Feature decoupling and cross domain alignment with transfer learning for cross working condition mechanical fault diagnosis Authors: Honglian Xiao, Jianhua Xiao Published: 2026 ABSTRACT AbstractAiming at the problems of poor adaptability of heterogeneous sensing signals and feature distribution shift caused by working condition changes in cross-working-condi tion mechanical fault diagnosis, a WDH-Net model based on transfer learning is proposed for cross-working-condition heterogeneous mechanical fault diagnosis. The model integrates three core modules: time-frequency feature enhancement, dual-stream decoupling encoding and heterogeneous feature alignment. Specifically, the Wavelet-ConvNet+Cross-Attention module is used to perform adaptive wavelet packet decomposition and cross-attention feature optimization on heterogeneous time-series signals such as vibration and current to enhance the extraction of fault-sensitive features; the dual-stream decoupling encoder is adopted to separate fault features from working condition features to reduce working condition interference; a heterogeneous common feature mapping space is constructed, and the heterogeneous multi-kernel maximum mean discrepancy algorithm is combined to achieve cross-domain feature alignment and improve the generalization ability of the model. Experimental results on the public CWRU and PU datasets show that WDH-Net exhibits excellent performance in cross-working-condition fault diagnosis tasks, providing a new methodological reference for the intelligent fault diagnosis of industrial mechanical equipment.Keywords: Fault diagnosis, Heterogeneous sensing signals, Feature decoupling, Transfer learning, Cross-working conditionSubject terms: Engineering, Mathematics and computing Introduction In the modern industrial system, mechanical equipment is the core production carrier, and the stability of its operating state directly determines production efficiency, operational safety and maintenance costs1. Key rotating components such as bearings and gears long operate under complex conditions with variable loads, rotating speeds and strong noise interference, and are prone to failure due to wear, fatigue and impact. Untimely fault diagnosis may cause equipment shutdown, production interruption and even safety accidents2. Therefore, research on cross-working condition mechanical fault diagnosis to achieve accurate fault identification under variable working conditions has important theoretical value and academic significance for ensuring the continuous and stable operation of industrial systems. The development of cross-working condition mechanical fault diagnosis technology has always centered on eliminating operating condition interference and enhancing the distinguishability of fault features. Early research relied on traditional time-frequency analysis methods such as wavelet transform and empirical mode decomposition, combined with machine learning algorithms like support vector machines (SVM) and random forests to achieve fault identification3,4. These methods, with their concise principles and low computational costs, have been applied to a certain extent in simple operating condition scenarios. However, they highly depend on manually designed features, show poor adaptability to complex working condition coupling and signal noise, and hardly satisfy high-precision diagnostic requirements. With the rise of deep learning technology, models such as convolutional neural networks (CNN), long short-term memory (LSTM), and temporal convolutional networks (TCN) have been widely applied in this field5,6. Their capability for automatic feature extraction effectively overcomes the limitations of manual feature engineering, and significantly improved the diagnostic accuracy under single operating conditions by mining deep semantic features of fault signals. Among them, CNN is good at capturing spatial features and is suitable for the analysis of fault signals in the form of time-frequency graphs, while LSTM and TCN perform excellently in mining long-range dependency relationships of time-series signals, providing stronger feature expression capabilities for cross-working condition diagnosis. In addition, graph neural networks (GNNs) have shown great potential in fault diagnosis7,8, especially interpretable graph modeling, physics-informed graph networks, and the DCMLDF framework, which provide new ideas for relational modeling and cross-domain learning9. The introduction of transfer learning has provided core technical support for solving the problem of feature distribution shift across working conditions and has become a research hotspot in recent years10,11. According to the consistency of signal modalities between the source domain and the target domain, existing methods can be divided into two categories: homogeneous transfer and heterogeneous transfer12,13. Homogeneous transfer assumes that the signal modalities of the source domain and the target domain are the same, and minimizes the distribution difference between them through domain adaptation algorithms. Typical representatives include Domain-Adversarial Neural Networks (DAN) and Joint Distribution Adaptation (JDA)14. Such methods have achieved good diagnostic results in homogeneous cross-working condition scenarios and are widely used in cross-working condition fault diagnosis tasks of single vibration signals. Heterogeneous transfer focuses on cross-domain adaptation of different modal signals (such as vibration, current, temperature), and its core is to construct a common feature mapping space to achieve unified expression and distribution alignment of heterogeneous features, effectively adapting to the actual scenario of multi-sensor heterogeneous data in industrial sites15,16. In recent years, the application of heterogeneous transfer learning in the field of fault diagnosis has gradually increased. Researchers have continuously improved the cross-domain transfer effect of heterogeneous signals by designing modality-specific projection heads and optimizing distribution alignment loss functions17,18. As an effective way to mitigate the coupling between working conditions and fault features, feature decoupling technology has become a research focus in the field of cross-working condition diagnosis in recent years. Its core idea is to separate the extracted mixed features into core features strongly related to faults and redundant features strongly related to operating conditions. By weakening the interference of operating condition features and strengthening the domain invariance of fault features, the cross-working condition generalization ability of the model is improved19. Existing feature decoupling methods are mainly implemented based on mutual information constraints, variational autoencoders (VAE), generative adversarial networks (GAN), etc., forcing the model to learn decoupled feature representations by constructing specific constraint loss functions20. Some studies have attempted to combine feature decoupling with transfer learning, further improving the cross-working condition diagnostic performance by first decoupling features and then performing domain alignment21. However, most of them adopt a single-stream decoupling structure, resulting in limited decoupling effects, and lack targeted design for heterogeneous signal scenarios. At present, a mature technical framework has been formed in the field of cross-working condition mechanical fault diagnosis, covering core links such as signal processing, feature extraction and cross-domain transfer. However, in complex heterogeneous data scenarios, existing state-of-the-art methods still have prominent technical defects: they optimize feature extraction, domain adaptation and feature decoupling independently without collaborative optimization mechanisms. The single-branch decoupling structure cannot completely separate fault features from working condition interference, resulting in residual distribution shifts that affect diagnostic accuracy. These inherent defects restrict the practical application of existing methods, highlighting the urgent need for an integrated framework to break through the current technical bottleneck. This paper proposes the WDH-Net (Wavelet-ConvNet+Dual-stream decoupling+Heterogeneous transfer Network) model, which constructs an integrated cross-working condition fault diagnosis architecture integrating time-frequency feature enhancement, dual-stream decoupling encoding and heterogeneous feature alignment. Firstly, Wavelet-ConvNet and cross-attention mechanism are used to accurately extract and enhance the time-frequency features of heterogeneous signals. Then, the dual-stream decoupling encoder is utilized to separate the core fault features from the redundant operating condition features. Finally, cross-domain transfer is completed through the heterogeneous common feature mapping space and optimized distribution alignment strategy, and the overall optimization of the model is realized combined with the joint loss function. The technical route follows theoretical analysis, architectural design, experimental validation and result discussion, sequentially completing model construction, parameter optimization and performance verification to ensure the scientificity and effectiveness of the method. The main contributions of this paper are as follows: An integrated end-to-end WDH-Net architecture is proposed for cross-working-condition heterogeneous mechanical fault diagnosis, which unifies wavelet-based time-frequency enhancement, dual-stream feature decoupling, and heterogeneous domain adaptation into a collaborative learning framework. Different from existing schemes that separately optimize feature extraction, decoupling or cross-domain alignment, this design provides a unified paradigm that simultaneously addresses feature distribution shift and modal mismatch under variable working conditions, rather than a simple incremental combination of existing modules. An integrated end-to-end WDH-Net architecture is proposed for cross-working-condition heterogeneous mechanical fault diagnosis, which unifies wavelet-based time-frequency enhancement, dual-stream feature decoupling, and heterogeneous domain adaptation into a collaborative learning framework. Different from existing schemes that separately optimize feature extraction, decoupling or cross-domain alignment, this design provides a unified paradigm that simultaneously addresses feature distribution shift and modal mismatch under variable working conditions, rather than a simple incremental combination of existing modules. A Wavelet-ConvNet+Cross-Attention module is developed to strengthen the extraction of fault-sensitive features from heterogeneous vibration and current signals. By combining adaptive wavelet packet decomposition and cross-attention weighting, the module improves the robustness of learned features against noise and working condition fluctuations, enabling more reliable perception of weak fault characteristics under complex operating scenarios. A Wavelet-ConvNet+Cross-Attention module is developed to strengthen the extraction of fault-sensitive features from heterogeneous vibration and current signals. By combining adaptive wavelet packet decomposition and cross-attention weighting, the module improves the robustness of learned features against noise and working condition fluctuations, enabling more reliable perception of weak fault characteristics under complex operating scenarios. A collaborative optimization mechanism is constructed by jointly introducing feature orthogonality constraints and heterogeneous multi-kernel maximum mean discrepancy, which realizes the coordinated optimization of feature decoupling and domain alignment. This mechanism is conceptually distinct from conventional sequential or independent designs: the dual-stream encoder eliminates working condition interference at the feature level, while the heterogeneous alignment module learns domain-invariant representations, leading to significantly improved diagnosis accuracy and generalization under large domain discrepancies. A collaborative optimization mechanism is constructed by jointly introducing feature orthogonality constraints and heterogeneous multi-kernel maximum mean discrepancy, which realizes the coordinated optimization of feature decoupling and domain alignment. This mechanism is conceptually distinct from conventional sequential or independent designs: the dual-stream encoder eliminates working condition interference at the feature level, while the heterogeneous alignment module learns domain-invariant representations, leading to significantly improved diagnosis accuracy and generalization under large domain discrepancies. This paper is divided into 6 chapters, and the core content of each chapter is as follows: Chapter “Related work” expounds the research background and significance, sorts out the research status at home and abroad, and clarifies the research content and main contributions; Chapter "Design of the proposed cross-working condition fault diagnosis architecture" introduces the core theoretical