Table 8RimNet Predictor results for a serviceable rim (Case A).ParameterValue/interpretationData fileACT_H2_ALU_6Jx14H2_2024_23_VC_GC_092_082_062_055_FR.wavTrueLabel (ground truth)OKPredLabel (model output)OK\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {file}}$$\end{document}0.03–0.11 (low defect confidence score)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {wheel}}$$\end{document} (probabilistic fusion)0.07Final decisionOKLogit-3.1Entropy0.27 (low)Decision margin0.86 (strong dominance of the OK class)Brier score0.006–0.02 (very good prediction consistency) Table 8RimNet Predictor results for a serviceable rim (Case A).ParameterValue/interpretationData fileACT_H2_ALU_6Jx14H2_2024_23_VC_GC_092_082_062_055_FR.wavTrueLabel (ground truth)OKPredLabel (model output)OK\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {file}}$$\end{document}0.03–0.11 (low defect confidence score)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {wheel}}$$\end{document} (probabilistic fusion)0.07Final decisionOKLogit-3.1Entropy0.27 (low)Decision margin0.86 (strong dominance of the OK class)Brier score0.006–0.02 (very good prediction consistency) RimNet Predictor results for a serviceable rim (Case A). The rim was correctly classified as OK, with consistently low file-level probabilities and a wheel-level probability of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {wheel}} = 0.07$$\end{document}. Low entropy, a large decision margin, and a low Brier score indicate a confident and stable prediction. This case represents the expected behaviour of the model for a nominal rim. Case B: explicit material defect Case B represents a rim with a material defect (VD); see Table 9. Its acoustic response is characterised by disturbed modal behaviour and atypical energy distribution.Table 9RimNet Predictor results for a rim with a material defect (Case B).ParameterValue/interpretationData fileNis_Sun_ALU_615Jx17H2_2018_210_VD_GD_092_066_087_091_FR.wavTrueLabel (ground truth)BAD (structural/material defect — VD)PredLabel (network output)BAD\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {file}}$$\end{document}0.82–0.97 (very high defect confidence)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {wheel}}$$\end{document} (probabilistic fusion)0.94Final decisionBADLogit+2.8Entropy0.11 (very low)Decision margin0.76 (strong dominance of the BAD class)Brier score0.03–0.07 (very good prediction consistency) Table 9RimNet Predictor results for a rim with a material defect (Case B).ParameterValue/interpretationData fileNis_Sun_ALU_615Jx17H2_2018_210_VD_GD_092_066_087_091_FR.wavTrueLabel (ground truth)BAD (structural/material defect — VD)PredLabel (network output)BAD\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {file}}$$\end{document}0.82–0.97 (very high defect confidence)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {wheel}}$$\end{document} (probabilistic fusion)0.94Final decisionBADLogit+2.8Entropy0.11 (very low)Decision margin0.76 (strong dominance of the BAD class)Brier score0.03–0.07 (very good prediction consistency) RimNet Predictor results for a rim with a material defect (Case B). The model correctly detected the defect, producing consistently high file-level probabilities and a wheel-level probability of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {wheel}} = 0.94$$\end{document}. The low entropy and large decision margin indicate a confident and internally consistent BAD decision. This case demonstrates the ability of RimNet to detect material degradation with high reliability. Case C: challenging geometric defect (GD) Case C represents a borderline example from the GD domain; see Table 10. Although the rim was formally non-compliant due to geometric deviation, its acoustic response did not contain unambiguous signatures of material damage, placing it close to the decision boundary.Table 10RimNet Predictor results for the borderline geometric case (Case C).ParameterValue/interpretationData fileVW_Gol_ALU_65Jx16H2_2006_280_VC_GC_088_101_095_099_RL.wavTrueLabel (ground truth)OK (atypical geometry — GD domain)PredLabel (model output)BAD (false positive)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {file}}$$\end{document}0.42–0.61 (values close to the decision threshold)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {wheel}}$$\end{document} (probabilistic fusion)0.58 (slightly above \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau _{\text {wheel}}$$\end{document})Final decisionBAD (low decision stability)Logit-0.11 to +0.19Entropy0.57 (high)Decision margin0.05–0.18 (very low dominance of the winning class)Brier score0.23–0.29 (degraded prediction consistency – uncertainty indicator) Table 10RimNet Predictor results for the borderline geometric case (Case C).ParameterValue/interpretationData fileVW_Gol_ALU_65Jx16H2_2006_280_VC_GC_088_101_095_099_RL.wavTrueLabel (ground truth)OK (atypical geometry — GD domain)PredLabel (model output)BAD (false positive)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {file}}$$\end{document}0.42–0.61 (values close to the decision threshold)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {wheel}}$$\end{document} (probabilistic