Single-source-domain generalization motor fault diagnosis method, system and device based on cochlea perception enhancement and parameter space integration and medium

Abstract

The invention relates to a single-source-domain generalization motor fault diagnosis method, system and device based on cochlea perception enhancement and parameter space integration and a medium, and the method comprises the steps: carrying out the multiband modal decomposition of an original vibration signal, and generating a two-dimensional time-frequency image through the combination of Mel frequency mapping; disturbance is applied to a frequency domain amplitude spectrum of the two-dimensional time-frequency image, a phase spectrum is kept unchanged, and an enhanced image is generated through time domain reconstruction and dynamic range normalization; constructing a fault diagnosis model, inputting the two-dimensional time-frequency image into a main branch for fault feature modeling, and inputting an enhanced image into an auxiliary branch to learn potential features so as to update parameters of the auxiliary branch; parameters of the auxiliary branch are obtained through periodic snapshot sampling and fused to the main branch to form a fault diagnosis model. The robustness and generalization ability of a single-source domain fault diagnosis model are improved through the synergistic effect of signal representation optimization inspired by biological perception, physical consistency disturbance enhancement and a double-branch progressive knowledge fusion mechanism.

Description

Technical Field

[0001] This invention relates to the field of motor fault diagnosis technology, and in particular to a single-source domain generalized motor fault diagnosis method, system, device, and medium based on cochlear perception enhancement and parameter space integration. Background Technology

[0002] Motor fault diagnosis (FD) is a crucial component of intelligent industrial operation and maintenance, and its accuracy directly impacts the safe operation and maintenance efficiency of production systems. In recent years, deep learning technology has made significant progress in mechanical condition identification tasks due to its powerful nonlinear feature extraction capabilities. However, this technology faces fundamental limitations in practical applications: due to the dynamic changes in the operating conditions of industrial equipment, there is a general distributional offset between the labeled data obtained during the training phase and the actual operating environment. This inter-domain difference significantly reduces the generalization performance of diagnostic models built based on deep learning under unknown operating conditions, severely limiting its deployment value in complex industrial scenarios.

[0003] To address the issue of data distribution offset, existing technologies primarily employ domain adaptation methods, aligning the feature spaces of the source and target domains through adversarial learning or distribution metrics. However, the high deployment costs stem from the need to continuously acquire target domain samples and retrain the model for each new operating condition. While domain generalization techniques do not require the target domain, multi-source domain solutions necessitate joint training on multi-condition data. Since industrial equipment typically operates under fixed baseline conditions, acquiring sufficient source domain data with significant distributional differences is challenging, limiting practical applications.

[0004] Existing single-source domain generalization schemes attempt to simulate potential distributional differences by constructing pseudo-domain samples through data augmentation strategies, such as using random feature perturbations or adversarial generative networks to expand the training data distribution. However, due to the lack of effective modeling of the frequency domain characteristics and physical evolution laws of mechanical fault signals, the generated samples may destroy the key feature structures closely related to the fault mode in the original signal. This non-physically constrained augmentation method is prone to introducing pseudo-features unrelated to the actual working conditions, which may introduce noise and damage the integrity of fault features, leading to a decrease in the stability of the feature learning process and fluctuations in model performance. Summary of the Invention

[0005] (a) Technical problems to be solved In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a single-source domain generalized motor fault diagnosis method, system, device and medium based on cochlear perception enhancement and parameter space integration. It solves the technical problem that the existing single-source domain generalization method has limited model generalization ability because it does not model the physical evolution characteristics of mechanical fault signals, which leads to the destruction of key feature structures in the data enhancement process, easy introduction of noise and damage to feature integrity.