basis of time-frequency feature extraction, attention mechanism, dual-stream decoupling and heterogeneous transfer; Chapter “Experimental design” details the design of WDH-Net model architecture, each module structure and training strategy; Chapter "Experimental results and analysis" verifies the diagnostic performance of the model in cross-working condition heterogeneous data scenarios through comparative experiments and ablation experiments; Chapter “Discussion” analyzes the core advantages, limitations and future research directions of the model; Chapter “Conclusion” summarizes the full text work and core conclusions. Related work Time-frequency feature extraction and attention mechanism The non-stationary and nonlinear properties of mechanical fault signals make time-frequency feature extraction a core prerequisite for fault diagnosis. In existing research, wavelet transform and its improved variants are widely adopted for time-frequency domain analysis of fault signals22. Related works enhance the extraction of sensitive frequency bands corresponding to fault impact characteristics by optimizing wavelet basis selection, adaptive decomposition layers and threshold adjustment, mitigate mode mixing inherent in conventional wavelet transform with fixed parameters, and strengthen the time-frequency feature representation capability for fault identification23,24. On this basis, lightweight convolutional neural networks have been extensively introduced into deep time-frequency feature mining. Improved lightweight architectures such as MobileNet and ShuffleNet achieve multi-level extraction of local spatial features from time-frequency graphs, while controlling model parameters and computational complexity via depthwise separable convolution and group convolution, making them suitable for industrial scenarios with limited computing resources25. The integration of attention mechanisms and convolutional networks has become a mainstream direction for time-frequency feature enhancement26. Related studies assign dynamic weights to different dimensions of time-frequency features through channel, spatial or temporal attention modules, prioritize key features highly contributive to fault recognition, and suppress redundant interference from background noise and working condition fluctuations27. Attention-based models including CBAM and SE-Net greatly improve the feature distinguishability of single-modal signals through weighted optimization of feature channels and spatial dimensions28; other efforts further explore the combination of attention and time-frequency analysis, adopting attention weights to guide time-frequency decomposition and prioritize feature learning of fault-sensitive frequency bands29,30. Nevertheless, most existing methods adopt a serial pipeline of time-frequency analysis, feature extraction and attention weighting, leading to insufficient fusion between time-frequency decomposition and attention modeling and limiting the ability to capture cross-dimensional correlations in the time-frequency domain. Meanwhile, current approaches lack adaptive designs for heterogeneous sensor signals, without differentiated extraction strategies for time-frequency discrepancies among vibration, current, temperature and other modal signals31. Although CNN-LSTM hybrid models can model temporal and spatial patterns, they easily introduce information redundancy when fusing heterogeneous features, weakening extraction robustness for heterogeneous data32,33. Moreover, existing attention-based schemes mainly enhance single-modal features but fail to mine complementary time-frequency information across heterogeneous modalities, often losing unique fault characteristics inherent to each signal type. Recently, vision transformer (ViT) frameworks have been introduced into fault diagnosis to capture long-range dependencies in time-frequency representations, combined with semi-supervised learning and uncertainty quantification to reduce reliance on labeled data34. Meanwhile, ViT models built on time-frequency scalograms have achieved promising performance in milling tool fault diagnosis under limited annotations. Even so, existing ViT-based methods lack dedicated designs for heterogeneous signal decoupling and cross-working-condition domain adaptation. Aiming at the inadequate fusion of time-frequency analysis and attention modules, as well as poor adaptability to heterogeneous signals, the Wavelet-ConvNet+Cross-Attention module in WDH-Net achieves deep integration of adaptive wavelet packet decomposition and cross-attention, thereby enhancing time-frequency representation for heterogeneous signals and preserving complementary modal information. Two-stream decoupling and heterogeneous transfer learning Feature decoupling serves as a core strategy to address feature coupling between working conditions and fault patterns. Existing literature has extensively explored decoupling network architectures and constraint loss functions35. VAE- and GAN-based decoupling models separate mixed extracted features into fault-related intrinsic components and condition-induced redundant components using single-stream encoding combined with mutual information minimization36; further improved frameworks such as β-VAE and InfoGAN enforce decoupled feature learning via customized reconstruction loss and adversarial constraints37,38. Relevant studies have verified that decoupled features effectively boost cross-working-condition diagnosis accuracy, confirming that stripping condition-related components significantly enhances the domain invariance of fault features39. However, most single-stream encoding structures struggle to fully isolate working condition interference from fault semantics, leaving residual conditional bias within decoupled representations40. Targeting the incomplete separation limitation of single-stream decoupling, the dual-stream decoupling encoder in WDH-Net adopts parallel branches for fault and working condition modeling, and introduces feature orthogonal constraints to realize thorough disentanglement between fault representations and conditional interference. Heterogeneous transfer learning concentrates on cross-domain adaptation for multi-modal heterogeneous signals, with the core principle of learning a shared feature embedding space41. Domain adaptation algorithms including DAN, JDA and CORAL project heterogeneous signals such as vibration and current into a dimension-unified common space via modality-specific projection heads, and align cross-domain feature distributions using maximum mean discrepancy and domain adversarial training42. Advanced strategies further optimize common space alignment by adopting weighted multi-kernel discrepancy and class center matching to improve cross-domain consistency of heterogeneous features43. Nevertheless, existing methods generally perform direct distribution alignment on undifferentiated mixed features, inevitably embedding redundant conditional components into transfer optimization and degrading cross-domain diagnostic performance. In addition, current common space designs employ fixed modal weight allocation without dynamic adjustment according to the fault sensitivity of each signal modality. Heterogeneous transfer architectures such as MDAN and MMD-AAE cannot fully exploit complementary properties across heterogeneous modalities, restricting the upper bound of transferable diagnosis accuracy44. More importantly, most existing implementations treat feature decoupling and heterogeneous transfer as isolated procedures, lacking joint collaborative optimization and preventing targeted cross-domain alignment on purified fault features. Recent studies have proposed multi-sensor observer-based residual learning to enhance fault feature saliency under variable working conditions45, alongside Auto-Permutation Feature Importance for efficient discriminative feature selection. Lightweight TCN has also been deployed for predictive maintenance to support early fault detection with low-latency inference on industrial edge devices46. Still, these approaches rarely unify feature decoupling and heterogeneous transfer learning, and thus cannot well accommodate cross-working-condition scenarios with multi-modal heterogeneous inputs. To remedy insufficient alignment targeting and the absence of joint optimization, the heterogeneous common feature mapping space in WDH-Net enables targeted cross-domain alignment exclusively for decoupled fault features, and establishes a collaborative optimization paradigm bridging feature decoupling and heterogeneous transfer learning. Design of the proposed cross-working condition fault diagnosis architecture Overall architecture framework Aiming at the core problems of poor adaptive capability of heterogeneous signal features and serious coupling between working condition interference and fault semantics in cross-working condition mechanical fault diagnosis, this paper proposes an integrated WDH-Net diagnostic architecture integrating time-frequency feature enhancement, dual-stream decoupling encoding and heterogeneous feature alignment, with the overall structure illustrated in Fig. 1. The workflow of the proposed framework consists of three key stages. In the first stage of heterogeneous signal time-frequency feature enhancement, input multi-modal signals are transformed into time-frequency feature matrices via adaptive wavelet packet decomposition. Lightweight convolution is adopted for local feature extraction, and the cross-attention module further achieves adaptive weighted fusion across multi-modal and multi-band components, so as to screen out highly distinguishable fault-sensitive features. The second stage corresponds to dual-stream decoupling encoding. Enhanced features are fed into a dual-stream encoder, where the fault feature branch strengthens intrinsic fault representation learning and the working condition branch captures conditional dependency information. Combined with explicit loss constraints, the two categories of features are effectively disentangled. The third stage implements cross-domain alignment for heterogeneous features. Decoupled fault features are projected into a shared embedding space through modality-specific projection heads. Distribution alignment and dynamic modal weight optimization are subsequently realized via heterogeneous multi-kernel maximum mean discrepancy, and diagnostic results are finally output by the classifier. By coordinating the above three stages, WDH-Net establishes an end-to-end diagnostic pipeline, achieving tight integration of time-frequency enhancement, dual-stream feature decoupling and heterogeneous transfer learning, and provides an effective solution for mechanical fault diagnosis under complex cross-working-condition and heterogeneous data scenarios. Overall architecture diagram of the WDH-Net model for cross-working-condition heterogeneous mechanical fault diagnosis. Wavelet-ConvNet+Cross-Attention module As the core feature enhancement component of WDH-Net, the Wavelet-ConvNet+Cross-Attention module aims to achieve accurate extraction and weighted enhancement of time-frequency features from heterogeneous signals through deep fusion of adaptive wavelet packet decomposition and cross-attention mechanism, thereby providing high-quality feature inputs for subsequent dual-stream decoupling encoding. The module first transforms one-dimensional heterogeneous time-series signals into multi-band time-frequency features via adaptive wavelet packet decomposition. Subsequently, a lightweight convolutional network is employed to mine local feature correlations within the time-frequency domain. Finally, the cross-attention mechanism dynamically optimizes the weights of multi-modal and multi-band features, strengthening the representation of fault-sensitive features while suppressing redundant interference from irrelevant components. This integrated design is theoretically based on the time-frequency analysis theory of non-stationary signals and the ability of attention mechanisms to model inter-feature dependencies, effectively addressing the inherent challenge of feature inconsistency across heterogeneous signals and improving the discriminability of fault-related features. The core of adaptive wavelet packet decomposition is to determine the optimal decomposition layers and wavelet basis through parameter optimization to achieve accurate splitting of fault signals in the time-frequency domain, and its decomposition process can be expressed as: where represents the time-frequency feature component of the k-th frequency band in the j-th layer, is the adaptively optimized wavelet packet filter coefficient, x(t) is the input one-dimensional heterogeneous time-series signal, and is the signal sampling period. Through this formula, heterogeneous signals of different modalities (vibration, current, etc.) can be uniformly converted into multi-dimensional time-frequency feature matrices, retaining the detailed information of fault features in the time-frequency domain, and alleviating the mode mixing problem caused by fixed parameters in traditional wavelet