Fault diagnosis of ORC expanders using a multiscale residual network with convolutional block attention Authors: Jing Li, Zexu Gao Published: 2026 ABSTRACT AbstractExpanders in organic Rankine cycle systems serve as critical energy-conversion components in low-grade waste heat recovery installations, yet their reliable operation is threatened by faults such as bearing defects, rotor imbalance, and blade cracking. Conventional diagnostic methods often struggle with non-stationary vibration characteristics, class imbalance, and low signal-to-noise ratios inherent to these working environments. This paper proposes an improved deep residual network, referred to as multi-scale convolutional block attention module residual network, that integrates a multi-scale parallel feature extraction module with convolutional block attention mechanisms for intelligent fault diagnosis. The multi-scale module employs three parallel convolutional branches with different kernel sizes to simultaneously capture transient impulses, periodic modulation, and low-frequency envelope features across multiple temporal scales. Attention-enhanced residual blocks sequentially recalibrate channel and spatial responses to emphasize fault-sensitive features while suppressing noise interference. A training optimization scheme combining Focal Loss, cosine annealing, and targeted data augmentation is further introduced to address the small-sample imbalanced-data challenge. Five-fold cross-validation experiments conducted on a 10 kW single-screw expander test rig demonstrate that the proposed model achieves 98.11 ± 0.34% diagnostic accuracy across four health states, surpassing the standard deep residual network baseline by 6.57 percentage points, with only 3.27% relative accuracy degradation at 10 dB signal-to-noise ratio. Ablation studies confirm a multiplicative synergy between the multi-scale and attention modules, statistical significance tests validate the robustness of the observed improvements, and comparative evaluations against six benchmark methods demonstrate the superiority and generalizability of the proposed approach.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-50794-7.Keywords: Organic Rankine cycle, Expander fault diagnosis, Deep residual network, Multi-scale feature extraction, Attention mechanism, Vibration signal analysisSubject terms: Energy science and technology, Engineering, Mathematics and computing Introduction The organic Rankine cycle (ORC) has emerged as one of the most promising thermodynamic frameworks for recovering low-to-medium grade waste heat in industrial processes and harnessing renewable energy sources such as geothermal and solar thermal power1. Unlike conventional steam Rankine systems, ORC technology employs organic working fluids with lower boiling points, which makes it particularly suited to distributed energy applications where heat source temperatures fall below 300 °C2. Over the past decade, growing pressure to improve industrial energy efficiency and reduce carbon emissions has driven widespread deployment of ORC systems across sectors ranging from cement manufacturing to biomass cogeneration3. Within the ORC power generation loop, the expander functions as the central energy-conversion component, directly transforming the enthalpy of high-pressure organic vapor into mechanical shaft work. Scroll expanders, screw expanders, and radial-inflow turbines each present distinct structural characteristics, yet they share a common vulnerability: prolonged operation under harsh thermal and mechanical loads inevitably gives rise to degradation phenomena such as blade erosion, bearing fatigue, and seal leakage4. Blade wear, for instance, alters rotor balance and reduces isentropic efficiency, while bearing defects introduce broadband vibration signatures that propagate through the drivetrain. Seal deterioration, on the other hand, permits internal leakage of the working fluid, which not only undermines thermodynamic performance but may also pose environmental and safety hazards when flammable organic media are involved5. Early and accurate identification of these fault modes is therefore critical to ensuring reliable ORC operation and preventing costly unplanned shutdowns. Fault diagnosis of rotating machinery has a long research history. Classical signal-processing techniques—fast Fourier transform, wavelet packet decomposition, and empirical mode decomposition—remain valuable for extracting frequency-domain and time–frequency features from vibration data6. These methods, however, rely heavily on expert knowledge to select appropriate decomposition parameters and interpret spectral patterns, which limits their scalability in complex industrial settings. To reduce this dependence on manual feature engineering, researchers introduced machine-learning classifiers such as support vector machines and random forests, achieving notable improvements in diagnostic accuracy for well-defined fault categories7. Nevertheless, shallow models struggle to capture the highly nonlinear mappings between raw sensor signals and incipient fault states, especially when operating conditions fluctuate. To bridge this gap, the present work puts forward an improved deep residual network—termed MS-CBAM-ResNet—that systematically integrates channel attention, spatial attention, and a multi-scale parallel convolution module. It is worth emphasising at the outset that this integration is not a straightforward stacking exercise; rather, each module is specifically motivated by the physical characteristics of ORC expander vibration signals. The multi-scale front end addresses the co-existence of short-duration bearing impulses and slow rotor modulation within a single architecture, while the attention modules exploit the diversified multi-scale features to suppress fluid-induced noise channels that would otherwise dilute fault signatures. Ablation evidence presented later in this paper shows that these components interact multiplicatively rather than additively, a finding that underscores the non-trivial nature of their combination. The principal contributions of this study are as follows. (1) A multi-scale feature extraction block tailored to ORC expander vibration signals is designed, employing parallel kernels of different sizes to capture fault-related patterns across multiple temporal resolutions. (2) A deep fusion strategy that embeds both channel and spatial attention into the residual learning path is developed, enabling the network to adaptively emphasise discriminative features while suppressing irrelevant noise components. (3) A training optimisation scheme combining data augmentation, Focal Loss, and cosine annealing is introduced to address the small-sample imbalanced-data scenario that typifies real ORC field applications. (4) Comprehensive five-fold cross-validation experiments with statistical significance testing are conducted on both an ORC expander test rig and the CWRU benchmark, demonstrating robust generalisation8. The remainder of this paper is organized as follows. “Related works” section reviews related work on deep learning-based fault diagnosis, attention mechanisms, and multi-scale feature extraction. “Theoretical foundation” section presents the theoretical foundation, covering ORC expander fault mechanisms, deep residual networks, and attention theory. “Construction of the deep residual network” section details the construction of the proposed MS-CBAM-ResNet architecture. “Experimental verification” section describes the experimental setup, ablation studies, and comparative evaluations. “Discussion” section provides an extended discussion of the results, limitations, and practical implications. “Conclusion” section concludes the paper with a summary of findings and directions for future research. Related works Deep learning has fundamentally reshaped the landscape of intelligent fault diagnosis over the last several years. Convolutional neural networks can automatically learn hierarchical feature representations from one-dimensional vibration signals or two-dimensional time–frequency images, while long short-term memory networks excel at modeling temporal dependencies in sequential monitoring data9. Among various architectures, the deep residual network deserves particular attention because its skip-connection topology alleviates the gradient degradation