Research on a belt roller fault diagnosis method based on adaptive band-time KAN Authors: Zhiyong Yang, Guangxu Luo, Guilin Hu, Jun Wang, Hongyu Zhang, Fucheng Yang, Lei Li, Wenhan Zhao Published: 2026 ABSTRACT AbstractBelt conveyor idlers frequently fail under high-load harsh conditions, causing system shutdowns. Existing deep learning-based fault diagnosis methods suffer from insufficient frequency resolution and poor dynamic adaptability. To address this, this paper proposes a fault diagnosis framework based on adaptive frequency-band KAN: First, an adaptive frequency-band Mel filter bank designed based on fault mechanisms enhances resolution in critical fault frequency bands through non-uniform frequency-axis remapping and third-order peak detection. Second, a temporal convolutional network is integrated to expand the receptive field and capture cross-period features. A Kolmogorov-Arnold Networks (KAN) layer is introduced to dynamically analyze nonlinear coupling relationships in the frequency domain using learnable B-spline basis functions. This model achieves synergistic optimization of feature resolution enhancement and dynamic modeling, significantly improving diagnostic accuracy and cross-condition generalization capability for roller faults. Under actual conveyor roller operating conditions, fault prediction accuracy reaches 81.25%, fully validating the model’s adaptability to real-world industrial scenarios.Keywords: Idler, Fault diagnosis, Deep learning, Mel spectrogramSubject terms: Engineering, Mathematics and computing Introduction In heavy industrial sectors such as coal mining, mining operations, and port logistics, belt conveyors serve as the core arteries for material transportation. Their operational stability directly impacts production safety and economic efficiency. As critical load-bearing components of the conveying system, idlers endure prolonged exposure to high loads, intense impacts, and corrosive environmental conditions. This leads to frequent failures including bearing damage, roller eccentricity, and insufficient lubrication. Research indicates that idler failures are a primary cause of belt conveyor system downtime. In mining, port, and power generation sectors, such failures can halt entire production lines, resulting in substantial economic losses1. Abnormal vibration serves as the primary manifestation of idler malfunction; noticeable changes in vibration signals occur when idler bearings deteriorate or the idler becomes eccentric2. However, traditional roller fault diagnosis primarily relies on vibration signal analysis and manual inspections, employing signal processing techniques like short-time Fourier transforms for fault detection3. These methods perform poorly when handling non-stationary mechanical vibration signals. The high-noise environment in industrial settings often masks early-stage fault signals. Furthermore, they require extensive expert experience for feature selection and fault judgment, resulting in low automation levels and high false alarm rates4,5. In recent years, deep learning technology has demonstrated tremendous potential in the field of mechanical fault diagnosis. Convolutional neural networks can automatically extract deep features from vibration signals, eliminating the cumbersome manual feature engineering required by traditional methods. Research indicates that deep learning-based fault diagnosis methods can achieve high diagnostic accuracy in bearing fault detection6,7. At the feature extraction level, the Mel-spectrogram, as a time-frequency analysis tool, exhibits unique advantages in mechanical fault diagnosis. It effectively captures the time-frequency characteristics of vibration signals and outperforms traditional MFCC features in pump equipment fault detection8. Mel-frequency mapping technology converts linear Hertz frequencies into nonlinear Mel frequencies, better adapting to the changing fault characteristics of mechanical equipment under different operating conditions9. Deep learning methods possess advantages such as automatic feature learning, end-to-end learning, and strong generalization capabilities, demonstrating good adaptability under various operating conditions10. Despite significant advancements in deep learning technology for fault diagnosis, numerous technical bottlenecks and limitations persist in practical applications for roller fault detection. At the feature extraction level, the design of traditional Mel filter banks still suffers from insufficient resolution in critical mechanical fault frequency bands, particularly when processing narrowband fault features. This limits their ability to effectively capture the heterogeneous frequency domain patterns of diverse roller faults. At the model architecture level, while CNNs can effectively capture the spatial-temporal correlations in spectrograms, their fixed convolutional kernels struggle to adaptively adjust frequency band weights, and pooling operations lead to the loss of low-frequency fault details11. RNNs and their variants can model temporal dependencies, but their recurrent structures suffer from low training efficiency and difficulty in capturing long-term dependencies12. Furthermore, existing approaches face challenges including severe data imbalance, poor adaptability to varying operating conditions, limitations of single-feature extraction methods in comprehensively capturing fault information, and the difficulty of balancing real-time requirements with diagnostic accuracy13–16. A more fundamental limitation lies in the use of fixed activation functions across existing models. Their static nonlinear mapping capabilities fail to adapt to the complex and variable response patterns of fault features in the frequency domain, resulting in insufficient feature decoupling. Distinct from a simple combination of existing techniques, the core contributions of this research are manifested at four synergistic levels. 