transform. Adaptive parameter optimization is theoretically justified by the uncertainty principle of time-frequency analysis, ensuring that the decomposition process adapts to the dynamic characteristics of different heterogeneous signals and avoids information loss or redundancy. To further mine the local spatial correlation of time-frequency features, the decomposed time-frequency feature matrix is sent to a lightweight convolutional network, and local features are extracted through depthwise separable convolution. Its feature mapping process can be expressed as: where represents the depthwise separable convolution operation, which is used for independent spatial feature extraction of time-frequency features of each frequency band, represents the pointwise convolution operation, which is used to integrate local features of different channels, ⊙ is the element-wise multiplication, and is the local enhanced feature matrix processed by lightweight convolution. This formula strengthens the spatial correlation of time-frequency features while controlling the number of model parameters, providing a more distinguishable feature basis for subsequent attention weighting. The adoption of depthwise separable convolution is theoretically supported by the principle of feature disentanglement, which separates spatial feature extraction and channel fusion to reduce computational complexity without sacrificing feature representation ability, making it suitable for lightweight design of heterogeneous fault diagnosis models. Finally, adaptive weighted fusion of local enhanced features of multiple modalities and multiple frequency bands is performed through the cross-attention mechanism to focus on fault-sensitive features and weaken redundant information. The calculation process of attention weight can be expressed as: where represents the attention weight of the m-th frequency band feature of the i-th modality, and are the query vector and key vector of the feature respectively, and M is the total number of frequency bands. Through this formula, the weights of features of different modalities and different frequency bands can be dynamically allocated, making the module pay more attention to feature components with high contribution to fault identification. The cross-attention mechanism is theoretically derived from the self-attention framework, which effectively models the cross-modal and cross-frequency dependencies, ensuring that the model can adaptively focus on fault-sensitive features in complex heterogeneous signal environments. Based on the attention weights, the final fusion enhanced feature can be expressed as: where I is the total number of input modalities, is the local enhanced feature of the m-th frequency band of the i-th modality, and is the final enhanced feature output by the module. The above four formulas are progressive layer by layer, forming a complete feature enhancement link from time-frequency decomposition and local feature extraction to attention weighted fusion, effectively solving the problems of poor adaptability and insufficient fusion depth of heterogeneous signal features, and providing strong support for subsequent dual-stream decoupling encoding. This progressive feature enhancement process is consistent with the hierarchical feature learning theory of deep neural networks, ensuring that low-level time-frequency features are gradually transformed into high-level discriminative features, which lays a theoretical foundation for the effectiveness of subsequent dual-stream decoupling. Design of dual-stream decoupling encoder The dual-stream decoupling encoder is the core module of WDH-Net to realize effective separation of operating condition and fault features. It adopts a parallel design of fault feature stream and operating condition feature stream, focusing on the enhanced transmission of core fault features and the accurate capture of redundant operating condition features respectively. The parallel structure is theoretically motivated by the disentangled representation learning principle, which encourages the model to learn semantically independent latent factors corresponding to fault patterns and working condition variations. The two streams adopt completely differentiated network structures and learning objectives, which form independent feature extraction paths and avoid structural interference in the feature learning process. It achieves efficient decoupling of the two types of features through differentiated network structures, and minimizes the interference of operating condition fluctuations on fault diagnosis. Taking the enhanced features output by the Wavelet-ConvNet+Cross-Attention module as input, the encoder outputs fault features with high purity and strong domain invariance through targeted structural design, laying a foundation for subsequent heterogeneous cross-domain alignment. The fault feature stream adopts a residual dense convolution structure, which strengthens the extraction and transmission of core fault features through feature reuse and cross-layer fusion, avoids the gradient vanishing problem in deep network training, and fully retains the details of fault features. Residual dense learning is theoretically supported by deep network optimization theory, which enhances information flow and improves the discriminative capacity of fault-sensitive features. Its feature output can be expressed as: where represents the residual dense convolution operation, is the input enhanced feature, are the intermediate features output by each convolutional layer, represents the feature concatenation operation, represents the ordinary convolution operation, and is the core fault feature output by the fault feature stream. This formula strengthens the representation ability of fault features through deep cross-layer feature fusion, making the features more suitable for the needs of fault type discrimination. To further improve the domain invariance of fault features, cross-layer residual fusion optimization is performed on the output core fault features to make up for the limitations of feature extraction by a single convolutional layer. Residual connection follows the identity mapping principle, which preserves low-level details while enabling the network to focus on refining invariant fault characteristics. The optimization process can be expressed as: where is the optimized core fault feature, γ is the residual fusion coefficient used to balance the weights of the original enhanced features and residual dense convolution features, and represents the batch normalization operation used to normalize the feature distribution and improve the stability of network training. Through residual connection and normalization processing, this formula enables fault features to have stronger cross-working condition adaptability while retaining original information. The operating condition feature stream adopts a temporal convolutional network structure. Taking advantage of dilated convolution, it expands the receptive field without increasing the number of parameters, and accurately captures the long-range dependency of operating condition changes. The use of temporal convolutions is theoretically justified by the temporal coherence of working condition variations, which exhibit slow and continuous transitions in mechanical systems. This stream is specially used to model working condition information and does not participate in the learning of fault-related features, ensuring the independence of feature learning objectives. Its feature output can be expressed as: where represents the temporal convolutional network operation, represents the dilated convolution operation with a dilation coefficient of d, D is the set of dilation coefficients, and is the redundant operating condition feature output by the operating condition feature stream. This formula comprehensively captures the operating condition change information at different time scales through the combination of temporal convolutions with multiple dilation coefficients, providing a clear reference for the separation of fault features and operating condition features. Based on the output characteristics of the two types of features, feature orthogonalization processing is used to weaken the correlation between fault features and operating condition features to achieve preliminary decoupling. Orthogonal projection is derived from linear subspace theory, which enforces structural independence between fault and condition representations. The orthogonalization processing can be expressed as: where is the orthogonalized core fault feature, · represents the vector dot product operation, and represents the -norm squared operation. This operation projects fault features onto the orthogonal space of operating condition features, completely eliminating the linear correlation between the two feature streams and fundamentally preventing mutual interference. This formula further improves the purity of fault features by eliminating the operating condition feature components mixed in the fault features. In addition, the feature orthogonality loss is constructed and integrated into the joint loss function to constrain the training process, forcing the two streams to learn relatively independent feature representations and ensuring no interference during the whole learning process. The orthogonality loss provides explicit statistical regularization, ensuring that the two distributions remain uncorrelated in the learned feature space. The above formulas are progressive layer by layer, forming a complete link from feature extraction and optimization to orthogonal decoupling, effectively solving the problem of incomplete separation in traditional single-stream decoupling structures, and outputting core fault features with high purity and strong domain invariance. The fault features processed by this module have stripped off operating condition interference to the greatest extent, providing a high-quality foundation for accurate cross-domain feature alignment in the subsequent heterogeneous common feature mapping space, and also laying a key support for the entire WDH-Net architecture to achieve the core goal of cross-working-condition heterogeneous fault diagnosis. Design of heterogeneous common feature mapping space The heterogeneous common feature mapping space is the core carrier for WDH-Net to achieve cross-domain feature alignment. Its core goal is to map the multi-modal fault features output by the dual-stream decoupling encoder to a feature space with unified dimensions, eliminate heterogeneous modal differences and cross-domain distribution deviations through dynamic weight optimization and distribution alignment strategies, and provide consistent feature input for the final fault classification. Taking the orthogonalized high-purity fault features as input, this space realizes efficient fusion and cross-domain adaptation of heterogeneous features through hierarchical design of modality-specific projection, dynamic weight allocation and cross-domain distribution alignment, connecting the decoupling module and the classifier to complete the diagnostic closed loop. This hierarchical design follows the theory of heterogeneous domain adaptation, which emphasizes unified representation learning and distribution matching across different modalities and domains. First, aiming at the dimensional heterogeneity of fault features of different modalities, a modality-specific projection head is designed to map the features of each modality to the same dimension to achieve feature dimension unification. Modality-specific projection is theoretically derived from subspace learning, which maps heterogeneous features into a shared latent space while preserving modality-specific structural information. The projection process can be expressed as: where is the projected feature of the i-th modality, is the specific projection operation of the i-th modality, is the orthogonalized fault feature of the i-th modality, and and are the projection weight matrix and bias term respectively. This formula retains the unique information of fault features of each modality through targeted projection design, and realizes dimension unification at the same time, laying a foundation for subsequent fusion and alignment. Considering that the contribution of different modal features to fault diagnosis is different, a dynamic weight allocation mechanism is introduced to adaptively adjust the weights according to the discriminative ability of each modal feature, so as to strengthen the role of effective modal features. Dynamic weight allocation is based on multi-modal fusion theory, which assigns higher weights to more discriminative modalities to improve overall diagnostic reliability. The weight calculation can be expressed as: where is the dynamic weight of the i-th modal feature, is the similarity calculation function used to measure the consistency between the modal feature and the reference feature, is the cross-modal reference feature, and I is the total number of modalities. This formula dynamically allocates weights through