problem, enabling the training of substantially deeper models without sacrificing convergence stability10. ResNet-based diagnostic frameworks have demonstrated competitive performance on benchmark bearing and gearbox datasets, yet their capacity to discriminate subtle fault signatures under severe background noise remains an open question. Recent studies have shown that embedding attention mechanisms into deep networks can sharpen the model’s focus on fault-sensitive feature channels and spatial regions. The squeeze-and-excitation network recalibrates channel-wise responses through a compact gating structure, while the convolutional block attention module further incorporates spatial attention to highlight informative locations within a feature map11. Meanwhile, the self-attention paradigm underlying Transformer architectures has been adapted for vibration-based diagnosis, offering a global receptive field that captures long-range signal dependencies12. In parallel, multi-scale feature extraction strategies—typically realized through parallel convolution branches with varying kernel sizes—have proven effective at simultaneously encoding both localized transient impulses and broader modulation patterns in machinery vibration signals13. Beyond the fault diagnosis domain, the synergy between deep learning architectures and metaheuristic optimization has attracted growing attention in broader engineering applications. Kumpf et al.14 proposed a two-level cooperative framework combining CNNs with boosting classifiers optimized by adapted metaheuristics for smart city IoT security, demonstrating that hybrid architectures can substantially outperform standalone deep models when the optimization pipeline is carefully designed. In a related direction, Jovanovic et al.15 introduced a two-tier framework integrating feature selection with XGBoost hyperparameter tuning via an improved firefly algorithm for phishing website detection, illustrating the value of joint feature-architecture optimization. Petrovic et al.16 applied metaheuristic-tuned CNNs to fault classification in direct current motors using Mel spectrograms, establishing that audio-domain feature representations coupled with optimized network configurations can achieve strong diagnostic accuracy in rotating machinery. Antonijevic et al.17 developed a CNN-enhanced attack detection framework with explainable artificial intelligence for IoT-based metaverse security, highlighting the importance of model interpretability in safety–critical applications. Furthermore, Jovanovic et al.18 demonstrated that attention-augmented recurrent architectures, when optimised by modified particle swarm optimisation, yield competitive performance in solar energy production forecasting—a finding that directly motivates the integration of attention mechanisms with residual networks pursued in the present study for vibration-based diagnosis. Despite these advances, several challenges persist when the target application shifts to ORC expanders specifically. First, ORC systems often operate under variable-speed and variable-load conditions driven by fluctuating heat source availability, which introduces significant non-stationarity into vibration measurements. Second, genuine fault samples from ORC expanders are scarce because these machines are relatively new in commercial service and catastrophic failures are rare, leading to severely imbalanced training datasets. Third, the coexistence of mechanical, thermodynamic, and fluid-induced noise sources in an ORC drivetrain creates a hostile signal environment where early-stage fault features can be easily masked19. Existing deep-learning approaches have not adequately addressed the interplay of these difficulties, often exhibiting degraded generalization and poor sensitivity to weak, incipient anomalies. Theoretical foundation Operating principles and fault mechanisms of ORC expanders An ORC system converts low-grade thermal energy into electricity through four sequential thermodynamic processes: isobaric evaporation in which the organic working fluid absorbs heat and transitions to superheated vapor, expansion across a work-producing device, isobaric condensation that rejects residual heat to a cooling medium, and pressurization by a feed pump that returns the liquid to the evaporator20. The expander sits at the heart of this loop—it is the sole component responsible for transforming the enthalpy drop of the vapor stream into rotational shaft power delivered to a generator. Two expander architectures dominate commercial ORC installations. Screw expanders employ a pair of helical rotors whose intermeshing profiles create sealed cavities that progressively enlarge along the axial direction, converting pressure energy into torque through volume expansion. Scroll expanders, by contrast, rely on an orbiting spiral element moving within a fixed scroll wrap; crescent-shaped gas pockets form at the periphery and migrate inward, shrinking continuously until discharge21. In both designs, bearings sustain radial and axial rotor loads, while lip seals or labyrinth seals minimize internal leakage. Prolonged thermal cycling and mechanical fatigue subject these components to non-uniform stress fields that accelerate localized material degradation. From classical thermodynamics, the isentropic efficiency of an expander is defined as the ratio of actual specific work output to the ideal isentropic enthalpy drop (a standard definition based on the first and second laws): where is the inlet specific enthalpy, is the actual outlet enthalpy, and denotes the outlet enthalpy under isentropic conditions. When the real expansion follows a polytropic path with index , the polytropic efficiency takes the well-known form: with representing the specific heat ratio of the working fluid. The net mechanical power transmitted through the drivetrain can then be expressed as22: where is the mass flow rate and accounts for friction and windage losses in bearings, seals, and couplings. Fault-induced changes manifest distinctly in vibration signals. Rotor imbalance generates a dominant spectral component at the shaft rotating frequency . Bearing inner-race defects produce a characteristic fault frequency given by the standard kinematic relation23: where is the number of rolling elements, and are the ball and pitch diameters, and is the contact angle. Outer-race faults follow an analogous expression with the sign reversed. Gear tooth wear introduces sideband modulation around the mesh frequency, whereas blade cracking excites transient impulses whose energy concentrates in high-frequency bands and broadens in time–frequency representations. These distinct spectral and temporal signatures constitute the physical basis upon which subsequent feature extraction and classification modules are constructed. Fundamentals of deep residual networks Convolutional neural networks extract local patterns from input data through learnable filter kernels, pooling operations, and nonlinear activations stacked in successive layers24. As network depth increases, however, two well-documented pathologies emerge: vanishing gradients, where error signals attenuate exponentially during back-propagation, and the degradation phenomenon, in which deeper plain networks paradoxically yield higher training error than their shallower counterparts25. These issues are not caused by overfitting but rather reflect the difficulty of optimizing highly nonlinear mappings through many composed layers. He et al. addressed this obstacle by reformulating the learning objective10. Instead of fitting a desired mapping directly, a residual block learns the residual function , so that the block output becomes: where denotes the trainable weight set and the identity shortcut passes gradient information unimpeded. A standard residual block comprises two convolutional layers with a ReLU activation interposed: For deeper architectures, the bottleneck variant reduces computational cost by sandwiching a narrow convolution between two projections25: Batch normalisation stabilises residual network training by normalising activations within each mini-batch using learnable affine parameters, as originally formulated by Ioffe and Szegedy26. The standard derivation is omitted here for brevity; what matters specifically for ORC expander signals is that the substantial amplitude