1. A fault-mechanism-driven adaptive frequency band remapping method. To avoid masking mechanical fault features by fixed or uniform frequency partitioning, we design a nonlinear frequency-axis remapping function and a dynamic bandwidth adjustment mechanism based on the inherent frequency characteristics of faults such as idler bearing damage and eccentricity. This method achieves targeted resolution enhancement in critical narrowband fault frequency bands, effectively remedying the insufficient frequency resolution of traditional Mel-spectrograms in mechanical fault diagnosis. 1. A fault-mechanism-driven adaptive frequency band remapping method. To avoid masking mechanical fault features by fixed or uniform frequency partitioning, we design a nonlinear frequency-axis remapping function and a dynamic bandwidth adjustment mechanism based on the inherent frequency characteristics of faults such as idler bearing damage and eccentricity. This method achieves targeted resolution enhancement in critical narrowband fault frequency bands, effectively remedying the insufficient frequency resolution of traditional Mel-spectrograms in mechanical fault diagnosis. 2. A deep model synergizing temporal awareness and nonlinear decoupling. To simultaneously capture long-range periodic dependencies and complex frequency-domain coupling relationships in fault signals, this paper innovatively integrates a Temporal Convolutional Network (TCN) with a Kolmogorov-Arnold Network (KAN). The TCN module expands the receptive field through dilated convolutions to efficiently capture cross-cycle impact temporal patterns. The KAN layer, leveraging learnable B-spline basis functions, adaptively parses complex nonlinear couplings between frequency-domain features, replacing traditional static activation functions to enhance model expressiveness. 2. A deep model synergizing temporal awareness and nonlinear decoupling. To simultaneously capture long-range periodic dependencies and complex frequency-domain coupling relationships in fault signals, this paper innovatively integrates a Temporal Convolutional Network (TCN) with a Kolmogorov-Arnold Network (KAN). The TCN module expands the receptive field through dilated convolutions to efficiently capture cross-cycle impact temporal patterns. The KAN layer, leveraging learnable B-spline basis functions, adaptively parses complex nonlinear couplings between frequency-domain features, replacing traditional static activation functions to enhance model expressiveness. 3. An end-to-end optimization loop from feature preprocessing to dynamic modeling. The adaptive filter bank and the TKAN network do not operate in isolation but form a synergistic optimization mechanism. The filter bank provides high signal-to-noise ratio input features for the target fault frequency bands, upon which the TKAN further performs nonlinear decoupling and discrimination of the enhanced features. They mutually reinforce each other, jointly improving the model’s diagnostic accuracy and cross-condition generalization capability in high-noise industrial environments. 3. An end-to-end optimization loop from feature preprocessing to dynamic modeling. The adaptive filter bank and the TKAN network do not operate in isolation but form a synergistic optimization mechanism. The filter bank provides high signal-to-noise ratio input features for the target fault frequency bands, upon which the TKAN further performs nonlinear decoupling and discrimination of the enhanced features. They mutually reinforce each other, jointly improving the model’s diagnostic accuracy and cross-condition generalization capability in high-noise industrial environments. 4. A complete validation based on a real industrial scenario. This paper conducted field data collection and experiments on a belt conveyor at the South Open-pit Coal Mine, verifying the effectiveness of the proposed method under actual operating conditions. Through module ablation studies and comparative analysis with various state-of-the-art models, the comprehensive performance of AFB-TKAN is thoroughly evaluated. 4. A complete validation based on a real industrial scenario. This paper conducted field data collection and experiments on a belt conveyor at the South Open-pit Coal Mine, verifying the effectiveness of the proposed method under actual operating conditions. Through module ablation studies and comparative analysis with various state-of-the-art models, the comprehensive performance of AFB-TKAN is thoroughly evaluated. Based on the above principles, this paper constructs the Adaptive Frequency-band Temporal KAN (AFB-TKAN) fault diagnosis framework. First, a Gaussian weighting function is constructed based on the failure mechanisms of idlers to design an adaptive frequency-band Mel filter bank. Through nonlinear frequency-axis remapping, scale stretching is achieved in key fault frequency bands, increasing local filter density, while a third-order peak detection algorithm is employed for dynamic delineation of fault frequency bands. Subsequently, the optimized Mel spectrogram is fed into the fusion architecture. The TCN is utilized to capture cross-cycle impact evolution features, and simultaneously, the KAN layer is introduced to dynamically analyze the nonlinear coupling relationships among frequency-domain features. This establishes a closed-loop optimization from feature enhancement to dynamic decoupling, ultimately leading to a significant improvement in the accuracy and generalization capability of fault diagnosis. Fault diagnosis method based on AFB-TKAN To address the core challenges in traditional