fusion)0.58 (slightly above \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau _{\text {wheel}}$$\end{document})Final decisionBAD (low decision stability)Logit-0.11 to +0.19Entropy0.57 (high)Decision margin0.05–0.18 (very low dominance of the winning class)Brier score0.23–0.29 (degraded prediction consistency – uncertainty indicator) RimNet Predictor results for the borderline geometric case (Case C). This case illustrates a borderline false-positive decision. File-level probabilities remain close to threshold, and the wheel-level probability only slightly exceeds \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau _{\text {wheel}}$$\end{document}. At the same time, high entropy, low decision margin, degraded Brier score, and logits close to zero clearly indicate reduced confidence. This confirms that RimNet does not fail silently and that uncertainty measures can be used to flag such cases for additional inspection. Case D: heterogeneous condition of a wheel set (Renault Clio) To illustrate practical deployment, a complete set of four rims from a Renault Clio vehicle was analysed. For each wheel, the RimNet Predictor returned a binary decision together with uncertainty measures; results are summarised in Table 11.Table 11RimNet Predictor results for the Renault Clio wheel set.Pos.Pred\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {p(OK)}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {p(BAD)}$$\end{document}LogitDistNormMarginEntropyBrierRRBAD0.00250.99755.990.4970.99500.02530.00001RLBAD0.00090.99916.980.4990.99810.01070.00000FROK0.68990.31010.800.1900.37980.89330.9519FLBAD0.00260.99745.960.4970.99480.02590.00001 Table 11RimNet Predictor results for the Renault Clio wheel set.Pos.Pred\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {p(OK)}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf {p(BAD)}$$\end{document}LogitDistNormMarginEntropyBrierRRBAD0.00250.99755.990.4970.99500.02530.00001RLBAD0.00090.99916.980.4990.99810.01070.00000FROK0.68990.31010.800.1900.37980.89330.9519FLBAD0.00260.99745.960.4970.99480.02590.00001 RimNet Predictor results for the Renault Clio wheel set. Three rims (RR, RL, FL) were classified as BAD with very high confidence, reflected by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p(\text {BAD}) \approx 0.997$$\end{document}–0.999, low entropy, and large normalised margins. In contrast, the FR rim was formally classified as OK but exhibited high entropy, low margin, and poor Brier score, indicating a borderline acoustic state. Figure 9 summarises the uncertainty profile of the wheel set.Fig. 9Radar-style visualization of normalized uncertainty measures for the Renault Clio wheel set (RR, RL, FR, FL). Defective rims (RR, RL, FL) exhibit wide, stable profiles, whereas the FR rim shows a distorted shape indicative of a borderline diagnostic state. Radar-style visualization of normalized uncertainty measures for the Renault Clio wheel set (RR, RL, FR, FL). Defective rims (RR, RL, FL) exhibit wide, stable profiles, whereas the FR rim shows a distorted shape indicative of a borderline diagnostic state. The Renault Clio example demonstrates that RimNet can assess a complete wheel set during routine service and distinguish clearly defective rims from borderline cases. In particular, uncertainty measures enable identification of suspicious rims requiring repeated measurement or additional inspection. Summary of the case studies The presented case studies confirm the practical applicability of RimNet under real workshop conditions. The system correctly identifies both serviceable rims and rims with material defects, while borderline cases are accompanied by elevated uncertainty measures that clearly indicate reduced decision reliability. The results show that RimNet combines accurate classification with interpretable diagnostics. In addition to the final class label, the system provides auxiliary information such as entropy, decision margin, and Brier score, which facilitates the identification of ambiguous cases and supports informed decision-making in service practice. The analysed examples also demonstrate that the proposed pipeline is robust to natural acoustic variability and that its predictions are driven primarily by defect-related acoustic signatures rather than by incidental fluctuations of measurement conditions. As a result, RimNet can operate as a practical service-support tool for rapid and objective assessment of wheel rims during routine inspection and seasonal tyre replacement. Future work will focus on several directions. First, the current framework will be extended towards continuous assessment of rim condition, including regression-based estimation of geometric deformation and tracking of degradation trends over time. Second, transfer learning and continual learning strategies will be investigated to enable adaptive improvement of the model without full retraining. Third, additional calibration techniques, such as temperature scaling and expected calibration error analysis, will be incorporated to further improve prediction reliability and interpretability. Finally, the system will be extended towards operation with heterogeneous acquisition