variation across operating conditions can cause unnormalised activations to shift dramatically between speed settings, destabilising gradient flow in deeper layers. BN mitigates this internal covariate shift and permits higher learning rates, both of which prove essential for convergence on the relatively small dataset used in this study. where and are the mini-batch statistics, is a small constant for numerical stability, and , are learnable affine parameters26. During back-propagation, the gradient with respect to the normalized input propagates as: This normalization suppresses internal covariate shift, permitting higher learning rates and smoother loss landscapes, which is particularly beneficial when training on limited ORC fault data where optimization stability is paramount. ResNet has proven highly effective for vibration-based machinery fault diagnosis, yet certain limitations surface when it is applied to one-dimensional expander signals27. The fixed kernel size within each residual block constrains the receptive field to a single temporal scale, making it difficult to simultaneously capture both short-duration impulse transients and slowly varying modulation envelopes. Equally problematic, all feature channels receive identical weighting through standard convolution, which means noise-dominated channels contribute as much as fault-informative ones. These shortcomings motivate the attention-augmented, multi-scale architecture proposed in the present work. These observations directly informed the architectural choices made in the present study. Three parallel kernel sizes (8, 16, 32 samples) were chosen because they map onto the distinct temporal scales present in ORC expander vibrations: bearing defect impulses typically span fewer than 10 sampling points at 25.6 kHz, rotor-frequency modulation extends over roughly 15–20 points at typical operating speeds, and thermally induced envelope trends persist across 30 or more points. The CBAM module was preferred over simpler SE-only attention because ORC signals contain spatially localised transient events (blade cracks) whose temporal position carries diagnostic information—a dimension that channel-only attention is unable to capture. Attention mechanisms and multi-scale feature extraction theory Attention mechanisms enable neural networks to selectively emphasize informative components within feature representations. Three broad categories have gained prominence in fault diagnosis research: channel attention, spatial attention, and self-attention. The squeeze-and-excitation (SE) module, introduced by Hu et al., exemplifies channel attention by first aggregating spatial information through global average pooling and then producing per-channel excitation weights via two fully connected layers28: where denotes the feature at position , is the signal length, is the ReLU function, and is the sigmoid activation. The resulting vector rescales each channel according to its estimated importance. Spatial attention, by contrast, compresses the channel dimension and highlights critical temporal or spatial positions. The convolutional block attention module (CBAM) arranges these two mechanisms in series—channel attention followed by spatial attention—to refine features along both axes sequentially11. Given an intermediate feature map , CBAM first applies the channel descriptor and then computes a spatial attention map: where is the channel-refined output, denotes a convolution with kernel size , and represents concatenation along the channel axis. Self-attention, rooted in the Transformer paradigm, computes pairwise affinities across all positions through query-key-value projections12: This formulation captures long-range dependencies but carries quadratic computational cost, which can be prohibitive for lengthy vibration sequences. Multi-scale feature extraction addresses the limitation of fixed receptive fields discussed in "Fundamentals of deep residual networks" section. Parallel convolutional branches with distinct kernel sizes—for instance , , and —simultaneously capture fine-grained transient impulses and broader modulation envelopes within the same layer29. The outputs of these branches are fused through concatenation and a projection: where is the projection weight matrix. Several recent studies have demonstrated that multi-scale convolutional architectures outperform single-scale baselines on bearing and gearbox fault diagnosis benchmarks, particularly under variable-speed operating conditions where fault-characteristic frequencies shift continuously30. The complementary strengths of attention-based channel recalibration and multi-scale receptive field diversity form the theoretical cornerstone of the improved residual architecture proposed in “Experimental verification” section. Construction of the deep residual network Overall model framework design Building on the theoretical foundations established in “Theoretical foundation” section, this paper puts forward an improved deep residual network—termed MS-CBAM-ResNet—specifically tailored to the fault diagnosis of ORC expander vibration signals. The architecture addresses three shortcomings of standard ResNet identified earlier: the rigid single-scale receptive field, the indiscriminate treatment of channel features, and the lack of spatial selectivity. Figure 1 illustrates the end-to-end processing pipeline from raw one-dimensional vibration input to fault category prediction. Overall architecture of the proposed MS-CBAM-ResNet model for ORC expander fault diagnosis. The model consists of five sequential stages. First, an input preprocessing layer performs z-score normalization on the raw vibration signal and segments it into fixed-length samples of 2048 points with 50% overlap, a windowing strategy commonly adopted for non-stationary rotating machinery signals31. Second, the normalized segments enter a multi-scale parallel feature extraction module comprising three convolutional branches with kernel sizes of , , and , whose outputs are concatenated along the channel dimension. This front-end module, detailed further in "Multi-scale parallel feature extraction module design" section, captures both localized impulse transients and broader periodic modulation patterns before any residual learning takes place. Third, the concatenated multi-scale features flow into the attention-enhanced residual backbone, which stacks four groups of modified residual blocks. Each block embeds a CBAM unit between the second convolutional layer and the identity shortcut addition, allowing the network to recalibrate channel and spatial responses at every depth level11. Batch normalization and ReLU activation follow each convolution within the block, consistent with the pre-activation residual design that has shown favorable gradient behavior in deep architectures25. The number of feature channels doubles at each group transition—from 64 to 128, 256, and finally 512—while strided convolutions halve the temporal dimension accordingly. Fourth, a global average pooling layer collapses the final 512-channel feature map into a compact descriptor vector, eliminating the need for fully connected hidden layers and thereby reducing overfitting risk on small ORC datasets. Fifth, a Softmax classifier maps this descriptor to a probability distribution over the predefined fault categories. Table 1 summarises the key hyperparameter settings of the proposed model. The network depth was set at 18 convolutional layers (four residual groups with two blocks each, plus the initial multi-scale module), a configuration that was found to strike a good balance between representational capacity and the limited sample volume typical of ORC field data. The initial learning rate follows a cosine annealing schedule starting from 0.001, and dropout with a rate of 0.3 is applied before the Softmax layer to further regularise training32. Table 1Hyperparameter configuration of the MS-CBAM-ResNet model.HyperparameterValueSelection basisRemarksInput segment length2048 pointsNyquist criterion at 10 kHz sampling rate50% overlap between segmentsMulti-scale kernel sizes7, 15, 31Cover fault impulse to rotational period scalesThree parallel branchesResidual backbone depth18 layers (4 groups × 2 blocks)Trade-off between capacity and sample sizeChannel doubling per groupInitial learning rate0.001Cosine annealing