fault diagnosis—namely the coupling of fault features with noise and the difficulty in capturing long-range temporal dependencies—an AFB-TKAN-based fault diagnosis framework has been established. The overall algorithmic architecture is illustrated in Fig. 1. Fault diagnosis algorithm architecture based on AFB-TKAN. The adaptive feature extraction module in the model processes the power spectral density obtained via discrete Fourier transform. By adjusting the frequency response of the filter bank, it selectively enhances the energy in frequency bands strongly correlated with faults while suppressing high-frequency noise and irrelevant interference. The TKAN module takes the enhanced Mel spectrogram as input. It captures long-range temporal dependencies within fault signals through temporal convolution and attention mechanisms, deeply integrating frequency-domain enhanced features with temporal information to enhance fault pattern discrimination capabilities. Adaptive mel filter Traditional Mel filter banks are optimized based on human auditory perception, with their central logarithmic frequency distribution following the Fletcher-Munson response curve. Consequently, they exhibit severely inadequate resolution in the frequency bands characteristic of mechanical faults17. To address this issue, this paper develops an adaptive frequency-band Mel spectrogram based on roller failure mechanisms, aiming to achieve precise enhancement and analysis of the characteristic frequencies of roller damage. Mechanism of roller failure and frequency band analysis Different damage types in belt conveyor idlers exhibit distinct frequency domain characteristics in vibration signals, stemming from fundamental differences in their physical failure mechanisms: localized damage generates transient impacts that excite the system’s inherent vibration modes, while continuous wear causes shifts in steady-state vibration energy distribution18. Therefore, accurately capturing changes in energy distribution within specific frequency bands is essential for achieving fault diagnosis. (1)(1)(1)Bearing damage failure frequency bands. Bearing damage failure frequency bands. Bearing damage primarily occurs on the outer ring, inner ring, rolling elements, and cage. Bearing failure frequency correlates with rotational speed and bearing geometric parameters, calculated using the following formula. where fr is the shaft rotational frequency, Nb is the number of rolling elements, db is the diameter of the rolling elements, Dp is the bearing pitch diameter, and θ is the contact angle. (2) Calculation of roller body eccentricity fault frequency bands. (2) Calculation of roller body eccentricity fault frequency bands. Eccentricity in roller bodies can cause faults, typically resulting from manufacturing defects or deformation due to impact during operation. This manifests primarily as the shaft rotation frequency and its harmonics. Where fr is the roller rotation frequency, n=1,2,3,... is the harmonic order, k=1,2,3,... and vbelt are the belt linear speeds, and droller is the roller spacing. (3) Insufficient lubrication fault frequency band. (3) Insufficient lubrication fault frequency band. Insufficient lubrication accelerates bearing wear, leading to increased fault frequency amplitude and the emergence of sidebands. Where k = 1,2,3,… and fbearing correspond to the failure frequencies fBPFO, fBPFI, fBSF, and Vslip of the outer ring, inner ring, and rolling elements, respectively; δ represent the sliding speed of the rolling elements and the elastic deformation of the contact zone. Optimization of mel filter bank Calculations based on fault mechanisms indicate that the core characteristic frequency bands of roller faults are primarily distributed in the low-frequency region. Therefore, filter design should prioritize enhancing characteristic resolution in the low-frequency range. However, traditional Mel filter banks exhibit uniform distribution across Mel scales from 0 to 8 kHz, rendering them unsuitable for adapting to the narrowband frequency domain characteristics of roller faults19. To address this issue, this paper proposes an adaptive Mel filter bank based on fault feature frequency bands. By increasing filter density in critical frequency bands, it enhances the resolution of fault features while reducing redundant coverage in non-critical bands, thereby more efficiently focusing on fault-sensitive information. (1) Fault band segmentation. (1) Fault band segmentation. An autonomous third-order peak detection algorithm based on vibration signal power spectrum identifies core fault frequency domains. By traversing the preprocessed power spectral density, it locates the positions of the first three significant spectral peaks fp1, fp2, and fp3 within non-rotational frequency intervals. Calculate the span between peaks. Constructing an adaptive fault band. (2) Fault feature band weighting function. (2) Fault feature band weighting function. To quantify the importance differences among fault feature bands, a frequency-domain weight distribution model based on Gaussian functions is constructed. This model forms a weight peak at the core fault frequency band and transitions smoothly to non-critical bands. In the formula, fl and fh correspond to the boundaries of the fault core frequency band, fc represents the center of the band, and k>0 denotes the resolution enhancement factor, forming a Gaussian weight distribution within the fault band. (3) Nonlinear frequency-axis remapping. (3) Nonlinear frequency-axis remapping. Reconstruct the Mel-frequency mapping relationship based on a weighting function to achieve a scale stretching effect in the fault frequency band. The improved Mel mapping function is defined as. The derivative of the