similarity evaluation, suppresses redundant modal interference, and improves the discriminative ability of fused features. After completing the preliminary fusion of multi-modal features based on dynamic weights, it is necessary to further eliminate the feature distribution deviation between the source domain and the target domain. The heterogeneous multi-kernel maximum mean discrepancy algorithm is used to measure and minimize the cross-domain distribution distance. HMK-MMD is theoretically grounded in reproducing kernel Hilbert space theory, which enables robust distribution alignment for heterogeneous cross-domain data. The distribution difference calculation can be expressed as: where represents the heterogeneous multi-kernel maximum mean discrepancy, and are the fused features of the source domain and target domain respectively, and are the number of samples in the source domain and target domain respectively, Φ is the kernel mapping function, is the reproducing kernel Hilbert space, and is the -norm squared in this space. This formula accurately measures the cross-domain distribution difference through multi-kernel mapping, providing a quantitative basis for distribution alignment. Finally, combining the results of dynamic weight fusion and cross-domain distribution alignment, the final aligned common features are output to complete the feature processing flow of the entire mapping space. The joint optimization of fusion and alignment follows the theory of domain-invariant representation learning, ensuring consistent and transferable features across working conditions. The aligned features can be expressed as: where is the final output aligned common feature, δ is the alignment coefficient used to balance the influence of fused features and distribution alignment, and is the gradient direction of cross-domain distribution difference. The above formulas are progressive layer by layer, forming a complete processing link from dimension unification and weight optimization to distribution alignment, effectively solving the problems of difficult adaptation of heterogeneous modalities and large cross-domain distribution deviation. The aligned common features output by this mapping space have both consistency and discriminability, and can be directly sent to the classifier to complete fault type discrimination, providing key guarantee for WDH-Net to realize cross-working condition heterogeneous fault diagnosis. Loss function The overall loss function of WDH-Net adopts a multi-objective collaborative optimization strategy, which comprehensively considers the training requirements of feature enhancement, feature decoupling and cross-domain alignment. By weighted combination of classification loss, cross-domain alignment loss and feature orthogonality loss, it guides the network to learn fault features with high distinguishability, strong domain invariance and low redundancy, ensuring the diagnostic accuracy and generalization ability of the model in cross-working condition heterogeneous scenarios. Based on the local optimization objectives of each module, the overall loss function achieves global training balance through reasonable weight allocation, promoting the entire architecture to form a collaborative optimization closed loop. The classification loss is used to constrain the consistency between the aligned common features and fault labels, ensuring the model has accurate fault discrimination ability. The cross-entropy loss function is adopted to measure the deviation between the prediction results and real labels, and its calculation can be expressed as: where is the classification loss, N is the total number of samples, C is the number of fault categories, is the real label of the i-th sample corresponding to the c-th fault category (in one-hot encoding form), σ is the Softmax activation function, f(· ) is the classifier mapping function, and is the aligned common feature of the i-th sample. This formula guides the model to learn common features with strong discriminability by penalizing the deviation between predicted probabilities and real labels. The cross-domain alignment loss is used to minimize the distribution difference of common features between the source domain and target domain, and strengthen the domain invariance of features. It is constructed based on heterogeneous multi-kernel maximum mean discrepancy, and its calculation can be expressed as: where is the cross-domain alignment loss, and are the aligned common features of the source domain and target domain respectively, and is the heterogeneous multi-kernel maximum mean discrepancy function, which is used to accurately measure the cross-domain feature distribution distance. This formula echoes the distribution difference calculation in Section "Design of heterogeneous common feature mapping space", and improves the generalization ability of the model under the target domain working conditions by constraining the cross-domain distribution deviation. The feature orthogonality loss is used to strengthen the decoupling effect of fault features and operating condition features, ensuring the purity of fault features. It is constructed based on the orthogonality measurement of the two types of features, and its calculation can be expressed as: where is the feature orthogonality loss, and are the orthogonalized fault feature and operating condition feature of the i-th sample respectively, · represents matrix multiplication, and represents the Frobenius norm, which is used to measure the overall size of the matrix. This formula further consolidates the decoupling effect of the dual-stream decoupling encoder by minimizing the correlation between the two types of features, avoiding the interference of mixed operating condition features on fault discrimination. The overall loss function of WDH-Net is the weighted sum of the above three types of losses. By introducing weight coefficients to balance the influence of each optimization objective, the network is ensured to be optimized collaboratively, and its calculation can be expressed as: where is the overall loss, α, β and γ are the weight coefficients of classification loss, cross-domain alignment loss and feature orthogonality loss respectively, and satisfy , which are used to adjust the contribution of each loss according to the training priority. The above formulas are progressive layer by layer, forming a complete optimization system from local objective constraint to global loss fusion, effectively guiding the network to learn features that meet the requirements of cross-working condition heterogeneous diagnosis, and ensuring the diagnostic performance and stability of the model. Experimental design Dataset To verify the effectiveness and generalization ability of the proposed WDH-Net in cross-working condition heterogeneous fault diagnosis tasks, two types of public benchmark datasets for mechanical fault diagnosis were adopted in the experiments: the Case Western Reserve University (CWRU) dataset and the Paderborn University (PU) dataset. These datasets cover different equipment types, fault modes and operating conditions, ensuring the comprehensiveness and reliability of experimental verification. CWRU Dataset The CWRU dataset is derived from the bearing fault simulation experiment platform of Case Western Reserve University, taking the 6205-2RS JEM SKF deep groove ball bearing as the research object. Single-point faults with different diameters (0.1778mm, 0.3556mm, 0.5334mm) were prefabricated on the surfaces of the inner ring, outer ring and rolling elements of the bearing by electric discharge machining technology, and the normal fault-free working condition was set as a control. During the experiment, multi-working condition scenarios were constructed by adjusting the motor load (0HP, 1HP, 2HP, 3HP), the sampling frequency was set to 12kHz and 48kHz, and vibration signal data were collected. Each combination of fault type and working condition corresponds to multiple groups of continuous time-series samples, and the sample length was uniformly intercepted to 1024 data points, providing abundant labeled data support for cross-working condition fault diagnosis. PU Dataset The PU dataset is provided by the Faculty of Mechanical Engineering of Paderborn University, based on an experimental platform built with different types of rolling bearings (6203-2RSH/C3, 6204-2RSH/C3). It includes multiple fault modes such as no fault, outer ring fault, inner ring fault and rolling element fault, with clear distinctions in fault degree and fault location. The experiment set different motor speeds (1500rpm, 1797rpm, 2000rpm, 2400rpm, 2800rpm) and load conditions (0N, 100 N, 200 N, 300 N) to form diversified cross-working condition scenarios. Meanwhile, two types of heterogeneous sensor data, vibration signals and current signals, were collected with a sampling frequency of 64kHz, and the sample length was intercepted to 2048 data points, providing high-quality multi-modal data for heterogeneous signal fusion and cross-domain transfer diagnosis experiments. Dataset preprocessing To suppress noise and redundant information and guarantee rigorous model training, unified preprocessing pipelines are implemented for the CWRU bearing dataset and PU gear dataset according to their inherent signal characteristics. First, a Butterworth low-pass filter is applied to eliminate high-frequency noise. The cutoff frequency is set to 1000 Hz for the CWRU dataset, matching the main frequency range of bearing fault signatures. For the PU dataset, the cutoff frequency is configured at 800 Hz to accommodate the lower characteristic frequency of gear vibration signals. This setting filters irrelevant high-frequency interference induced by background noise and sensor clutter without distorting effective fault-related components. Second, the sliding window strategy is adopted to partition long time-series sequences. For the CWRU dataset, the window length is set to 2048 points, corresponding to a duration of 0.2048 s under a 10 kHz sampling rate, which guarantees that each segmented sample contains complete fault characteristic information. For the PU dataset, the window length is set to 1024 points to fit its 5 kHz sampling frequency. A fixed sliding step of 512 points is employed for both datasets to balance sample quantity and feature overlap, ensuring adequate data volume while preventing redundant sampling. Finally, min–max normalization is conducted to standardize signal amplitude distributions. For the CWRU dataset, normalization bounds are determined from the global statistical range of raw vibration signals. For the PU dataset, normalization parameters are configured independently for each signal modality to mitigate modal discrepancy interference, maintaining consistent data distribution and reducing amplitude heterogeneity during model training. After preprocessing, samples are partitioned by working conditions and fault categories to construct training and test sets, providing a reliable data basis for subsequent model training and diagnostic evaluation. Experimental environment The experimental hardware platform adopted an Intel Core i7-12700H processor, equipped with 32GB DDR5 memory and an NVIDIA RTX 4090 graphics card (24GB video memory), providing sufficient computing power support for model training and inference; the software environment was built based on Python 3.9, with the deep learning framework PyTorch 1.18.0, and auxiliary libraries including NumPy 1.24.3, SciPy 1.10.1, Matplotlib 3.7.1, etc., used for data processing, model construction and result visualization. Model parameter settings During model training, the Adam optimizer is used to update network parameters. It effectively balances convergence speed and training stability by adaptively adjusting the learning rate for different parameters, making it suitable for complex deep networks such as WDH-Net. The initial learning rate is set to , a value optimized for heterogeneous fault diagnosis to avoid gradient explosion or slow convergence caused by inappropriate learning rate initialization. To balance stability and convergence efficiency, a cosine annealing learning rate scheduler is adopted, with the learning rate decayed every 5 epochs and a decay factor of 0.9. This strategy dynamically adjusts the learning rate, helping the model escape local optima while ensuring stable convergence in the late training phase. The batch size is set to 32, which matches the memory capacity of the training hardware and maintains sufficient sample diversity in each batch to improve generalization without memory overflow. The total training epoch is set to 100, determined according to the model convergence behavior to ensure sufficient feature learning while avoiding redundant computation. An early stopping mechanism is applied when the validation loss does not improve for 10 consecutive epochs, effectively alleviating overfitting and enhancing the model generalization ability on the target domain. To ensure rigorous experimental reproducibility, a fixed random seed of 42 is employed for all experiments including