scheduleMinimum rate 1 × 10Dropout rate0.3Regularization for small-sample scenariosApplied before Softmax layer Hyperparameter configuration of the MS-CBAM-ResNet model. As presented in Table 1, the kernel size triplet of 7, 15, and 31 was chosen so that the smallest kernel targets high-frequency bearing defect impulses spanning roughly 0.7 ms at a 10 kHz sampling rate, while the largest kernel encompasses nearly a full shaft revolution period, capturing rotor imbalance and low-frequency modulation signatures. Compared with a vanilla ResNet-18 baseline, MS-CBAM-ResNet introduces only a modest increase in parameter count—approximately 8%—because the CBAM modules and the projection after multi-scale concatenation are lightweight by design33. This architectural efficiency is particularly relevant for ORC monitoring systems deployed on embedded industrial hardware where computational budgets are constrained. Multi-scale parallel feature extraction module design ORC expander vibration signals carry fault information across disparate temporal scales—bearing defect impulses span only a few sampling points, rotor-related periodic modulation extends over tens of points, and thermal drift or load-induced trends persist across even longer windows. A single fixed-size kernel, no matter how carefully tuned, inevitably privileges one scale at the expense of others. This observation motivated the design of the multi-scale feature extraction module (MSFEM), which sits immediately after the input preprocessing layer and feeds its output into the attention-enhanced residual backbone described in "Overall model framework design" section. MSFEM splits the normalized input signal into three parallel one-dimensional convolutional branches, each configured with a different kernel size: , , and . The small-kernel branch targets short-duration impact features characteristic of localized bearing faults, where energy concentrates in sharp transient bursts lasting less than one millisecond at typical ORC sampling rates. The medium-kernel branch is designed to capture periodic amplitude modulation associated with meshing or rotational-frequency components. The large-kernel branch, meanwhile, spans a receptive field broad enough to encode low-frequency envelope trends and slowly varying operational signatures34. Each branch follows an identical internal pipeline—a one-dimensional convolution, batch normalization, ReLU activation, and max pooling—ensuring that all three branches produce feature maps of the same temporal length despite differing kernel widths. Stride and padding parameters are adjusted per branch to guarantee dimensional consistency, a prerequisite for the subsequent fusion step. After the three branches process the input independently, their feature maps are concatenated along the channel dimension. This concatenation preserves scale-specific representations without forcing premature information mixing, a strategy that has shown clear advantages over element-wise addition in multi-scale fault diagnosis architectures35. However, channel concatenation triples the feature dimensionality relative to any single branch, which would impose a heavy computational burden on the downstream residual blocks. To remedy this, a convolutional layer immediately after concatenation to compress the channel count back to 64, the entry width expected by the first residual group. This pointwise projection not only reduces parameter overhead but also introduces a learned linear combination across scales, allowing the network to weight each branch’s contribution adaptively29. Table 2 summarizes the structural parameters of the three convolutional branches and the subsequent channel compression layer. As presented in Table 2, the output channel count of each branch is set to 32, yielding 96 channels after concatenation, which the convolution then reduces to 64. Table 2Structural parameters of the multi-scale convolutional branches in MSFEM.BranchKernel sizeOutput channelsStride/paddingMax pool sizeSmall-scale321/same2Medium-scale321/same2Large-scale321/same2Channel compression641/0– Structural parameters of the multi-scale convolutional branches in MSFEM. Alternative kernel combinations were also evaluated during preliminary experiments—including a two-branch variant () and a four-branch variant (). The two-branch scheme yielded noticeably weaker recognition of medium-frequency gear-related faults, while the four-branch arrangement provided only marginal accuracy gains at the cost of a 35% increase in floating-point operations. The three-branch configuration struck the most favorable balance between diagnostic coverage and computational economy, consistent with findings reported for other rotating machinery applications36. At the module’s output boundary, the 64-channel compressed feature map passes through a second batch normalization layer that standardizes the activation distribution before entering the first CBAM-enhanced residual block. This interface design ensures smooth gradient flow across the transition from multi-scale extraction to residual learning, preventing scale mismatch or distribution shift from destabilizing early training epochs37. Attention-enhanced residual module design and model optimization strategies The core building block of the MS-CBAM-ResNet backbone is the attention-enhanced residual block (AER-Block), which inserts a CBAM unit into the main path of a standard residual block so that feature recalibration occurs before the identity shortcut addition. Figure 2 depicts the internal structure of one AER-Block. Schematic of the attention-enhanced residual block (AER-Block) showing the integration of CBAM with the residual shortcut. As illustrated in Fig. 2, given an input feature map , the block first applies two successive convolutional layers with batch normalization and ReLU activation: The CBAM module then refines sequentially along the channel and spatial dimensions. The channel attention sub-module aggregates spatial information through both global average pooling and global max pooling, producing two compact descriptors and 11: These descriptors pass through a shared two-layer perceptron with a reduction ratio , and their outputs are summed before sigmoid gating: where and are shared weights, and throughout the proposed network. The channel-refined feature is obtained as: The spatial attention sub-module then operates on by computing channel-wise average and maximum values at each temporal position, concatenating them, and convolving with a kernel11: The final AER-Block output combines the attention-refined feature with the identity shortcut: This formulation ensures that gradient signals propagate through the shortcut even when attention gates suppress certain channels or positions, preserving the training stability inherent to residual learning. To handle the severe class imbalance characteristic of ORC fault datasets—where normal-state samples vastly outnumber minority fault categories—the standard cross-entropy objective is replaced with Focal Loss38: Here is the predicted probability for the true class, is a class-balancing weight, and controls down-weighting of well-classified examples. The learning rate follows a cosine annealing schedule (a standard formulation): with , , and denoting total training epochs. This schedule permits aggressive early optimization while fine-tuning parameters gently in later stages39. Given the scarcity of genuine ORC expander fault samples, three data augmentation techniques are employed during training: random cropping that extracts sub-segments of length 1792 from each 2048-point window, additive Gaussian noise injection to simulate sensor interference, and temporal flipping that reverses the signal sequence to double the effective dataset without altering spectral content40. Table 3 lists the specific parameter settings for each augmentation strategy. Table 3Data augmentation strategy parameter settings.Augmentation methodKey parameterValueRandom croppingCrop length1792 points (87.5% of original)Gaussian noise injectionSignal-to-noise ratio20 dB–40 dB (uniform random)Temporal flippingApplication probability0.5 per sample Data augmentation strategy parameter settings. As presented in Table 3, the noise injection range of 20–40 dB was selected to reflect realistic sensor noise levels observed in ORC test rigs while avoiding excessive