mapping function at the fault frequency band is. In particular, the derivative significantly increases in the fault frequency band, causing the Mel scale to stretch. (4) Adaptive filter distribution. (4) Adaptive filter distribution. Based on the nonlinear mapping , distribute the center frequencies of filters non-uniformly along the Mel axis. Let N denote the total number of filters. The sequence of center frequencies satisfies the optimization condition. This model ensures that the filter density within the fault frequency band satisfies. (5) Dynamic bandwidth adjustment. (5) Dynamic bandwidth adjustment. To enhance spectral resolution in critical frequency bands, establish a bandwidth adaptive adjustment mechanism. Where γ = 0.7 is the compression factor. The bandwidth compression ratio within the fault band is . The comparison between the optimized filter and the original filter, as shown in Fig. 2, demonstrates how fault-sensitive information is extracted by increasing the filter weights within the fault frequency band. The red shaded area in the figure indicates the core fault frequency band determined based on fault mechanisms and third-order peak detection. It can be seen that the adaptive design significantly increases the filter density in the key bands while compressing the bandwidth, and reduces filter coverage in non-critical frequency bands. Comparison of the original mel filter and the adaptive mel filter. The adaptive Mel filter bank proposed in this section has its core contribution in combining prior knowledge of fault mechanisms with data-driven peak detection, dynamically adjusting the allocation of frequency domain analysis resources. Serving as a pre-feature enhancer, it specifically addresses the issue of ’insufficient resolution at the feature extraction level,’ providing subsequent deep learning models with cleaner and more discriminative time-frequency representations. Parameter settings and implementation details To ensure the reproducibility of our method, the key implementation parameters of the adaptive Mel filter bank are detailed as follows. Total number of filters (N): Set to 20, consistent with the number in traditional Mel filter banks to ensure fair comparison, while providing sufficient degrees of freedom for frequency band reallocation. Peak detection parameters: In the third-order peak detection algorithm, the positions of the first three (k = 3) most significant non-rotational frequency harmonic peaks in the power spectrum are selected to define the core fault frequency band . The peak significance threshold is set to twice the global average power spectrum amplitude to filter out noise-induced pseudo-peaks. Weighting function hyperparameter: The resolution enhancement factor γ in the Gaussian weighting function is empirically set to 2.0 to form a sufficiently steep weight distribution within the fault frequency band . The band center μ is automatically taken as . Bandwidth compression factor (α): In the dynamic bandwidth adjustment mechanism, the compression factor α is set to 0.7. This means that within the core fault frequency band , the bandwidth of the filters is compressed to 70% of the original design bandwidth, thereby increasing the frequency resolution by approximately 1.43 times. Specifically, the selection of hyperparameters γ and α was determined through sensitivity analysis. A grid search was conducted within the ranges of γ ∈ [1.0,3.0] and α ∈ [0.5,0.9] . The results indicated that the energy concentration in the fault frequency band reached its peak when γ = 2.0 ; conversely, an excessively large γ resulted in an overly narrow frequency band, leading to the loss of edge features. For α , a value of 0.7 achieved the optimal balance between frequency resolution and noise immunity. Consequently, γ and α were fixed at 2.0 and 0.7, respectively, in this study to ensure the robustness of feature extraction. The selection of the third-order peak detection is based on a trade-off between the physical characteristics of idler vibration signals and computational efficiency. According to impact vibration theory, the impulse response generated by an idler fault typically contains the fundamental frequency and its first two harmonics. In the frequency domain, this manifests as three clusters of concentrated energy centered around the fault characteristic frequency. While selecting a higher order could capture more harmonic details, it would introduce substantial high-frequency noise and increase computational complexity. Conversely, choosing a lower-order parameter might lead to the omission of critical fault modulation sidebands. Therefore, setting k=3 represents the optimal solution for balancing feature integrity with noise immunity. Feature extraction based on adaptive mel filters The Mel-spectrogram, as a vital tool for time-frequency analysis, effectively characterizes the time-varying properties of non-stationary signals. Conventional methods generate Mel-spectrograms with insufficient resolution for narrowband fault features. This paper employs an adaptive Mel filter bank to optimize the feature extraction process. Let the original vibration signal be denoted as s(n), with a sampling frequency of fs. The signal is divided into K overlapping frames, each of which is processed using a Hamming window to eliminate edge effects. Where L is the frame length, R is the frame offset, k=0,1,...,K-1. Calculate the power spectral density for each frame using the Discrete Fourier Transform (DFT). Where N is the number of FFT points, corresponding to frequency . Apply the adaptive filter bank described above to enhance the feature frequency band. In the equation, denotes the adaptive Mel filter transformation operator, and represents the set of optimization