data splitting, model initialization and batch sampling. All samples are constructed using non-overlapping sliding windows with a step size of 512 points to guarantee complete independence among samples. For each transfer task, the source domain contains 2100 samples and the target domain contains 900 samples with a unified 7:3 train-test split ratio. A 10% held-out validation set is uniformly used for hyperparameter selection across all models, and each transfer task is independently repeated 5 times to obtain statistically reliable mean ± standard deviation results. All baseline models follow the identical data split, validation protocol and evaluation procedure to ensure fair comparison. The key parameters of each module are set as follows. In the Wavelet-ConvNet+Cross-Attention module, the number of adaptive wavelet packet decomposition layers is set to 4, which fully decomposes fault signals into multi-scale frequency bands and captures sensitive features without excessive computation. The wavelet basis is set to db4, due to its superior time-frequency localization for transient fault feature extraction from non-stationary mechanical signals. The lightweight convolutional network uses 3 layers of depthwise separable convolution, balancing feature extraction performance and model efficiency by reducing parameters while maintaining deep time-frequency feature representation. The query and key vector dimensions of the cross-attention module are both set to 256, ensuring adequate feature representation while avoiding excessive computational cost. In the dual-stream decoupling encoder, the residual dense convolution block contains 6 convolutional layers with a kernel size of 3× 3. The 6-layer structure ensures sufficient feature reuse and deep feature learning, while the 3× 3 kernel effectively captures local feature correlations without introducing extra computational burden. The dilation coefficients of the temporal convolutional network are set to , expanding the multi-scale receptive field to capture long-range dependencies of time-series fault signals, which is critical for fault recognition under variable working conditions. In the heterogeneous common feature mapping space, the output dimension of the modality-specific projection head is set to 512, mapping heterogeneous vibration and current signals into a high-dimensional shared space with strong discriminability for subsequent cross-domain alignment. The heterogeneous multi-kernel maximum mean discrepancy uses a combination of Gaussian and polynomial kernels, integrating the Gaussian kernel’s ability to model nonlinear distributions and the polynomial kernel’s strength in fitting linear correlations to improve alignment accuracy. For the overall loss function, the weights are set as α =0.6 for classification loss, β =0.3 for cross-domain alignment loss, and γ =0.1 for feature orthogonality loss. This configuration prioritizes classification accuracy while maintaining balanced domain alignment and feature decoupling. Training process To clearly illustrate the learning pipeline and standardize the model training procedure, the complete training workflow of WDH-Net is formally presented in Algorithm 1. Training Protocol of WDH-Net. Evaluation metrics setting To comprehensively and objectively evaluate the fault diagnosis performance of WDH-Net and comparison models, four core classification evaluation metrics are adopted in the experiments, including Accuracy (Acc), Precision (P), Recall (R) and F1-Score (F1). Accuracy reflects the overall diagnostic effect of the model as the proportion of correctly diagnosed samples to the total number of samples; Precision evaluates the accuracy of the model’s positive predictions to avoid misdiagnosing normal samples as fault samples; Recall measures the coverage of fault samples by the model’s predictions to ensure all fault samples are identified; F1-Score is the harmonic mean of Precision and Recall, which comprehensively balances the two metrics to avoid one-sided evaluation results. In addition, the training time and parameter quantity of each model are recorded to conduct a comprehensive assessment from both performance and efficiency perspectives. Comparison models selection To comprehensively verify the superiority of WDH-Net, seven state-of-the-art fault diagnosis models covering various technical routes (transfer learning, domain adaptation, multi-source transfer, contrastive learning, etc.) were selected as benchmarks. All models share the same input data, preprocessing methods, and experimental environment to ensure fair comparison. In addition, the hyperparameters of all models were tuned systematically using the same optimization strategy to ensure that the comparison is conducted under equivalent conditions. Brief introductions are as follows: MMDAN47: A multi-scale and multi-view domain adversarial network, designed for partial set and cross-working condition fault diagnosis via knowledge transfer between source and target conditions. RSTCG48: An enhanced dual-channel feature fusion approach for rolling bearing fault diagnosis, converting temporal sequences to 2D feature maps and fusing multi-scale features from Swin Transformer and CNN2D with GAM. CDPDTLN49: A cross-domain deep transfer network, combining cross-domain data processing, improved multi-scale residual network, and combined domain distribution adaptation strategy. DANN50: An enhanced domain-adversarial neural network, fusing mechanism-simulated data and measured data with subdomain discriminator for cross-condition fault diagnosis. SSCTLN51: A semi-supervised contrastive transfer network, integrating LMMD, semi-supervised contrastive learning, and domain adversarial training for cross-condition bearing diagnosis. Proto-DISFNet52: A prototype-guided dual-feature transfer method, realizing domain feature disentanglement and prototype construction for cross-condition fault diagnosis. MSDF-DANN53: A multi-source domain fusion network, transforming signals into 2D features and adopting convolutional fusion module for bearing fault diagnosis. NaCdNet54: A cross-condition fault diagnosis framework that integrates convolutional neural networks, attention mechanisms, and real-time style transfer to adapt to data distribution variations caused by changing working conditions. MSCLDA55: A multi-source domain adaptation method based on supervised contrastive learning, which alleviates negative transfer and aligns cross-domain distributions by using prototype-guided pseudo-label optimization and adaptive domain weighting. Proto-ADDA-X56: A prototype-attention domain adaptation method that aligns class centroids and adaptively weights target samples based on similarity and confidence for cross-working-condition bearing fault diagnosis. It also provides interpretable visualization of feature alignment and prototype evolution during transfer. Experimental results and analysis The performance of each model in separate detection on CWRU and PU datasets, with consistent test conditions to ensure the fairness of comparison. As shown in Table 1, WDH-Net achieves remarkable performance advantages on both datasets, significantly outperforming all comparison models. On the CWRU dataset, WDH-Net obtains an accuracy of 94.18%, precision of 93.87%, recall of 93.72%, and F1-score of 93.79%. On the PU dataset, its performance is further optimized, with accuracy, precision, recall, and F1-score reaching 94.85%, 94.53%, 94.38%, and 94.45% respectively. Compared with MSCLDA (the second-best model), WDH-Net improves accuracy by 2.34 percentage points on both datasets, with precision, recall and F1-score increased by 2.0–3.0 percentage points. Such stable improvement benefits from the synergies of time-frequency enhancement, dual-stream decoupling and heterogeneous multi-kernel alignment, which help capture domain-invariant fault features. In terms of performance stability, the standard deviations of all indicators of WDH-Net are strictly controlled within 0.6, which is far lower than those of comparison models. Most comparison models have indicator standard deviations exceeding 1.0; among them, MMDAN shows the largest fluctuation, with the recall standard deviation reaching 1.70 on the CWRU dataset, indicating poor adaptability to the inherent characteristics of different datasets. MSCLDA and NaCdNet achieve competitive results (90%–92.5%) via style transfer and contrastive learning, but still lag behind WDH-Net due to the lack of effective feature decoupling and heterogeneous multi-kernel alignment. SSCTLN and Proto-DISFNet follow closely, with core indicators between 86% and 88%, and their performance is slightly inferior due to insufficient ability in handling complex distribution discrepancies. RSTCG, MSDF-DANN, and Improved DANN show moderate performance, with core indicators concentrated between 84% and 86% on both datasets. These models are limited by simple feature extraction and domain adaptation designs, failing to disentangle fault features from working condition interference. MMDAN and CDPDTLN perform the worst, with all indicators below 84%, which is attributed to their shallow domain adaptation designs that cannot cope with complex data distributions in bearing fault datasets. The PU dataset contains heterogeneous signals, stronger noise and more fault types, raising diagnosis difficulty. WDH-Net still performs stably well, verifying its strong ability in decoupling and heterogeneous alignment. The performance difference between models on the two datasets also reflects the adaptability of model structures to different data characteristics. The consistent high performance on both datasets verifies that the core design of WDH-Net can effectively capture the essential characteristics of bearing faults, regardless of data source differences and fault category variations. Table 1Performance comparison of different models on CWRU and PU datasets.ModelCWRU DatasetPU DatasetAcc (%)P (%)R (%)F1 (%)Acc (%)P (%)R (%)F1 (%)MMDAN82.45 ± 1.5681.78 ± 1.6281.56 ± 1.7081.67 ± 1.5083.17 ± 1.4882.51 ± 1.5582.28 ± 1.6282.39 ± 1.42RSTCG85.79 ± 1.2485.12 ± 1.3084.95 ± 1.3685.03 ± 1.1886.42 ± 1.1885.75 ± 1.2585.58 ± 1.3185.66 ± 1.12CDPDTLN83.62 ± 1.4282.95 ± 1.4882.74 ± 1.5582.84 ± 1.3584.29 ± 1.3283.62 ± 1.3883.41 ± 1.4583.51 ± 1.25Improved DANN84.38 ± 1.3583.71 ± 1.4283.50 ± 1.4883.60 ± 1.2885.03 ± 1.2584.36 ± 1.3284.15 ± 1.3884.25 ± 1.18SSCTLN87.26 ± 1.0586.61 ± 1.1286.45 ± 1.1886.53 ± 1.0087.91 ± 0.9887.26 ± 1.0587.09 ± 1.1287.17 ± 0.92Proto-DISFNet86.89 ± 1.1086.24 ± 1.1686.08 ± 1.2286.16 ± 1.0587.54 ± 1.0286.89 ± 1.0886.72 ± 1.1586.80 ± 0.98MSDF-DANN85.14 ± 1.2884.47 ± 1.3584.30 ± 1.4184.38 ± 1.2285.81 ± 1.2285.14 ± 1.2884.97 ± 1.3585.05 ± 1.15NaCdNet90.36 ± 0.8289.84 ± 0.8689.65 ± 0.9189.74 ± 0.7991.02 ± 0.7690.46 ± 0.8090.27 ± 0.8590.36 ± 0.73MSCLDA91.84 ± 0.7391.31 ± 0.7791.12 ± 0.8291.21 ± 0.7092.51 ± 0.6891.95 ± 0.7291.76 ± 0.7791.85 ± 0.66WDH-Net94.18 ± 0.5593.87 ± 0.5093.72 ± 0.5893.79 ± 0.4894.85 ± 0.5294.53 ± 0.4694.38 ± 0.5494.45 ± 0.42 Performance comparison of different models on CWRU and PU datasets. Taking the CWRU dataset as the experimental object, four groups of cross-load condition transfer tasks (source domain/target domain: 0HP?1HP, 0HP?2HP, 0HP?3HP, 1HP?3HP) were constructed to verify the performance of WDH-Net in the scenario of single-modal (vibration signal) cross-working condition fault diagnosis. The average diagnostic indicators and standard deviations of each model are shown in Table 2. Table 2Performance comparison of cross-working condition single-modal diagnosis on CWRU dataset.ModelAcc (%)P (%)R (%)F1 (%)MMDAN89.24 ± 1.2588.76 ± 1.1888.51 ± 1.3288.63 ± 1.21RSTCG92.57 ± 0.9891.89 ± 1.0591.74 ± 1.1291.81 ± 1.02CDPDTLN90.18 ± 1.1689.65 ± 1.2189.42 ± 1.2889.53 ± 1.15Improved DANN91.05 ± 1.0990.48 ± 1.1390.31 ± 1.2090.39 ± 1.08SSCTLN93.12 ± 0.8592.67 ± 0.9292.54 ± 0.8892.60 ± 0.81Proto-DISFNet92.89 ± 0.9292.35 ± 0.9692.18 ± 1.0192.26 ± 0.90MSDF-DANN91.76 ± 1.0591.22 ± 1.1091.08 ± 1.1591.15 ± 1.03NaCdNet94.06 ± 0.7693.61 ± 0.8093.45 ± 0.8593.53 ± 0.73MSCLDA94.68 ± 0.7094.23 ± 0.7494.07 ± 0.7994.15 ± 0.68WDH-Net97.85 ± 0.4297.62 ± 0.3897.58 ± 0.4597.60 ± 0.36 Performance comparison of cross-working condition single-modal diagnosis on CWRU dataset. It can be seen from Table 2 that WDH-Net exhibits significant performance advantages and stability in the cross-working condition single-modal diagnosis task on the CWRU dataset, with an average accuracy of up to 97.85%, which is 3.17 percentage points higher than that of the sub-optimal model MSCLDA, and the precision, recall and F1-score are also improved by more than 3 percentage points simultaneously. This substantial improvement demonstrates the superiority of the proposed method in cross-load transfer tasks. Moreover, the standard deviations of all indicators are strictly controlled within 0.5, far lower than those of other comparison models, indicating that its performance fluctuation is extremely small in different cross-load condition tasks and it has strong adaptability. In-depth analysis shows that the performance differences of the