corruption that could obscure weak fault signatures. Together, these optimization strategies—Focal Loss, cosine annealing, and targeted augmentation—form an integrated training scheme that enables the MS-CBAM-ResNet to converge reliably even when fewer than 100 samples per fault class are available. Experimental verification Experimental platform setup and dataset construction The experimental platform centers on a single-screw expander rated at 10 kW, driven by an ORC loop that circulates R245fa as the organic working fluid. The loop comprises four principal subsystems: an electric heating module that raises the thermal oil temperature to simulate waste heat sources between 80 and 120 °C at the evaporator outlet, the screw expander coupled to a magnetic powder brake serving as an adjustable resistive load, a plate-type condenser with a controlled cooling water circuit, and a hermetic diaphragm pump that returns subcooled liquid to the evaporator41. Shaft rotational speed is regulated between 1200 and 2400 rpm by varying the heating power and load torque simultaneously, creating a range of realistic operating conditions. Vibration data were acquired through a set of piezoelectric accelerometers (sensitivity 100 mV/g) mounted at three orthogonal positions on the drive-end bearing housing—horizontal, vertical, and axial—following the guidelines recommended in ISO 10816 for rotating machinery monitoring42. All channels were sampled synchronously at 25.6 kHz using a 16-bit data acquisition system, with each recording session lasting 60 s per operating condition to ensure statistical sufficiency. Four health states were investigated: normal operation, outer-race bearing defect, rotor mass imbalance, and blade root crack. The bearing fault was introduced by electrical discharge machining of a 0.3 mm wide groove on the outer race, a technique widely adopted in bearing diagnostics research43. Rotor imbalance was created by attaching a 5 g trial mass at a 120° angular position on the coupling disk. Blade cracking was simulated through wire-cut notches of 2 mm depth at the root of a selected rotor lobe. Each fault was tested independently across three speed settings (1200, 1800, 2400 rpm) and two load levels, yielding 12 distinct operating points per fault category. Raw signals underwent a preprocessing pipeline consisting of a 10–10,000 Hz bandpass Butterworth filter to remove DC offset and high-frequency aliasing artifacts, followed by z-score normalization. A sliding window of 2048 points with 50% overlap segmented the continuous recordings into individual samples. Table 4 summarizes the resulting dataset statistics. Table 4Statistical summary of the experimental dataset.Health stateNo. of samplesTraining set (60%)Validation set (20%)Test set (20%)Normal1200720240240Bearing outer-race fault960576192192Rotor imbalance960576192192Blade crack840504168168Total39602376792792Bold values indicate column totals. Statistical summary of the experimental dataset. Bold values indicate column totals. As presented in Table 4, the total dataset comprises 3960 samples distributed across four classes with a moderate imbalance ratio of approximately 1.43:1 between the largest and smallest categories. The 60/20/20 stratified split preserves class proportions across all three subsets, and no sample from the same continuous recording segment appears in more than one subset to prevent data leakage44. Figure 3 displays representative time-domain waveforms for the four health states recorded at 1800 rpm under medium load. The normal signal exhibits low-amplitude quasi-stationary fluctuations, whereas the bearing fault introduces repetitive high-frequency impulse bursts at intervals consistent with the outer-race characteristic frequency. Rotor imbalance manifests as a dominant sinusoidal modulation at the shaft frequency, and the blade crack produces irregular transient spikes superimposed on a slightly elevated baseline vibration level. Time-domain vibration waveform comparison for four health states at 1800 rpm. Figure 4 presents the corresponding frequency spectra obtained via 4096-point FFT. The bearing fault spectrum shows distinct harmonic peaks at multiples of the ball-pass frequency outer race, while the imbalance condition concentrates energy sharply at 30 Hz (the shaft frequency at 1800 rpm). The blade crack spectrum, by contrast, elevates broadband energy in the 2000–5000 Hz range without producing clearly isolated peaks—a characteristic that makes this fault particularly challenging to detect through conventional spectral analysis alone. Frequency spectrum comparison for four health states at 1800 rpm. To quantify the noise contamination across different operating conditions, the signal-to-noise ratio (SNR) of each sample was computed by treating the fault-related periodic or transient component as signal and the residual as noise. Figure 5 shows the SNR distribution across all samples grouped by fault type. Normal-state recordings cluster around 15–20 dB, bearing faults range from 8 to 18 dB depending on speed, and blade crack samples concentrate below 10 dB—a finding that confirms this fault mode produces the weakest discriminative signature relative to background noise. Signal-to-noise ratio distribution across fault categories and operating conditions. Figure 6 depicts the class-wise sample distribution, visually reinforcing the imbalance noted in Table 4. The blade crack category contains roughly 30% fewer samples than the normal class, motivating the Focal Loss and augmentation strategies described in "Attention-enhanced residual module design and model optimization strategies" section. Sample quantity distribution across four fault categories in the dataset. Ablation experiment analysis To isolate the contribution of each proposed module, a systematic ablation study was conducted with four model variants trained under identical conditions: the same dataset split from Table 4, Focal Loss objective, cosine annealing schedule, and 150 training epochs. The four variants are: (a) Base-ResNet—a standard 18-layer residual network with single-scale input convolution; (b) MS-ResNet—Base-ResNet preceded by the MSFEM module from "Multi-scale parallel feature extraction module design" section but without any attention mechanism; (c) CBAM-ResNet—Base-ResNet with CBAM inserted into each residual block but retaining the single-scale front end; and (d) the complete MS-CBAM-ResNet. Diagnostic performance was evaluated using four standard metrics (standard definitions in multi-class classification): All metrics are reported as macro-averages across the four fault classes. To quantify robustness under noise corruption, a noise degradation index is further defined to measure the relative accuracy drop when test samples are contaminated with additive white Gaussian noise at a specified SNR level relative to the clean-signal baseline45: A smaller indicates stronger resilience to noise interference. Table 5 summarizes the test-set results for all four variants under clean conditions and at two noise levels (10 dB and 0 dB SNR). Table 5Diagnostic performance comparison of ablation variants.Model variantAccuracy (%)Precision (%)Recall (%)F1 (%) at 10 dB (%)Base-ResNet91.5490.8791.0290.9412.36MS-ResNet95.0894.6394.9194.777.82CBAM-ResNet94.3293.9594.1894.068.45MS-CBAM-ResNet98.1197.8697.9497.903.27 Diagnostic performance comparison of ablation variants. The results in Table 5 reveal several noteworthy patterns. Introducing the multi-scale module alone (MS-ResNet) raises accuracy by 3.54 percentage points over the baseline, which confirms that parallel kernels of different sizes capture complementary fault signatures that a single-scale front end misses—particularly for the blade crack category, whose broadband energy distribution benefits from the large-kernel branch. Adding CBAM without multi-scale input (CBAM-ResNet) yields a somewhat smaller gain of 2.78 points, suggesting that attention recalibration is most effective when it operates on already diversified feature representations rather than on a homogeneous single-scale map. The full MS-CBAM-ResNet achieves 98.11% accuracy, outperforming both partial variants by a clear margin. What is perhaps more telling is the noise robustness column: drops from 12.36% for Base-ResNet to just 3.27% for the complete model, meaning that diagnostic accuracy