parameters. To simulate the human ear’s logarithmic perception characteristics, the logarithmic energy of the filter bank output is calculated as the Mel-spectrogram. Where Hm(l) is the transfer function of the mth filter, and is the numerical stability factor. Fig. 3 compares the spectrograms of the Mel filter before and after optimization. It is evident that the adaptive design significantly enhances feature extraction performance for fault diagnosis by dynamically adjusting the frequency resolution distribution. The standard Mel filter uniformly distributes filter bandwidth across the Mel axis based on human auditory characteristics, resulting in identical resolution for both critical fault bands and non-fault bands. This prevents subtle variations in fault features from being highlighted. After optimization, while maintaining the total number of filters, more resources are concentrated on fault-sensitive frequency bands. By compressing the bandwidth and enhancing the density of triangular filters, the frequency resolution in this range is improved, enabling the neural network to capture subtle fault features. Simultaneously, reducing the number of filters in non-fault bands achieves dimensionality reduction. This preserves the main vibration energy trends while reducing redundant interference, significantly improving the signal-to-noise ratio of the spectrogram and effectively enhancing the distinguishability of fault features. Mel spectrum before and after optimization. TKAN fault diagnosis network model architecture The fault diagnosis network architecture proposed in this paper centers on the deep integration of Time-Convolutional Networks20 and Kolmogorov-Arnold Networks (KAN)21, with the network model architecture illustrated in Fig. 4. This model takes Mel-spectrograms as input. Through the temporal feature extraction module, it captures long-range temporal dependencies and local frequency-domain correlations. The nonlinear feature enhancement module amplifies low-frequency fault details, and the fault classification module ultimately outputs diagnostic results. This hierarchical, progressive feature processing workflow achieves precise capture and adaptive modeling of low-frequency narrowband features in Mel-spectrograms of roller faults. TKAN fault diagnosis network model partial architecture. Traditional convolutional feature extraction suffers from fixed, limited receptive fields that struggle to capture long-range dependencies, while dimensionality reduction operations result in loss of low-frequency details. To address this, the temporal feature extraction module in this model employs dilated convolutions with residual connections. This significantly expands the receptive field without increasing parameters while preserving spatio-temporal resolution, as shown in Fig. 4a. However, the extracted coarse-grained features still exhibit coupling between low-frequency fault characteristics and high-frequency noise, with complex nonlinear distributions. Traditional convolution enhancement methods struggle to adaptively adjust frequency-band weights due to fixed kernel weights. Therefore, a nonlinear feature enhancement module employs a two-layer convolution structure (as shown in Fig. 4b). This leverages the local receptive field advantage of standard convolutions to strengthen low-frequency details, suppress high-frequency noise, and enhance feature discriminability. To address the limitation of traditional activation functions in modeling the complex nonlinear patterns of fault features, the Fault Classification Module (FCM) employs the Kolmogorov-Arnold Representation Theorem 22. It replaces fixed activation functions with learnable B-spline basisfunctions to construct a KAN layer, enabling adaptive nonlinear transformation. This ultimately outputs multi-class classification probabilities to accomplish fault type identification. The structure of its KAN layer is shown in Fig. 5. KAN layer structure. Experimental verification Data description To validate the effectiveness of the proposed AFB-TKAN-based fault diagnosis model, a roller failure experiment was conducted on the No. 1 belt conveyor at Nanlu Tianyu Coal Mine. The selected roller and bearing parameters are shown in Table 1. Table 1.Experimental parameters.StructureParametersRoller bearingDeep groove ball bearing 6205-2RS (GB/T 276)Roller assemblyDiameter Φ60mm, length 250mmConveyor BeltWidth 240mm, thickness 5mm, Belt speed 4m/s Experimental parameters. The experiment involved laying optical fibers on the roller bridge structure, with fiber terminals connected to an industrial control computer to monitor phase changes in the fibers caused by vibrations. Signal acquisition software received these phase changes, which were then converted into real-time vibration signals through phase demodulation. These signals provided the raw experimental data. The entire vibration signal acquisition process is illustrated in Fig. 6. Vibration signal acquisition process. Vibration signals from normal idlers were collected as baseline data, followed by sequential installation and testing of various faulty components. For each fault type, vibration signals were captured under identical belt speed conditions, with each condition recorded continuously for 150 minutes. To enhance data analyzability, the raw vibration signals were segmented into 5-second intervals. Each operating condition yielded 1,800 valid samples, which were divided into training, validation, and test sets in a 10:4:4 ratio. Model parameter configuration To achieve high-precision diagnosis of roller failures, model parameter design must balance fault feature extraction capability with computational efficiency. Considering the time-frequency characteristics of Mel-spectrograms