comparison models mainly stem from the limitations of feature extraction and cross-domain alignment strategies: MSCLDA and NaCdNet achieve competitive results by using contrastive learning and style transfer to reduce domain shift, but they lack an explicit dual-stream structure to decouple fault features and working condition features, resulting in residual interference in feature representation. SSCTLN can strengthen the aggregation of homogeneous features by virtue of the semi-supervised contrastive learning mechanism, so its indicators are relatively excellent and the standard deviation is also at a low level; while models such as MMDAN and CDPDTLN only adopt traditional domain adaptation methods, failing to effectively separate the load condition interference from fault features, resulting in not only low accuracy but also standard deviation exceeding 1.0, and their performance decreases significantly in high load difference tasks. The superior performance of WDH-Net is attributed to the Wavelet-ConvNet and cross-attention module that enhances weak fault features in the time-frequency domain, and the dual-stream decoupling structure that completely eliminates working condition interference, enabling the model to learn pure and domain-invariant fault features. As shown in Table 3, Proto-ADDA-X achieves an accuracy of 100% on the CWRU dataset, while WDH-Net reports 97.85 ± 0.42. This difference can likely be attributed to the fact that Proto-ADDA-X uses a fixed source-to-target transfer setting (1HP→3HP), while WDH-Net is evaluated across multiple cross-working-condition tasks, leading to potential error accumulation across different task types. Therefore, WDH-Net demonstrates superior generalization ability, as it handles a wider range of target conditions. On the PU dataset, WDH-Net achieves an accuracy of 96.92 ± 0.48 in the heterogeneous modality task, whereas Proto-ADDA-X also reports 100% accuracy on the Paderborn dataset. It should be noted that the Paderborn dataset used by Proto-ADDA-X differs significantly from WDH-Net’s PU dataset, as Proto-ADDA-X utilizes only vibration signals, while WDH-Net combines vibration and current signals. This difference in input modalities leads to a slight decrease in WDH-Net’s performance. However, in real industrial applications, WDH-Net’s ability to handle more complex input signals may provide better adaptability. Overall, Proto-ADDA-X provides extremely high accuracy on well-known benchmark datasets, but it is limited to vibration signal input, and its transfer protocol is relatively fixed, reducing its adaptability. In contrast, WDH-Net demonstrates strong cross-modal and cross-task generalization capabilities through more complex signal fusion and cross-working-condition learning, although its accuracy may slightly decrease in certain tasks. Table 3Supplementary benchmark comparison between WDH-Net and the recent Proto-ADDA-X method on public cross-domain bearing fault diagnosis tasks.MethodDatasetSignal modalityTransfer settingAccuracy (%)WDH-NetCWRUVibrationAverage over 0HP→1HP, 0HP→2HP,0HP→3HP, 1HP→3HP97.85 ± 0.42Proto-ADDA-XCWRUVibration1HP (1772 rpm)→3HP (1750 rpm)100.00WDH-NetPUVibration + CurrentAverage over five cross-speed/load transfer tasks96.92 ± 0.48Proto-ADDA-XPaderbornVibrationCondition 1→Condition 2100.00 Supplementary benchmark comparison between WDH-Net and the recent Proto-ADDA-X method on public cross-domain bearing fault diagnosis tasks. Average over 0HP→1HP, 0HP→2HP, 0HP→3HP, 1HP→3HP Based on the PU dataset, five groups of cross-rotation speed-load condition transfer tasks (source domain/target domain: 1500rpm-0N?1797rpm-100N, 1500rpm-0N?2000rpm-200N, 1500rpm-0N?2400rpm-300N, 1797rpm-100N?2800rpm-300N, 2000rpm-200N?2800rpm-300N) were constructed. Vibration + current heterogeneous modal signals were input to verify the model’s adaptability to heterogeneous features and generalization performance across working conditions. The experimental results are shown in Table 4. In the heterogeneous-modal cross-working condition diagnosis task, WDH-Net still maintains an absolutely leading position, with an average accuracy of 96.92%, which is 3.75 percentage points higher than that of the sub-optimal model MSCLDA, demonstrating its remarkable superiority in handling heterogeneous cross-domain tasks. The standard deviations of all indicators are maintained at around 0.5, showing particularly prominent stability advantages. The differences in feature distribution of heterogeneous modal signals (vibration + current) and the domain shift caused by the coupled changes of rotation speed and load are the core reasons for the limited performance of comparison models: MSCLDA and NaCdNet achieve competitive performance through contrastive learning and style transfer, but lack specialized heterogeneous modal alignment and dynamic weight optimization, thus cannot fully utilize the complementary information of multi-modal signals. Although RSTCG adopts a dual-channel architecture for feature fusion, it only focuses on optimization at the feature extraction level and does not design targeted cross-domain alignment and heterogeneous modal adaptation strategies, resulting in obvious performance fluctuations (standard deviation exceeding 1.0) in high rotation speed and high load difference tasks (e.g., 1500rpm-0N?2800rpm-300N); MSDF-DANN supports multi-source domain fusion but fails to fully explore the complementarity of vibration and current signals, and domain shift still exists after feature fusion, leading to inferior performance compared with SSCTLN and Proto-DISFNet; Models such as MMDAN and CDPDTLN are limited by a single domain alignment strategy, making it difficult to adapt to the complex distribution of heterogeneous modalities, with accuracy generally lower than 88% and standard deviation even exceeding 1.2. In contrast, WDH-Net effectively unifies the feature dimensions of vibration and current signals through the heterogeneous common feature mapping space, achieves accurate cross-domain feature alignment combined with multi-kernel MMD, and strengthens high-contribution modal features through dynamic weight assignment. Thus, in complex heterogeneous and multi-working condition coupling scenarios, it not only ensures diagnostic accuracy but also achieves excellent stability. Table 4Performance comparison of cross-working condition heterogeneous-modal diagnosis on PU dataset.ModelAcc (%)P (%)R (%)F1 (%)MMDAN86.17 ± 1.3885.52 ± 1.4585.36 ± 1.5285.44 ± 1.32RSTCG89.83 ± 1.0289.15 ± 1.0888.97 ± 1.1589.06 ± 0.98CDPDTLN87.59 ± 1.2486.94 ± 1.3186.78 ± 1.3686.86 ± 1.20Improved DANN88.24 ± 1.1587.61 ± 1.2287.45 ± 1.2887.53 ± 1.10SSCTLN90.58 ± 0.9089.97 ± 0.9689.81 ± 1.0289.89 ± 0.85Proto-DISFNet90.21 ± 0.9589.58 ± 1.0189.42 ± 1.0889.50 ± 0.92MSDF-DANN89.06 ± 1.0888.41 ± 1.1588.25 ± 1.2188.33 ± 1.05NaCdNet92.35 ± 0.8291.82 ± 0.8791.65 ± 0.9391.73 ± 0.78MSCLDA93.17 ± 0.7592.64 ± 0.8092.48 ± 0.8692.56 ± 0.72WDH-Net96.92 ± 0.4896.71 ± 0.4296.65 ± 0.5196.68 ± 0.39 Performance comparison of cross-working condition heterogeneous-modal diagnosis on PU dataset. Combined with the specific working condition combinations corresponding to domain discrepancy intensities noted in Table 5, the diagnostic accuracy of all models showed varying degrees of decline with the increase of domain discrepancy intensity, but there were significant differences in the decline range and absolute performance. On the CWRU dataset, the accuracy of WDH-Net reached 98.56% under low domain discrepancy intensity (0HP→1HP), which was 2.04 percentage points higher than that of the second-best model MSCLDA. As the domain discrepancy intensity increased to the highest level (0HP→3HP), the accuracy of WDH-Net still remained at 97.36%, with a decline of only 1.2 percentage points, while most other models had a decline of 3–5 percentage points. Among them, the accuracy of MMDAN decreased from 91.35% to 87.65%, with a decline of 3.7 percentage points. This trend was more obvious in the heterogeneous modal cross-domain task of the PU dataset. Under high domain discrepancy intensity (1500rpm-0N→2800rpm-300N), the accuracy of WDH-Net was 95.87%, which was 5.11 percentage points higher than that of MSCLDA and 11.95 percentage points higher than that of MMDAN (the worst-performing model). Moreover, the standard deviation of its accuracy under all intensities was controlled within 0.5, much lower than that of other models, reflecting extremely strong performance stability. NaCdNet and MSCLDA outperform traditional methods via style transfer and contrastive learning, yet suffer severe performance degradation under high discrepancy due to the absence of effective feature decoupling and heterogeneous domain alignment. This result fully indicates that WDH-Net can effectively weaken the impact of feature distribution shift caused by domain discrepancy through the integrated architecture design of time-frequency feature enhancement, dual-stream decoupling encoding and heterogeneous feature alignment. Even in scenarios where the working condition difference between the source domain and the target domain is severe, it can accurately extract core fault features with domain invariance. In contrast, other models are difficult to cope with the challenges brought by the change of domain discrepancy intensity due to the lack of targeted feature decoupling mechanisms or single domain alignment strategies, leading to rapid performance attenuation with the increase of domain discrepancy. The dual-stream decoupling structure eliminates working condition interference fundamentally, and the heterogeneous multi-kernel MMD ensures robust alignment even under large distribution shifts, which jointly enables WDH-Net to maintain stable and high-precision diagnosis. Table 5Cross-domain diagnostic performance of models under different domain discrepancy intensities.DatasetDiscrepancy IntensityMMDANRSTCGCDPDTLNImproved DANNSSCTLNProto-DISFNetMSDF-DANNNaCdNetMSCLDAWDH-NetCWRULow91.35 ± 1.1894.26 ± 0.9292.47 ± 1.1093.15 ± 1.0395.48 ± 0.7895.12 ± 0.8593.02 ± 1.0196.15 ± 0.7096.52 ± 0.6598.56 ± 0.35Medium88.72 ± 1.2592.13 ± 0.9890.05 ± 1.1690.83 ± 1.0993.76 ± 0.8293.45 ± 0.9090.78 ± 1.0794.42 ± 0.7594.86 ± 0.7097.63 ± 0.41High87.65 ± 1.3291.35 ± 1.0587.92 ± 1.2288.96 ± 1.1591.15 ± 0.9090.83 ± 0.9788.95 ± 1.1392.08 ± 0.8092.51 ± 0.7697.36 ± 0.48PULow88.42 ± 1.3092.17 ± 1.0189.83 ± 1.2590.56 ± 1.1893.28 ± 0.8892.95 ± 0.9591.03 ± 1.1094.16 ± 0.8094.63 ± 0.7597.89 ± 0.38Medium86.05 ± 1.3889.75 ± 1.0887.32 ± 1.3288.15 ± 1.2591.45 ± 0.9291.08 ± 1.0288.86 ± 1.1792.37 ± 0.8592.85 ± 0.8096.78 ± 0.45High83.92 ± 1.4587.53 ± 1.1585.17 ± 1.3986.02 ± 1.3289.32 ± 0.9888.97 ± 1.0886.75 ± 1.2490.24 ± 0.9090.76 ± 0.8595.87 ± 0.52Note: Low discrepancy: CWRU (0HP→1HP), PU (1500rpm-0N→1797rpm-100N); Medium discrepancy: CWRU (0HP→2HP), PU (1500rpm-0N→2000rpm-200N); High discrepancy: CWRU (0HP→3HP), PU (1500rpm-0N→2800rpm-300N). Cross-domain diagnostic performance of models under different domain discrepancy intensities. Note: Low discrepancy: CWRU (0HP→1HP), PU (1500rpm-0N→1797rpm-100N); Medium discrepancy: CWRU (0HP→2HP), PU (1500rpm-0N→2000rpm-200N); High discrepancy: CWRU (0HP→3HP), PU (1500rpm-0N→2800rpm-300N). Figure 2 presents the confusion matrix of WDH-Net for different fault types on the CWRU and PU datasets. It can be seen that in the diagnosis of various faults on the CWRU dataset, the diagnostic accuracy at the diagonal positions remains at a high level. Specifically, the diagnostic accuracy of normal state, inner race fault, outer race fault and roller fault reaches 98.2%, 97.8%, 97.5% and 97.8% respectively, while the values at non-diagonal positions are generally low, indicating that misjudgment between different fault types is rare. On the PU dataset, the diagnostic accuracy of normal state, inner race fault, outer race fault, roller fault and cage fault reaches 96.5%, 95.7%, 96.2%, 95.6% and 95.2% respectively, and the proportion of cross misjudgment among various fault types is also in a low range. Overall, WDH-Net has stable and excellent diagnostic differentiation capabilities for various faults under different datasets. This result further verifies the effectiveness of the model in the fault feature extraction and classification recognition stages, and it can accurately distinguish different types of mechanical fault states. Confusion matrix of diagnostic results of WDH-Net for various fault types under different datasets. To theoretically illustrate the efficiency of the proposed model, the computational complexity of each core module in WDH-Net is analyzed from a mathematical perspective. For the Wavelet-ConvNet+Cross-Attention module, given an input feature map of size , a convolution kernel of size k × k, and an output channel number C', its time complexity can be expressed as . Since the cross-attention mechanism only performs linear projection and weighted summation on local feature channels, its complexity is , which is at the same order as conventional channel attention and does not introduce excessive overhead. For the dual-stream decoupling encoder, the fault stream adopts