degrades by only about 3% even at 10 dB SNR. This synergistic effect likely arises because multi-scale convolution distributes fault energy across multiple representational channels, and the CBAM module then selectively amplifies those channels while suppressing noise-dominated ones—a complementary interaction that neither module can fully realize on its own46. Figure 7 plots the training loss curves of the four variants over 150 epochs. Base-ResNet converges slowest and exhibits noticeable oscillation after epoch 80, indicative of the optimization difficulty caused by a narrow representational bottleneck. MS-ResNet and CBAM-ResNet both converge faster and reach lower plateau values, though CBAM-ResNet shows slightly sharper initial descent—an observation consistent with the regularizing effect of attention gating on gradient flow47. The complete MS-CBAM-ResNet achieves the lowest terminal loss and the smoothest convergence trajectory, reaching near-plateau performance by approximately epoch 60, roughly 30 epochs earlier than the baseline. Training loss convergence curves for the four ablation variants over 150 epochs. As depicted in Fig. 7, the computational overhead introduced by the combined modules remains modest. MS-CBAM-ResNet requires approximately 1.12 million parameters compared with 1.04 million for Base-ResNet—an increase of roughly 8%—while inference time per sample rises by only 0.3 ms on an NVIDIA RTX 3090 GPU. This favorable efficiency profile confirms that the performance gains documented above are not simply a consequence of increased model capacity, but rather reflect genuinely improved feature selectivity enabled by the multi-scale and attention mechanisms working in concert. To isolate the contribution of each training strategy component—and to address the question of whether Focal Loss is warranted at the mild 1.43:1 imbalance ratio—an additional ablation experiment was carried out with the architecture held constant at full MS-CBAM-ResNet while the training pipeline was varied. Table 6 reports results averaged over five random seeds. Replacing Focal Loss with standard cross-entropy reduces overall accuracy by 1.83 percentage points and, more critically, drops blade crack recall from 96.8 to 91.2%; this confirms that even at moderate imbalance ratios, Focal Loss provides meaningful benefit for minority classes whose samples occupy a narrower feature volume. Removing data augmentation causes a more severe 3.41-point accuracy decline, underscoring the importance of synthetic sample diversity when training data are scarce. Disabling cosine annealing in favour of a fixed learning rate degrades accuracy by 1.12 points, which is consistent with the known benefit of learning rate warm restarts in promoting escape from sharp local minima39. Table 6Training strategy ablation results (MS-CBAM-ResNet architecture held constant, five-seed average).Training configurationAccuracy (%)Blade crack recall (%)F1 (%)Full pipeline (Focal Loss + Augmentation + Cosine Annealing)98.11 ± 0.3496.8 ± 0.5297.90 ± 0.31Cross-entropy replacing Focal Loss96.28 ± 0.4891.2 ± 1.0595.87 ± 0.44No data augmentation94.70 ± 0.6788.5 ± 1.3294.12 ± 0.59Fixed learning rate (no cosine annealing)96.99 ± 0.4195.1 ± 0.7396.68 ± 0.38No augmentation + cross-entropy92.15 ± 0.8984.3 ± 1.6791.42 ± 0.81 Training strategy ablation results (MS-CBAM-ResNet architecture held constant, five-seed average). Comparative experiments and performance evaluation Six representative fault diagnosis methods were selected as benchmarks: SVM with radial basis function kernel and hand-crafted statistical features, a standard five-layer 1D-CNN, ResNet-18, a two-layer LSTM network, SENet-ResNet (ResNet-18 with SE channel attention), and a Transformer encoder with four self-attention heads. To ensure a fair comparison, each deep learning baseline underwent independent hyperparameter tuning via grid search over learning rate, batch size, and architecture-specific parameters (e.g., number of LSTM hidden units, number of Transformer heads), selecting the configuration that maximized validation accuracy. This avoids the potential confound of attributing performance gaps to training pipeline differences rather than architectural merit. Evaluation was conducted on both the ORC expander dataset described in "Experimental platform setup and dataset construction" section and the publicly available Case Western Reserve University (CWRU) bearing dataset to assess cross-domain generalization48. Table 7 presents the comprehensive results on the ORC expander test set. The macro-averaged precision, recall, and F1 score were computed following Eqs. (24)–(25), while the area under the receiver operating characteristic curve (AUC) was obtained via one-versus-rest averaging across all four classes (standard multi-class AUC formulation): where denotes the number of fault categories. Table 7Comprehensive performance comparison of all methods on the ORC expander dataset.MethodAccuracy (%)Precision (%)Recall (%)F1 (%)AUCParams (M)Inference time (ms)SVM82.4581.7380.9681.340.912—0.81D-CNN88.7688.1287.9488.030.9480.381.1ResNet-1891.5490.8791.0290.940.9671.041.5LSTM89.5288.9689.2389.090.9530.823.7SENet-ResNet94.9594.5194.7294.610.9811.071.7Transformer93.1892.8593.0492.940.9762.464.2MS-CBAM-ResNet98.1197.8697.9497.900.9981.121.8Bold values indicate the proposed method (MS-CBAM-ResNet). Comprehensive performance comparison of all methods on the ORC expander dataset. Bold values indicate the proposed method (MS-CBAM-ResNet). The results in Table 7 reveal that MS-CBAM-ResNet surpasses all competitors across every diagnostic metric. SVM suffers most on the blade crack category—its recall for this class drops below 72%—because hand-crafted time-domain and frequency-domain features fail to capture the broadband, non-stationary nature of crack-induced vibrations. The standard CNN and LSTM improve upon SVM substantially but still fall short of 90% accuracy, limited by shallow representational depth and fixed temporal modeling scope, respectively. ResNet-18 reaches 91.54% yet struggles with the blade crack class for reasons discussed in "Ablation experiment analysis" section. SENet-ResNet narrows the gap considerably by introducing channel recalibration, though the absence of multi-scale input constrains its ability to simultaneously resolve high-frequency impulses and low-frequency modulation. The Transformer achieves competitive accuracy at 93.18% but demands 2.46 million parameters and 4.2 ms per sample—roughly 2.3 times the inference cost of the proposed model—which makes it less practical for real-time ORC monitoring49. Figure 8 presents the ROC curves of all seven methods on the ORC test set. MS-CBAM-ResNet traces a curve that hugs the upper-left corner most tightly, with an AUC of 0.998, confirming near-perfect class separability. SENet-ResNet and Transformer cluster closely behind, while SVM and 1D-CNN exhibit noticeably earlier departure from the ideal boundary, particularly at low false-positive rates where early fault detection matters most. ROC curve comparison of all methods on the ORC expander test set. To verify that the observed performance differences are statistically meaningful rather than artefacts of a particular data split, all experiments were repeated using five-fold stratified cross-validation with five different random seeds, and results are reported as mean ± standard deviation across 25 total runs. MS-CBAM-ResNet achieved 98.11 ± 0.34% accuracy, whereas the closest competitor (SENet-ResNet) reached 94.95 ± 0.51%. A Wilcoxon signed-rank test between the per-fold accuracy vectors of MS-CBAM-ResNet and each competitor yielded p < 0.001 in all pairwise comparisons, which decisively confirms a statistically significant superiority. The McNemar test was additionally applied on the single best-split predictions: for every pair of classifiers producing discordant predictions, all pairwise comparisons yielded p < 0.001, corroborating the cross-validation findings50. Under the null hypothesis of equal error rates, follows a chi-squared distribution with one degree of freedom. All pairwise comparisons yielded (), decisively rejecting the null hypothesis and confirming statistically significant superiority of the proposed model. A further examination was carried out of how each deep learning method degrades under limited