and the noise interference patterns in mining environments, the parameter configurations for each module are as follows: Network hyperparameters are set based on computational capacity and training sample size. The mini-batch size is set to 64, the Adam optimizer’s learning rate employs exponential decay (initial learning rate 0.01, decay base 0.9), and the maximum training epoch is 100. Cross-entropy loss for fault classification serves as the optimization objective. Specific neural network parameter configurations are detailed in Table 2. Table 2.Neural network parameter c onfiguration.ModuleNetwork iayerParametersTime-domain feature extraction moduleHollow convolutional layer T1K = 5 × 1, d = 1, s = 1, p = 2 × 0, input = 1, output = 32Feature optimization layer DBN(momentum = 0.9, eps = 1e-5) ,ReLU, Dropout(p = 0.5) Hollow convolutional layer T2K = 5 × 1, d = 2, s = 1, p = 4 × 0, Input = 32, Output = 64Nonlinear feature enhancement moduleConvolution layer CK = 1 × 3, s = 1, p = 0 × 1, input = 64, output = 64Fault classification moduleKAN layer K3-layer KAN structure, with 10 splines per feature dimension and uniformly distributed nodesDimensional reduction layer SFully connected, input = 64 × T × F, output = 6 Neural network parameter c onfiguration. Experimental results To validate the effectiveness of this model in roller fault diagnosis tasks, 1,000 samples were randomly selected from all samples in each fault category to form the training set. Subsequently, 400 samples were randomly selected from the remaining samples to form the validation set, with the rest serving as the test set. The confusion matrix of the AFB-TKAN fault diagnosis model’s prediction results obtained from the experiment is shown in Fig. 7. where “A, B, C, D, E, F” correspond to normal, inner ring failure, outer ring failure, rolling element failure, roller eccentricity, and insufficient lubrication, respectively. Fault diagnosis model prediction results. The confusion matrix data indicates that the AFB-TKAN model can perform fault classification prediction in the roller fault diagnosis task, demonstrating strong classification capabilities on both the training and test sets. The average accuracy on the training set is 0.869, while on the test set it is 0.812. The model exhibits stable generalization performance with no significant overfitting observed. Regarding fault type discrimination, normal samples achieve the highest recognition accuracy (0.932 on the training set and 0.895 on the test set), indicating significant feature differences between normal and fault states. while identification accuracy for insufficient lubrication (training set: 0.825, test set: 0.758) and roller eccentricity (training set: 0.846, test set: 0.772) was relatively lower, reflecting the higher complexity of fault features in these two categories. Fig. 8 demonstrates the performance comparison of the AFB-TKAN fault diagnosis model against other state-of-the-art mainstream fault diagnosis models (MWK-CNN23, CNN-KAN24, KNN-GNN25) on the test dataset, thereby validating the superiority of the proposed model. Fault diagnosis comparative test results. To comprehensively evaluate the model’s strengths, we employ precision (Pr), Matthews Correlation Coefficient (MCC), and Foulkes-Mallows Index (FMI) as evaluation metrics for this task, thereby providing a more holistic reflection of model performance. To verify the statistical significance of the results, all models were independently run five times. Table 3 reports the mean and standard deviation of each metric. Table 3.Comparison of experimental results.ModelAccuracyMCCFMIAFB-TKAN0.812 ± 0.0040.773 ± 0.0110.811 ± 0.009MWK-CNN0.784 ± 0.0060.742 ± 0.0150.784 ± 0.012CNN-KAN0.763 ± 0.0050.717 ± 0.0130.764 ± 0.010KNN-GNN0.724 ± 0.0080.670 ± 0.0180.725 ± 0.014 Comparison of experimental results. (1) Accuracy measures the proportion of samples predicted as positive by the model that are actually positive, reflecting the reliability of positive predictions. Where TPi is a true positive example of class i, and FPi is a false positive example of class i. (2) The Matthews correlation coefficient measures the overall prediction quality of a classifier, where 1 indicates perfect prediction and 0 indicates random prediction. It is robust for multi-classification tasks and reflects the overall correlation of the confusion matrix. Where TPi is a true positive example of class i, TNi is a true negative example of class i, FPi is a false positive example of class i, and FNi is a false negative example of class i. (3) The F1 score is the geometric mean of precision and recall, making it more sensitive to class imbalance and outliers. It is well-suited for evaluating the robustness of classification tasks. Where TPi is a true positive example of class i, TNi is a true negative example of class i, FPi is a false positive example of class i, and FNi is a false negative example of class i. Based on the metric results, the proposed model demonstrates superior accuracy compared to other benchmark models, validating its effectiveness in roller fault prediction tasks. The MCC and FMI metrics exhibit consistent trends with accuracy, both outperforming other benchmark models, indicating that this model achieves optimal overall predictive correlation and stability. For the challenging samples of “insufficient lubrication” and "idler eccentricity," which exhibited relatively lower recognition rates in the confusion matrix, Table 4 further details the Precision, Recall, and F1-score. It can be observed that despite the subtle fault features of these two categories, the AFB-TKAN model maintains metrics above 75%, outperforming the comparative models. This indicates that the adaptive frequency band design effectively enhances