devices and towards automatic reporting for complete wheel sets in practical service environments. In its current form, RimNet v3s-X operates as a binary OK/BAD classifier, while complete diagnostic assessment combines material-state prediction with an independent geometric evaluation, reflecting actual workshop procedures. Overall, the case studies confirm that RimNet provides a repeatable, objective, and operationally relevant diagnostic solution for aluminium wheel rims. Ablation study of RimNet components The contribution of the main components of the RimNet pipeline—domain filtering (DF), hard-negative mining (HNM), and calibration (CAL)—was analysed through a qualitative ablation study. Four model configurations were considered: (i) a baseline CNN operating on individual acoustic recordings, (ii) CNN with domain filtering (DF), (iii) CNN with DF and hard-negative mining (HNM), and (iv) the full RimNet pipeline including wheel-level calibration (CAL). Baseline CNN The baseline model provides reasonable discrimination between OK and BAD rims but exhibits high variability of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\textrm{file}}$$\end{document} within the OK class. In several cases, geometrically compliant rims receive elevated defect probabilities, indicating sensitivity to geometric variability rather than material anomalies. This behaviour leads to unstable wheel-level aggregation and increased false alarms. CNN + DF Introducing domain filtering substantially stabilises the feature space. Removing recordings affected by pronounced geometric deviations (GD) and acoustically ambiguous X-type cases reduces non-diagnostic variability and shifts the distribution of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\textrm{file}}$$\end{document} for the OK class towards lower values. As a result, prediction consistency improves and wheel-level fusion becomes more reliable. CNN + DF + HNM Hard-negative mining further improves the model by exposing it to the most challenging OK recordings. This leads to a sharper separation between compliant rims and defective rims, reduces high-probability outliers, and increases the stability of the fused probability \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\textrm{wheel}}$$\end{document}. Full RimNet (CNN + DF + HNM + CAL) The complete RimNet pipeline additionally applies wheel-level calibration through the threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau _{\textrm{wheel}}$$\end{document}. After aggregation, the predicted confidence scores form two clearly separated distributions, with OK rims near \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\textrm{wheel}}\approx 0$$\end{document} and BAD rims near \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\textrm{wheel}}\approx 1$$\end{document}. The final model achieves no observed classification errors at the wheel level on the analysed dataset, with an AUC of 1.00 under the adopted evaluation protocol. This result confirms that the combination of DF, HNM, and calibration yields a stable and interpretable diagnostic system. Overall, the ablation study confirms that each component contributes to improved robustness: domain filtering removes non-diagnostic variability, hard-negative mining sharpens the decision boundary, and calibration stabilises the final wheel-level decision (Table 12).Table 12Qualitative ablation study of RimNet components.Model variantDFHNMCALObserved effect and interpretationBase CNN–––High variability of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {file}}$$\end{document} and an elevated false-alarm rate. The model overfits to rims with geometric deviations and misclassifies some compliant rims as borderline material defects.CNN + DF\checkmark––Clear reduction in spectral variability. Removing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {GD}$$\end{document} and acoustically ambiguous X-class recordings stabilises the feature space and decreases the number of false positives among compliant rims.CNN + DF + HNM\checkmark\checkmark–Hard-negative mining enforces separation between compliant acoustic outliers and true material defects. The decision boundary becomes sharper, reducing the misclassification of compliant rims with atypical spectra.CNN + DF + HNM +CAL (RimNet)\checkmark\checkmark\checkmarkBest overall performance. The thresholds \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau _{\text {file}}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau _{\text {wheel}}$$\end{document} optimise the F1 score of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {BAD}$$\end{document} class and yield stable and interpretable confidence estimates \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {file}}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {wheel}}$$\end{document}. No incorrect BAD decisions were observed at the wheel level in the analysed dataset.DF domain filtering; HNM hard-negative mining; CAL dual-level calibration. Table 12Qualitative ablation study of RimNet components.Model variantDFHNMCALObserved effect and interpretationBase CNN–––High variability of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {file}}$$\end{document} and