residual dense convolution with feature reuse, and its complexity is , where L is the number of layers. The operating condition stream uses temporal convolution with fixed dilation rates, and its complexity is , where D denotes the number of dilation scales. Since the two streams perform parallel computing and share the initial feature extraction layer, the overall complexity of the dual-stream structure does not show a linear superposition. For the heterogeneous feature alignment module, the complexity of mapping heterogeneous features to a common space is O(HW C C''), where C'' is the dimension of the common feature space. The H-MMD calculation only involves statistical distance measurement of feature distributions, with a complexity of O(N C''), where N is the batch size. Overall, the computational complexity of all core modules in WDH-Net remains at a linear or quadratic order with respect to feature dimensions, and no high-order complex operations are introduced, which guarantees its computational efficiency. The efficiency of each model is evaluated from two core aspects, parameter quantity and training time consumption, with tests conducted on the same hardware platform to ensure the fairness of comparison. The specific efficiency performance indicators of each model are shown in Table 6. WDH-Net achieves significant advantages in both parameter quantity and training time consumption compared with other comparison models. In terms of parameter quantity, WDH-Net has the smallest parameter volume among all models, reaching 11.3 M. It is 9.7 percent lower than MSCLDA (12.5 M) and 13.7 percent lower than NaCdNet (13.1 M), and is 29.0 percent lower than SSCTLN and 38.9 percent lower than RSTCG which has the largest parameter quantity. Such lightweight performance provides favorable conditions for its deployment in industrial embedded systems with limited resources. Regarding training time consumption, consistent test parameters on the same hardware platform ensure that the comparison results are reliable. WDH-Net consumes 27.9 seconds per epoch on the CWRU dataset and 30.8 seconds per epoch on the PU dataset, with an average training time of 29.35 seconds per epoch, which is also faster than MSCLDA, NaCdNet, SSCTLN and other models. The gap between its training time on the two datasets is only 2.9 seconds, indicating stable adaptation efficiency to different datasets. WDH-Net maintains high diagnostic accuracy above 94 percent in previous cross-working condition and cross-dataset tasks while optimizing both parameter quantity and training time consumption, breaking the inherent bottleneck between high performance and high resource consumption. In contrast, RSTCG has redundant architecture design, leading to the largest parameter quantity and longest training time among all models. MSCLDA and NaCdNet achieve competitive diagnostic performance but are inferior to WDH-Net in lightweight design and training efficiency. SSCTLN performs well in diagnostic accuracy but lags behind WDH-Net in both parameter quantity and training time consumption, failing to achieve a balanced performance between efficiency and accuracy. Other comparison models either have insufficient diagnostic accuracy or fail to optimize resource consumption, resulting in poor comprehensive application value compared with WDH-Net. Table 6Efficiency performance comparison of different models.ModelParameter Quantity (M)Training Time per Epoch (s)Training Time per Epoch (s)Average Training Time (s)(CWRU Dataset)(PU Dataset)MMDAN12.828.732.430.55RSTCG18.535.238.736.95CDPDTLN14.231.537.833.65Improved DANN13.629.833.631.70SSCTLN15.933.637.235.40Proto-DISFNet16.734.137.835.95MSDF-DANN14.830.434.332.35NaCdNet13.130.133.932.00MSCLDA12.529.533.231.35WDH-Net11.327.930.829.35 Efficiency performance comparison of different models. Figure 3 shows the diagnostic accuracy performance of each model under different SNR conditions. It can be observed that the accuracy of all models gradually increases with the rise of SNR, and the accuracy of WDH-Net remains at the highest level among all models, with a significant performance gap maintained across the entire SNR range. As two recent competitive methods, NaCdNet and MSCLDA achieve better performance than traditional domain adaptation models by employing style transfer and contrastive learning strategies, and their overall accuracy is slightly higher than SSCTLN and Proto-DISFNet. In low SNR scenarios (e.g., SNR = 5 dB), the advantage of WDH-Net is particularly prominent: its accuracy on both datasets is at least 3.2 percentage points higher than that of other models. In high SNR scenarios, the accuracy of WDH-Net approaches 97%, which further reflects its strong adaptability to noise interference. Meanwhile, the accuracy curve of WDH-Net has smaller fluctuations, indicating that its performance is less affected by changes in noise intensity and can maintain stable diagnostic capabilities in different noise environments. This result demonstrates that the time-frequency feature enhancement and dual-stream decoupling design adopted by WDH-Net can effectively extract fault-sensitive information and suppress noise interference, thus maintaining excellent diagnostic performance in complex noise scenarios. Diagnostic accuracy variation of different models on CWRU and PU datasets under different SNR levels. The table7 above presents the accuracy performance of WDH-Net on the CWRU and PU datasets under different loss function coefficient combinations. It can be clearly observed that when the coefficient combination is set as 0.6, 0.3, 0.1, the model achieves the highest accuracy on both datasets, with 94.18% on the CWRU dataset and 94.85% on the PU dataset. Meanwhile, the corresponding standard deviations are the smallest, at 0.55 and 0.52 respectively, which reflects the superiority and stability of the model’s diagnostic performance under this coefficient combination. This result verifies that this coefficient ratio can optimally balance the optimization objectives of classification loss, cross-domain alignment loss and feature orthogonality loss, enabling the model to simultaneously take into account the accuracy of fault classification, cross-domain adaptability and feature purity. Under other coefficient combinations, the model’s accuracy decreases to varying degrees. An excessively high weight of classification loss will lead to insufficient cross-domain alignment, limiting the generalization ability of the model in the target domain; an excessively high weight of cross-domain alignment loss will weaken the classification objective and reduce the accuracy of fault identification; an excessively high weight of feature orthogonality loss will overemphasize feature decoupling, which may lead to the loss of some key information related to faults. All these factors will result in a decline in the overall diagnostic accuracy. In addition, under the same coefficient combination, the accuracy of the PU dataset is slightly higher than that of the CWRU dataset overall. This is related to the fact that the PU dataset contains heterogeneous modal signals of vibration and current, and the complementarity of heterogeneous features provides the model with richer fault information, helping to improve diagnostic performance. On the whole, the reasonable ratio of loss function coefficients is crucial to the cross-working-condition heterogeneous fault diagnosis performance of WDH-Net, and the combination of 0.6, 0.3, 0.1 is the optimal choice verified by experiments. Table 7Accuracy comparison of WDH-Net under different loss function coefficient combinations.Loss Function Coefficients ()CWRU Dataset Acc (%)PU Dataset Acc (%)0.6, 0.3, 0.194.18±0.5594.85±0.520.7, 0.2, 0.192.85±0.6293.52±0.580.5, 0.4, 0.191.76±0.7092.38±0.650.6, 0.2, 0.293.24±0.5893.96±0.550.5, 0.3, 0.290.92±0.7591.57±0.700.8, 0.1, 0.190.35±0.8191.02±0.76 Accuracy comparison of WDH-Net under different loss function coefficient combinations. This Fig. 4 intuitively presents the convergence characteristics of various losses and the differences between datasets during the training process of WDH-Net. On the whole, all loss function values show a trend of rapid decline followed by gradual flattening with the increase of training epochs, reflecting the effective convergence of model training, and accompanied by slight random fluctuations throughout the process, which is consistent with the real process of gradient descent optimization for deep learning models. As the loss type with the highest weight, the classification loss has an initial value significantly higher than other losses and the fastest convergence rate, decreasing rapidly in the early stage of training and stabilizing at a low level in the later stage, indicating that the model can quickly focus on the core goal of fault classification. The cross-domain alignment loss converges slightly slower, reflecting that the adaptation to cross-working-condition domain distribution differences requires a longer training process, while the feature orthogonality loss has the lowest initial value and stable convergence, indicating that the decoupling process between fault features and working-condition features is easier to stabilize. Comparing the two datasets, it can be seen that all loss values of the PU dataset are consistently higher than those of the CWRU dataset. This is because the PU dataset contains heterogeneous modal signals of vibration and current, with more complex data distribution and higher difficulty in model optimization. At the same time, the fluctuation range of losses in the PU dataset is slightly larger than that in the CWRU dataset, further confirming the complexity of heterogeneous data training. As the weighted sum of the three basic losses, the variation trend of the total loss is highly consistent with the classification loss, verifying the leading role of classification loss in model optimization. Meanwhile, the reasonable proportion of cross-domain alignment loss and feature orthogonality loss also ensures that the total loss can balance the goals of cross-domain adaptation and feature decoupling. Ultimately, all types of losses of the model stabilize at a low level after 100 training epochs, providing a guarantee for the high accuracy of subsequent cross-working-condition mechanical fault diagnosis. Variation trends of various loss function values of WDH-Net with training epochs on CWRU and PU datasets, including the dynamic changes of total loss, classification loss, cross-domain alignment loss and feature orthogonality loss. As shown in Fig. 5, the proposed WDH-Net demonstrates distinct cross-domain alignment capability in the learned shared feature space. In Fig. 5 (a), source-domain and target-domain samples on the CWRU dataset form compact class-wise clusters with significant overlap between the two domains, while normal, inner-race, outer-race, and roller fault categories remain well separated. This indicates that the proposed Wavelet-ConvNet+Cross-Attention and dual-stream decoupling encoder can effectively suppress working-condition interference while retaining discriminative fault-related information, suggesting that the learned fault representations are not only separable across categories but also relatively invariant to domain shifts caused by varying loads. The class-center trajectories in Fig. 5 (b) further illustrate the dynamic adaptation process of WDH-Net on the CWRU dataset: at the early training stage, target-domain class centers are relatively scattered, reflecting the initial discrepancy between source and target feature distributions; as training proceeds, all target centers gradually move toward compact, stable positions with reduced movement amplitudes in later epochs. This trend confirms that the proposed heterogeneous alignment strategy progressively reduces domain discrepancy and improves the stability of class-level representations, while the smooth convergence behavior also implies a well-balanced optimization process that avoids severe oscillation during adaptation. A similar yet more challenging pattern is observed on the PU dataset in Fig. 5 (c). Compared with the CWRU dataset, the PU dataset involves more complex operating-condition variations and heterogeneous modal information, increasing the difficulty of cross-domain transfer. Even under this more demanding setting, source and target samples still exhibit clear class-wise aggregation in the shared feature space; although slight residual offsets remain in several categories—especially those under stronger domain discrepancy—the overall inter-class margin is well preserved, demonstrating that WDH-Net maintains both discriminability and cross-domain consistency in more complex heterogeneous diagnosis scenarios. Fig. 5 (d) provides additional evidence of stable adaptation on the PU dataset through class-center evolution: target-domain centers show a consistent convergence tendency toward compact regions during training, though their trajectories are relatively longer than those on the CWRU dataset. This behavior aligns with the higher difficulty of heterogeneous