the separability of subtle fault features. Table 4.Comparison of experimental results on challenging samples.Fault typeModelPrecisionMCCFMIRoller bodyEccentricityAFB-TKAN0.7720.7630.761KNN-GNN0.7040.7080.721Insufficient lubricationAFB-TKAN0.7580.7490.753KNN-GNN0.7100.7150.707 Comparison of experimental results on challenging samples. Roller body Eccentricity Ablation study To intuitively demonstrate the decoupling and clustering capabilities of the KAN layer for fault features, Fig. 9 presents the t-SNE visualization of the feature distribution at the input of the Fault Classification Module (FCM). It can be observed that the features extracted by AFB-TKAN form distinct clusters in the two-dimensional space, with clear boundaries between different fault types and high intra-class compactness. In contrast, the visualization results of the other models exhibit more class overlap. This proves that the KAN layer effectively resolves the nonlinear coupling relationships of frequency-domain features by leveraging B-spline basis functions. t-SNE visualization results. To clarify the individual contributions of the Adaptive Frequency Band (AFB), Temporal Convolution Network (TCN), and KAN layers, ablation experiments were conducted. The model variants used for the ablation study are as follows: AFB-CNN: Uses only the adaptive Mel filter bank with a standard CNN (replacing TCN and KAN). AFB-CNN: Uses only the adaptive Mel filter bank with a standard CNN (replacing TCN and KAN). FB-TKAN: Uses a fixed Mel filter bank with the TKAN network described in this paper. FB-TKAN: Uses a fixed Mel filter bank with the TKAN network described in this paper. AFB-TCN: Uses the adaptive Mel filter bank with a pure TCN network (replacing the KAN layer with a regular fully connected layer with ReLU). AFB-TCN: Uses the adaptive Mel filter bank with a pure TCN network (replacing the KAN layer with a regular fully connected layer with ReLU). CNN: Uses the baseline model (standard filtering and basic CNN). CNN: Uses the baseline model (standard filtering and basic CNN). Vibration signals collected from the conveyor belt of Line 1 at the Nan Open-pit Coal Mine were used as the raw input samples for the models. The experimental results are shown in Table 5 Comparison of experimental results. Table 5.Ablation experiment results.ModelAccuracyMCCFMIAFB-TKAN0.8120.7730.811AFB-TCN0.7810.7350.761FB-TKAN0.7680.7180.742AFB-CNN0.7420.6820.710Standard CNN0.7010.6400.665 Ablation experiment results. From the analysis of Table 5) Comparing AFB-TKAN with AFB-TCN, the KAN layer brings about an accuracy improvement of approximately 3 percentage points, demonstrating the effectiveness of its adaptive nonlinear mapping in decoupling complex fault features. 2) Comparing AFB-TKAN with FB-TKAN, adaptive frequency band optimization provides an improvement of about 4 percentage points, validating its contribution to enhancing key fault band features. 3) The comparison between AFB-TCN and AFB-CNN highlights TCN’s advantage in capturing long-range temporal dependencies. The ablation experiments clearly show that the three core components proposed in this paper are complementary and indispensable, together forming the framework with optimal performance. Computational efficiency and real-time feasibility analysis Industrial fault diagnosis systems have clear real-time requirements. To assess the deployment potential of AFB-TKAN, its computational complexity was quantitatively analyzed. The experiments were conducted on a workstation equipped with an NVIDIA GeForce RTX 3090 GPU and an Intel Core i9-12900K CPU. Model parameters and computations: The total number of parameters for the complete AFB-TKAN model is approximately 85.6K. The number of floating-point operations (FLOPs) required for one forward pass is estimated to be about 0.41 GFLOPs. Table 6 compares the parameters and computational efficiency of various model variants. Table 6.Comparison results of parameter count and computational efficiency.ModelNumber of parameters (K)Estimated FLOPs (G)Average inference time (ms)AFB-TKAN85.60.4150.9MWK-CNN82.10.4562.7CNN-KAN88.60.5072.9KNN-GNN120.50.6875.7 Comparison results of parameter count and computational efficiency. Inference time: For a single 5-second vibration signal sample, the average end-to-end inference time of the AFB-TKAN model is 50.9 ms. Among this, adaptive filtering feature extraction takes about 10.8 ms, while TKAN network inference takes about 40.1 ms. The table analysis indicates that AFB-TKAN maintains the highest diagnostic accuracy while having lower parameter counts and computational overhead compared to models like MWK-CNN. The single-sample inference time of 50.9 ms is much shorter than the sampling interval (5 seconds), meeting the timeliness requirements for real-time online monitoring (typically requiring diagnostic delays within seconds). Although the AFB-TKAN model demonstrates impressive inference speed (50.9 ms) on the workstation, actual industrial edge deployment typically involves strict resource constraints. To evaluate the model’s applicability in resource-constrained environments, we conducted a quantitative analysis. The model contains 85.6K parameters, and after INT8 quantization, the model size can be compressed to approximately 171 KB, which fully satisfies the storage limits of most industrial gateways. However, the computation of B-spline basis functions in the KAN layer involves table lookup and interpolation operations, which have a higher computational density than standard multiply-accumulate (MAC) operations. On low-end MCUs, this may lead to increased inference latency. Therefore, while the model theoretically possesses deployment potential, for practical implementation, it is recommended to perform