an elevated false-alarm rate. The model overfits to rims with geometric deviations and misclassifies some compliant rims as borderline material defects.CNN + DF\checkmark––Clear reduction in spectral variability. Removing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {GD}$$\end{document} and acoustically ambiguous X-class recordings stabilises the feature space and decreases the number of false positives among compliant rims.CNN + DF + HNM\checkmark\checkmark–Hard-negative mining enforces separation between compliant acoustic outliers and true material defects. The decision boundary becomes sharper, reducing the misclassification of compliant rims with atypical spectra.CNN + DF + HNM +CAL (RimNet)\checkmark\checkmark\checkmarkBest overall performance. The thresholds \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau _{\text {file}}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau _{\text {wheel}}$$\end{document} optimise the F1 score of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {BAD}$$\end{document} class and yield stable and interpretable confidence estimates \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {file}}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\text {wheel}}$$\end{document}. No incorrect BAD decisions were observed at the wheel level in the analysed dataset.DF domain filtering; HNM hard-negative mining; CAL dual-level calibration. Qualitative ablation study of RimNet components. CNN + DF + HNM + CAL (RimNet) DF domain filtering; HNM hard-negative mining; CAL dual-level calibration. Limitations and future work Although RimNet demonstrates high diagnostic performance under the evaluated conditions, several limitations should be acknowledged. First, the model operates in a single-instance learning regime and does not explicitly exploit the multi-instance structure of repeated impact measurements. Incorporating multi-instance learning (MIL) or attention-based aggregation could further improve robustness to impact-level variability. Second, the present study focuses exclusively on acoustic signals and does not include multimodal baselines combining acoustics with imaging or geometric inspection. Integrating complementary sensing modalities could improve defect localisation and diagnostic interpretability. Third, the ablation analysis presented in this work is primarily qualitative. A fully quantitative component-wise evaluation was not performed due to limitations in industrial data availability and acquisition constraints. Finally, although the dataset reflects realistic service conditions, it originates from a limited number of operational environments. Future work will therefore focus on expanding the dataset, incorporating multimodal information, and investigating advanced fusion and calibration strategies for robust large-scale deployment. The current study compares RimNet primarily with classical machine-learning approaches and does not include additional deep neural architectures such as ResNet, EfficientNet, or transformer-based models. Quantitative benchmarking of all intermediate RimNet variants was not systematically preserved during iterative model development. The reported results should be interpreted in the context of the analysed dataset and may not fully generalise to all possible rim types and operating conditions. Conclusions This work introduces RimNet, a complete acoustic diagnostic framework for aluminium wheel rims based on convolutional neural networks and uncertainty-aware prediction analysis. The proposed pipeline integrates impulse-based acoustic measurements, time–frequency representations, domain filtering, and hard-negative mining, enabling robust detection of material-related rim defects. Experimental results demonstrate high classification performance both at the file level and at the aggregated wheel level, with clear separation between serviceable rims and rims exhibiting structural defects. The analysis of uncertainty indicators—including entropy, decision margins, Brier score, and logit values—shows that high model confidence strongly correlates with correct decisions, whereas borderline cases are characterised by increased uncertainty. An important practical advantage of the proposed approach is its minimal hardware requirement. The system relies only on a microphone and a single mechanical impulse, which enables straightforward deployment in workshop environments without specialised laboratory equipment. The presented results demonstrate the potential of acoustic-based deep learning approaches for practical rim diagnostics under real-world conditions. At the same time, it should be noted that the performance is influenced by dataset characteristics, including class distribution and the availability of representative defect cases. Future work will focus on extending the dataset, improving robustness under varying measurement conditions, and further refining the integration of acoustic and geometric assessment to enhance diagnostic reliability. The proposed approach offers a promising direction for the development of low-cost, non-invasive diagnostic tools suitable for deployment in workshop environments.