vibration-current adaptation under coupled speed-load variations, yet the final center locations are still clearly separated, confirming that the proposed heterogeneous common feature mapping space and H-MK-MMD-based alignment mechanism can effectively narrow the source-target gap while preserving fault-specific structure. Overall, Fig. 5 visually verifies the effectiveness of WDH-Net from both sample-level and class-level perspectives. The feature distribution plots confirm that the proposed framework achieves domain-invariant yet class-discriminative representation learning, while the center-evolution plots reveal that the alignment process is progressive, stable, and interpretable. These observations provide intuitive support for the quantitative results reported in Tables 2–4, further demonstrating that WDH-Net improves cross-working-condition fault diagnosis performance by jointly enhancing feature separability and cross-domain consistency. Visualization of cross-domain feature alignment and class-center evolution of WDH-Net in the shared feature space. (a) Feature distribution of source- and target-domain samples on the CWRU dataset after adaptation. (b) Evolution trajectories of target-domain class centers on the CWRU dataset during training. (c) Feature distribution of source- and target-domain samples on the PU dataset after adaptation. (d) Evolution trajectories of target-domain class centers on the PU dataset during training. To quantitatively evaluate the interpretability and feature decoupling effect of WDH-Net, SHAP-based feature importance analysis is conducted. The contribution scores of different feature components are calculated to reflect their influence on diagnostic results. As shown in Table 8, the fault-sensitive features, including the wavelet-based fault features and decoupled fault-related features, have significantly higher importance scores compared to the working condition interference and noise components. This quantitatively demonstrates that WDH-Net successfully enhances the fault-related features and suppresses irrelevant components, which leads to improved model interpretability, reduced domain shift, and enhanced robustness. Table 8SHAP Feature Importance Analysis of WDH-Net.Feature CategoryImportance ScoreWavelet fault feature (vibration)0.36Wavelet fault feature (current)0.29Decoupled fault-related feature0.18Working condition interference feature0.09Environmental noise component0.08 SHAP Feature Importance Analysis of WDH-Net. Ablation experiments are conducted to validate the necessity of WDH-Net core modules, including Wavelet-ConvNet, Cross-Attention, dual-stream decoder, heterogeneous feature fusion, multi-kernel MMD alignment, and dynamic weight assignment. As shown in Table 9, the complete WDH-Net obtains the best overall results on both CWRU and PU datasets, achieving diagnostic accuracy of 94.18% and 94.85% respectively. All metrics including precision, recall and F1-score remain at high levels with standard deviations lower than 0.6, which reflects reliable and stable performance. The elimination of any core module induces an obvious accuracy drop, verifying the indispensable contribution of each component to the overall framework. Without Wavelet-ConvNet, model performance declines most noticeably, with accuracy falling to 89.72% on the CWRU dataset and 90.15% on the PU dataset. Such degradation originates from the inherent function of Wavelet-ConvNet in mining subtle time-frequency fault characteristics from vibration signals, and the lack of this module prevents the network from capturing fine-grained fault patterns and thereby weakens recognition capability. The removal of Cross-Attention also causes observable accuracy drops on both datasets. This module enables the model to concentrate on critical fault-related regions, and its absence reduces feature discriminability and further lowers precision and recall. The dual-stream decoder plays a vital role in isolating working condition components from intrinsic fault representations, and disabling this module aggravates feature coupling and degrades the ability to distinguish diverse fault modes. The absence of heterogeneous feature fusion undermines complementary information integration between vibration and current signals, while removing multi-kernel MMD alignment weakens cross-domain distribution matching across variable operating conditions. Discarding dynamic weight assignment further cancels adaptive modal contribution adjustment, leading to decreased diagnostic accuracy and stability. All ablation outcomes indicate that the involved components cooperate closely with one another. Effective fault feature extraction, key region enhancement, mixed feature decoupling, multi-modal information integration, robust cross-domain alignment and adaptive feature selection are jointly guaranteed by these modules, and their synergistic effect enables WDH-Net to maintain superior diagnostic performance on different bearing fault datasets. Table 9Ablation experiment results of WDH-Net on CWRU and PU datasets.Model (WDH-Net Variants)CWRU DatasetPU DatasetAcc (%)P (%)R (%)F1 (%)Acc (%)P (%)R (%)F1 (%)WDH-Net (Full)94.18 ± 0.5593.87 ± 0.5093.72 ± 0.5893.79 ± 0.4894.85 ± 0.5294.53 ± 0.4694.38 ± 0.5494.45 ± 0.42W/O Wavelet-ConvNet89.72 ± 0.8889.45 ± 0.8589.21 ± 0.9089.33 ± 0.8790.15 ± 0.9189.87 ± 0.8989.62 ± 0.9389.75 ± 0.90W/O Cross-Attention91.35 ± 0.7291.08 ± 0.7090.92 ± 0.7491.00 ± 0.7291.88 ± 0.7591.56 ± 0.7391.41 ± 0.7691.48 ± 0.74W/O Dual-stream Decoupling Encoder90.68 ± 0.7990.41 ± 0.7790.25 ± 0.8190.33 ± 0.7991.24 ± 0.8290.95 ± 0.8090.78 ± 0.8390.86 ± 0.81W/O Heterogeneous Fusion91.84 ± 0.7091.57 ± 0.6791.41 ± 0.7291.49 ± 0.6992.31 ± 0.7392.02 ± 0.7091.86 ± 0.7591.94 ± 0.72W/O Multi-kernel MMD Alignment91.12 ± 0.7690.85 ± 0.7390.69 ± 0.7890.77 ± 0.7591.66 ± 0.7991.37 ± 0.7691.21 ± 0.8191.29 ± 0.78W/O Dynamic Weight Assignment92.35 ± 0.6692.09 ± 0.6391.93 ± 0.6892.01 ± 0.6592.87 ± 0.6992.58 ± 0.6692.42 ± 0.7192.50 ± 0.68 Ablation experiment results of WDH-Net on CWRU and PU datasets. Discussion The WDH-Net model proposed in this paper effectively breaks through the bottlenecks of traditional methods in working condition adaptation and heterogeneous signal processing for the core problem of cross-working-condition heterogeneous mechanical fault diagnosis, via the deep integration of time-frequency feature enhancement, dual-stream decoupling encoding and heterogeneous feature alignment. Its high diagnostic accuracy and stability have been verified on the public CWRU and PU datasets, providing a new solution for cross-working-condition fault diagnosis. The proposed dual-stream decoupling structure successfully separates fault-relevant features from working condition interference, and the heterogeneous multi-kernel MMD enables robust distribution alignment between vibration and current signals, which reveals the importance of feature disentanglement and adaptive cross-domain alignment for real industrial diagnosis tasks. However, this study still has several limitations to be improved: the model training is highly dependent on publicly available datasets with complete annotations, while mechanical fault data in actual industrial scenarios often face problems such as high annotation costs, unbalanced samples and small sample sizes; its generalization ability in such real industrial scenarios has not been fully verified, and direct migration and application may lead to performance degradation; additionally, the model is validated under controlled laboratory conditions, without considering severe sensor drift, intermittent noise, or non-stationary operating shocks that are common in real industrial sites. While WDH-Net has achieved lightweight design with only 11.3M parameters, the computational logic of the wavelet packet decomposition module and cross-attention mechanism still has room for optimization, and its real-time inference efficiency on industrial embedded devices is difficult to meet the requirements of some high-frequency acquisition scenarios; the model also does not consider computational resource constraints on edge devices with limited memory and computing power, which may restrict its deployment in online monitoring systems. This study only focuses on the fusion of two types of heterogeneous signals (vibration and current), and does not incorporate more common types of sensing data in industrial scenarios such as temperature, pressure and acoustic emission, leaving the feature complementarity and fusion strategy of multi-source heterogeneous data under-explored. To date, the model has only completed offline experimental verification, without combining with the actual deployment environment of industrial sites, and engineering issues such as data transmission delay, device compatibility and environmental noise interference have not been taken into account; its scalability to multi-component rotating machinery (e.g., gearboxes, bearings, motors) and variable-speed non-stationary conditions also remains insufficiently validated. Based on the above limitations, future research work can be further expanded from multiple dimensions: combined with semi-supervised learning, self-supervised learning or meta-learning methods, a cross-working-condition fault diagnosis framework for few-annotation and small-sample scenarios can be built, reducing dependence on manually annotated data by mining features of unannotated data and improving the model’s adaptability in real industrial scenarios. To address the problem of model deployment efficiency, WDH-Net can be optimized using lightweight technologies such as quantization, pruning and knowledge distillation, simplifying the computational process of wavelet packet decomposition and exploring hardware-accelerated inference schemes to achieve low-latency and low-power deployment on embedded terminals. The fusion dimension of multi-source heterogeneous sensing data can be expanded by integrating multi-type industrial sensing data such as temperature and pressure, designing an adaptive modal weight allocation mechanism that dynamically adjusts feature weights according to the effectiveness of each modal data under different equipment and working conditions, thereby fully tapping the complementary value of multi-source data. The model can be promoted from offline verification to engineering application by building an end-edge-cloud collaborative online fault diagnosis system combined with the industrial internet platform, realizing a closed-loop full process of data collection, real-time analysis and fault early warning, accumulating actual fault data from industrial sites, and establishing a dynamically updated model iteration mechanism to continuously improve the engineering practicality and robustness of the model. Future research will also focus on improving the model’s generalization under extreme noise, sensor failure, and imbalanced fault data, and exploring its generalization ability across different mechanical components to meet the diverse requirements of industrial predictive maintenance. Conclusion The WDH-Net model proposed in this paper effectively addresses the core problems of heterogeneous signal adaptation and working condition distribution shift in cross-working-condition mechanical fault diagnosis by integrating time-frequency feature enhancement, dual-stream decoupling encoding and heterogeneous feature alignment mechanisms. This model exhibits superior diagnostic performance compared with existing methods on the CWRU and PU datasets, balancing high accuracy and lightweight characteristics, and thus provides an effective solution for intelligent fault diagnosis of machinery under complex working conditions. The relevant research results can provide theoretical reference and practical basis for the research and development of intelligent operation and maintenance technology for industrial mechanical equipment. Author contributions Honglian Xiao conceptualized the research, developed the methodology, and performed the primary data analysis. Jianhua Xiao provided critical insights into the experimental design, conducted supplementary analysis, and helped refine the manuscript. Both authors reviewed and approved the final manuscript. Funding Not applicable. Data availability The datasets used in this study are publicly available and can be accessed through the following links: CWRU Dataset (https://www.kaggle.com/datasets/astrollama/cwru-case-western-reserve-university-dataset) and PU Dataset: Paderborn University Dataset (https://www.kaggle.com/datasets/dippatel03/paderborn-db) 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. Contributor Information Honglian Xiao, Email: 18577399986@163.com. Jianhua Xiao, Email: xiaojianhua@hnu.edu.cn. References Associated Data Data Availability Statement The datasets used in this study are publicly available and can be accessed through the following links: CWRU Dataset (https://www.kaggle.com/datasets/astrollama/cwru-case-western-reserve-university-dataset) and PU Dataset: Paderborn University Dataset (https://www.kaggle.com/datasets/dippatel03/paderborn-db)