operator-level optimization of the KAN layer for specific edge chip architectures, or to employ knowledge distillation techniques to distill the nonlinear feature mappings into a lightweight CNN. Conclusion Computational efficiency and real-time feasibility analysis Limitations in generalization under variable working conditions Although the proposed AFB-TKAN model demonstrates excellent performance under fixed working conditions, we must honestly discuss its limitations under variable conditions. The characteristic frequency of mechanical faults is linearly positively correlated with the rotational speed, whereas the adaptive Mel filter bank designed in this paper is based on the fault characteristic frequency under a specific rotational speed. When the belt speed changes drastically, the fault characteristic frequency band shifts, which may lead to a mismatch between the preset center frequencies of the filters and the actual fault frequencies. Although the nonlinear fitting capability of the KAN layer can compensate for the feature distortion caused by this frequency shift to a certain extent, the physical limitations of the filter bank objectively exist. Future work will introduce a rotational speed adaptive mapping mechanism to dynamically adjust the stretching region of the Mel axis according to the real-time rotational frequency, thereby addressing this generalization bottleneck. Physical essence and design trade-offs of accuracy Regarding the accuracy of 0.812 on the test set, we need to conduct a more in-depth critical discussion from the perspective of physical mechanisms. This value is not merely a model defect but is determined by the physical nature of the fault signals. In the open-pit coal mine environment, idler fault signals are not only modulated by bearing damage but also interfered with by broadband noise such as belt tension fluctuations and material impact. The model converging at 0.812 indicates that under this signal-to-noise ratio (SNR), the model has extracted distinguishable feature boundaries, and the remaining errors mainly stem from the unpredictability of random noise. Furthermore, the confusion matrix shows a high misclassification rate for "insufficient lubrication." This is essentially because the change in vibration energy caused by the lubrication state is continuous rather than a discrete step change, which increases the ambiguity of the classification boundary. Parameter sensitivity and robustness Although the hyperparameters in the AFB module were optimized via grid search, we acknowledge that these parameters may exhibit sensitivity during cross-equipment applications. Ablation studies indicate that when α deviates from the optimal value of 0.7 by more than ±0.2, the model accuracy drops by over 5%. This demonstrates that the model performance is highly dependent on the design of the feature extractor. To address this issue, we introduced Dropout regularization in the KAN layer, which mitigates output fluctuations caused by minor perturbations in feature input to a certain extent and enhances the internal robustness of the model. Conclusion This study addresses the dual challenges of insufficient resolution in traditional Mel spectrograms and limited nonlinear expressive capabilities in deep learning models for roller fault diagnosis by proposing an Adaptive Frequency Band Time-Convolutional KAN Framework (AFB-TKAN). At the feature extraction level, it innovatively designs an adaptive frequency band Mel filter bank. Through fault mechanism-driven nonlinear frequency axis remapping and a third-order peak detection algorithm, it achieves resolution enhancement in critical frequency bands (e.g., bearing damage, roller eccentricity) while sparsifying non-critical bands, significantly improving the separability of narrowband fault features. At the model architecture level, it integrates a temporal convolutional network with a Kolmogorov-Arnold network. The temporal convolutional network employs dilated convolution to expand the receptive field, precisely capturing cross-cycle impact evolution features. The KAN utilizes learnable B-spline basis functions to replace fixed activation functions, adaptively decoupling the nonlinear coupling patterns of fault features in the frequency domain. Experimental validation demonstrates that this model achieves a diagnostic accuracy of 81.25%. Compared with mainstream fault detection models, AFB-TKAN exhibits higher overall precision in fault diagnosis and recognition, providing a novel approach for fault diagnosis tasks in conveyor belt idlers. Acknowledgements The authors would like to thank Letpub for the English language editing. Author contributions **Zhiyong Yang:** Writing-original draft, Methodology, Investigation,Conceptualization. **Guangxu Luo:** Writing-review amp; editing, Validation, Supervision, Methodology, Conceptualization. **Guilin Hu:** Validation, Methodology. **Hongyu Zhang** : Writing-review amp; edit-ing, Validation, Supervision, Methodology, Conceptualization. **Fucheng Yang:** Writing-review amp; editing,Validation, Supervision, Conceptualization. **Lei Li** : Supervision, Methodology, Conceptualization. **Wenhan Zhao** : Writingreview amp; editing, Validation. Funding This work were supported by “Tianchi Talent” Introduction Plan、Postdoctoral Program: 410713 and Natural Science Foundation of Xinjiang Autonomous Region 2024D01A101. Data availability The datasets used and analysed during the current study available from the corresponding author on reasonable request will be made available on request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The datasets used and analysed during the current study available from the corresponding author on reasonable request will be made available on request.