training data, by randomly subsampling 10, 30, and 50 instances per class and averaging results over five repetitions. At 10 samples per class, MS-CBAM-ResNet maintained 87.3% accuracy—a remarkable figure given the extreme data scarcity—while ResNet-18 dropped to 71.2% and the Transformer collapsed to 64.8%, most likely because of overfitting its large parameter space. The small-sample resilience of the proposed model can be attributed to the Focal Loss rebalancing, targeted augmentation strategies, and the inductive bias provided by multi-scale convolution, which constrains the hypothesis space more effectively than generic architectures. On the CWRU bearing dataset, the standard 10-class protocol was followed: fault diameters of 0.007, 0.014, and 0.021 inches for inner race, outer race, and ball faults plus the normal class, sampled at 12 kHz from the drive-end accelerometer under 1 HP load48. Signals were segmented into 2048-point windows with 50% overlap, yielding approximately 400 samples per class. The same five-fold cross-validation procedure was applied. MS-CBAM-ResNet achieved 99.46 ± 0.12% accuracy, which confirms that the architectural improvements transfer well beyond the ORC-specific domain; it should however be noted that achieving above 99% on CWRU is now commonplace, and the primary value of this experiment lies in validating transferability rather than in claiming state-of-the-art performance on this benchmark. Direct transfer from CWRU to the ORC dataset without fine-tuning yielded only 72.3% accuracy, which reflects the substantial distributional gap between laboratory bearing rigs and industrial expander environments. This gap arises from differences in working fluid coupling, thermal noise signatures, and shaft speed regimes—factors absent in the CWRU setup. Discussion The experimental results presented in “Discussion” section warrant a closer examination of why the proposed MS-CBAM-ResNet achieves such consistent gains over existing approaches. The performance improvement appears to stem from a synergy across three levels: signal representation, network architecture, and training strategy. At the signal processing level, ORC expander vibrations carry fault information distributed unevenly across temporal scales—bearing defect impulses are sharp and localized, while rotor imbalance manifests as sustained periodic modulation. The multi-scale parallel convolution front end addresses this heterogeneity directly. Rather than forcing a single kernel to compromise between transient sensitivity and periodic resolution, the three-branch MSFEM lets each kernel specialize and defers the fusion decision to a learned projection layer. The ablation results in Table 5 confirm that this design alone accounts for a 3.54-point accuracy gain, with the largest improvement observed on the blade crack category, whose broadband spectral signature demands wide receptive field coverage. At the architectural level, the CBAM attention module provides a complementary benefit that becomes most potent when operating on the already diversified multi-scale feature maps. Channel attention suppresses noise-dominated channels—a particularly valuable property in ORC systems where fluid pulsation and thermal-induced vibration contaminate certain frequency bands. Spatial attention, in turn, highlights temporal positions where fault-related transients concentrate, effectively sharpening the network’s focus on diagnostically relevant signal segments. The fact that CBAM-ResNet alone improves accuracy by only 2.78 points while the combined MS-CBAM-ResNet achieves a 6.57-point gain over the baseline suggests a multiplicative interaction: attention is far more effective when it has richer, multi-scale features to select from. The training optimisation strategy deserves separate acknowledgement. Focal Loss proved essential for preventing the model from ignoring the minority blade crack class, and the three augmentation techniques—random cropping, Gaussian noise injection, and temporal flipping—expanded the effective training distribution enough to sustain 87.3% accuracy even at 10 samples per class. Without these measures, severe overfitting was observed within 30 epochs during preliminary trials. To provide deeper insight into the learned representations, the CBAM attention maps were visualised for representative test samples drawn from each fault category. Figure 9 presents the channel attention weights averaged across all test samples per class. The bearing outer-race fault activates channels associated with high-frequency impulse features most strongly, whereas the rotor imbalance condition concentrates attention on channels encoding the shaft rotational frequency band. The blade crack category, by contrast, distributes attention more evenly across mid- and high-frequency channels, consistent with its broadband spectral signature. Normal-state samples yield nearly uniform channel weights, which suggests that no single frequency band dominates the healthy vibration pattern. These observations confirm that the CBAM module learns physically interpretable channel selectivity rather than arbitrary weighting. Channel attention weight distribution across fault categories, averaged over all test samples. Figure 10 displays the intermediate feature maps extracted from the third residual group for one sample of each fault type. The bearing fault sample shows sharply localised activation peaks at temporal positions corresponding to impact events, while rotor imbalance produces smooth sinusoidal activation patterns aligned with the shaft frequency. Blade crack features exhibit irregular, aperiodic activation spikes—a pattern that explains why single-scale architectures struggle with this fault category, since the irregularity demands broad receptive field coverage to capture the full transient envelope. These feature-level visualisations complement the attention maps and strengthen the case for the engineering interpretability of the proposed model, although it must be admitted that they fall short of formal causal explanations. Intermediate feature maps from the third residual group for representative samples of each fault type. Several practical challenges remain, however. The 1.8 ms inference latency per sample, while acceptable for most monitoring intervals, may become a bottleneck in applications requiring sub-millisecond response times on edge computing hardware. Sensor installation on compact ORC expanders also raises cost and accessibility concerns, particularly in retrofit scenarios where bearing housings offer limited mounting surfaces. Perhaps the most significant limitation is model interpretability—although attention maps provide some visual insight into which channels and positions the network prioritizes, they fall short of the causal explanations that maintenance engineers typically expect. As detailed in "Comparative experiments and performance evaluation" section, the CWRU transfer experiment (99.46% accuracy) confirmed cross-domain generalizability, while the 72.3% direct-transfer result from CWRU to ORC data highlights the need for domain adaptation techniques in future work. From a practical deployment perspective, the proposed framework offers several advantages for industrial ORC condition monitoring. The 1.8 ms inference latency per sample is well within the typical 100 ms monitoring interval used in commercial SCADA systems, permitting real-time diagnosis without dedicated GPU hardware—a standard CPU suffices for single-stream operation. The modest parameter footprint of 1.12 million weights translates to approximately 4.5 MB of storage, comfortably fitting on edge computing devices commonly deployed at distributed ORC installations. Sensor integration requires only a single triaxial accelerometer mounted on the drive-end bearing housing, which is a standard configuration in vibration monitoring practice and does not demand invasive modification of the expander enclosure. It should also be acknowledged that all faults investigated in this study are static, full-severity, single-point artificial defects introduced under controlled laboratory conditions. Real-world ORC failures typically evolve gradually through incipient, developing, and severe stages, and compound faults involving simultaneous degradation of multiple components are not uncommon. The current experimental design does not capture these progressive or compound scenarios, which limits the validity of any claim regarding suitability for industrial predictive maintenance. Future experiments should incorporate run-to-failure tests and multi-fault injection protocols to evaluate the model under more realistic degradation trajectories. Beyond this, the experimental dataset originates from a single 10 kW screw expander; validation on scroll expanders and radial-inflow turbines therefore remains an important open task before broader generalisation claims can be substantiated. Looking ahead, extending the framework to other rotating machinery—compressors, gas turbines, wind turbine drivetrains—appears straightforward, given that the multi-scale and attention mechanisms are agnostic to the specific fault physics. Two research directions appear particularly promising: graph neural networks for modelling spatial relationships among multiple sensors distributed across a machine, and federated learning frameworks that would allow multiple ORC plant operators to collaboratively train diagnostic models without sharing proprietary operational data. Conclusion This paper addressed the fault diagnosis problem of ORC drivetrain expanders by proposing MS-CBAM-ResNet, an improved deep residual network that integrates multi-scale parallel feature extraction with convolutional block attention mechanisms. The research encompassed four tightly connected tasks: construction of an experimental platform and a labeled vibration dataset covering four health states of a screw expander, design of a three-branch multi-scale convolution front end (MSFEM) tailored to the heterogeneous temporal characteristics of expander vibration signals, development of attention-enhanced residual blocks (AER-Blocks) that embed channel and spatial recalibration into the residual learning path, and formulation of a training optimization scheme combining Focal Loss, cosine annealing, and targeted data augmentation for small-sample imbalanced scenarios. The principal findings can be distilled into three points. First, MS-CBAM-ResNet achieved 98.11% overall accuracy on the ORC expander test set—a 6.57-percentage-point improvement over the standard ResNet-18 baseline—with particularly strong gains on the blade crack category, which conventional single-scale architectures consistently misclassified. Second, the model demonstrated notable robustness under noise-degraded conditions, sustaining only a 3.27% relative accuracy drop at 10 dB SNR compared with 12.36% for the baseline, a difference attributable to the CBAM module’s capacity to suppress noise-dominated feature channels. Third, under extreme data scarcity of just 10 labeled samples per class, the proposed model retained 87.3% accuracy, outperforming all six competing methods by substantial margins, which confirms the effectiveness of the joint augmentation and Focal Loss strategy in regularizing training on limited ORC field data. From a theoretical standpoint, this work contributes a systematic framework for fusing attention mechanisms and multi-scale feature extraction within residual networks for rotating machinery diagnostics. The ablation study revealed that the two modules interact multiplicatively rather than additively—attention recalibration is far more effective when applied to diversified multi-scale representations—an insight that may guide architectural decisions in related domains. On the practical side, the framework offers a deployable solution for intelligent condition monitoring and predictive maintenance of ORC waste heat recovery systems, with an inference cost of only 1.8 ms per sample and a modest parameter footprint of 1.12 million weights. Several limitations should be recognised that temper the scope of these conclusions. The experimental data originate from a single 10 kW screw expander test rig under three discrete speed settings; validation on scroll expanders, radial turbines, and continuously varying speed-load trajectories therefore remains pending. The four fault categories do not exhaust all possible degradation modes—coupling misalignment, working fluid contamination, valve leakage, and compound faults all deserve future investigation. All faults are artificially seeded at full severity, and the model’s sensitivity to incipient, gradually developing defects has not yet been evaluated. The dataset comprises 3960 samples from a single author-controlled test rig, and external validation on an independent ORC installation will be needed to confirm generalisation beyond this specific setup. Future research will pursue five directions. First, future work will incorporate physics-informed neural network principles that embed thermodynamic and rotor-dynamic governing equations into the loss function, which should enhance model interpretability and physical consistency. Second, self-supervised pretraining on large volumes of unlabelled ORC operational data could further alleviate the labelled-sample bottleneck that constrains industrial deployment. Third, multi-sensor information fusion combined with edge computing optimisation will be explored, to enable real-time, on-site diagnosis in distributed ORC installations where cloud connectivity cannot be guaranteed. Fourth, domain adaptation and transfer learning techniques should be investigated to bridge the distributional gap between laboratory and field data, potentially enabling models trained on one expander type to generalise to others with minimal fine-tuning. Fifth, run-to-failure experiments and compound fault scenarios must be incorporated to validate the framework under progressive degradation conditions representative of actual industrial operation. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements Not applicable. Abbreviations Organic Rankine cycle Multi-scale convolutional block attention module residual network Multi-scale feature extraction module Attention-enhanced residual block Convolutional block attention module Deep residual network Convolutional neural network One-dimensional convolutional neural network Long short-term memory Squeeze-and-excitation Support vector machine Batch normalization Rectified linear unit Fast Fourier transform Signal-to-noise ratio Receiver operating characteristic Area under the curve Case Western Reserve University Ball pass frequency inner race Physics-informed neural network True positive False positive False negative Author contributions Jing Li conceptualized the study, designed the network architecture, developed the methodology, supervised the research, and wrote the original manuscript. Zexu Gao contributed to the experimental platform setup, data acquisition and preprocessing, conducted the comparative and ablation experiments, and assisted with result analysis and manuscript revision. All authors read and approved the final manuscript. Funding This work is supported by the Natural Science Foundation of China (Grant No. 61873025). Data availability All data generated and analysed during the current study, along with the complete Python implementation source code, have been uploaded to the journal’s Supplementary Material section as accompanying files with this manuscript. The archive includes scripts for data preprocessing, model training, ablation experiments, and visualisation. Interested researchers may also contact the corresponding author for further details. Declarations Ethics approval and consent to participate Not applicable. This study involves computational modeling and mechanical vibration signal analysis only. No human or animal subjects were involved in this research. Consent for publication All authors have reviewed the manuscript and consent to its publication. No identifiable information regarding participants has been included. Competing interests The authors declare no competing interests. Clinical trial number Not applicable. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement All data generated and analysed during the current study, along with the complete Python implementation source code, have been uploaded to the journal’s Supplementary Material section as accompanying files with this manuscript. The archive includes scripts for data preprocessing, model training, ablation experiments, and visualisation. Interested researchers may also contact the corresponding author for further details.