[0006] (II) Technical Solution To achieve the above objectives, the main technical solutions adopted by the present invention include: In a first aspect, embodiments of the present invention provide a single-source domain generalized motor fault diagnosis method based on cochlear perception enhancement and parameter space integration, comprising: The acquired original vibration signal was subjected to multi-band mode decomposition based on the spiral cochlear structure to obtain the cochlear time spectrum; The Mel frequency mapping mechanism is introduced to reconstruct the frequency axis of the cochlear time spectrum in a perception-driven manner, generating a perceptually continuous two-dimensional time-frequency image; The frequency domain amplitude spectrum of the two-dimensional time-frequency image is perturbed while the phase spectrum remains unchanged. After time-domain reconstruction and dynamic range normalization, a physically consistent enhanced image is generated. A two-branch fault diagnosis model is constructed. Two-dimensional time-frequency images are input into the main branch to model fault features, and enhanced images are input into the auxiliary branch. Representation alignment regularization is used to learn latent features that are semantically consistent with the fault features of the main branch in order to update the parameters of the auxiliary branch. The parameters of the auxiliary branches are obtained by periodic snapshot sampling and gradually merged into the parameter space of the main branch to form a progressively optimized diagnostic model, which is used to perform motor fault diagnosis tasks.

[0007] Optionally, the acquired original vibration signal is subjected to multi-band mode decomposition based on the spiral cochlear structure to obtain the cochlear time spectrum, including: Based on the spatial distribution characteristics of the spiral cochlear structure, a Cochlear wavelet filter bank is constructed along a preset angle domain. Each wavelet filter corresponds to a preset angle position of the cochlea and has spatial locality and frequency response characteristics, which can be used for multi-band mode decomposition of signals. For the signal components in the acquired original vibration signal that are outside the cochlear sensing frequency range, an analytic signal is generated through Hilbert transform, and then multiplied with the modulated carrier to generate a frequency-shifted modulated propagation wave. Using a Cochlear wavelet filter bank, the modulated propagating wave is projected onto spatially supported Cochlear wavelet modes, and the response coefficients of each mode are calculated by inner product, where different modes represent the local frequency response at different angular positions of the cochlea; Hilbert analysis was performed on the response coefficients to obtain the instantaneous amplitude and phase information of each mode, and the cochlear temporal spectrum with color intensity proportional to amplitude was generated based on the amplitude-intensity mapping.

[0008] Optionally, a Mel frequency mapping mechanism is introduced to reconstruct the frequency axis of the cochlear time-frequency spectrum in a perception-driven manner, generating a perceptually continuous two-dimensional time-frequency image, including: Construct a frequency mapping relationship based on Mel perception characteristics, expand the resolution of linear frequencies below the preset threshold frequency, and perform redundant compression on linear frequencies above the threshold frequency to generate a nonlinear perception frequency scale aligned with the range of human hearing. Based on the nonlinear sensing frequency scale, a set of overlapping triangular bandpass filter banks is obtained by uniformly dividing the corresponding Mel frequency axis. Each triangular bandpass filter covers the corresponding sensing frequency range, and the boundaries of adjacent triangular bandpass filters overlap to achieve sensing continuity coverage of the frequency axis. The power spectral component of the cochlear frequency spectrum is input into a bandpass filter bank. The spectral energy is weighted and calculated through each filter to extract the perceptual alignment energy value of the corresponding Mel frequency band. The perceptual alignment energy values ​​of each Mel frequency band are spliced ​​and stacked along the time dimension to generate a two-dimensional time-frequency image that integrates cochlear modal decomposition characteristics and Mel perceptual consistency.

[0009] Optionally, each triangular bandpass filter has the following form: ; in, Let f(m) be the response value of the m-th triangular bandpass filter at frequency index n, where n is the frequency index of the power spectrum, and f(m) is the center frequency of the m-th filter, which is uniformly divided by the Mel frequency.

[0010] Optionally, by perturbing the frequency domain amplitude spectrum of the two-dimensional time-frequency image while keeping the phase spectrum unchanged, and then performing time-domain reconstruction and dynamic range normalization, a physically consistent enhanced image is generated, including: A fast Fourier transform is performed on a two-dimensional time-frequency image to separate the amplitude spectrum and phase spectrum in the frequency domain. Keeping the phase spectrum unchanged, a perturbation of controllable intensity is applied to the amplitude spectrum to generate a perturbed amplitude spectrum. The perturbed amplitude spectrum is then combined with the phase spectrum to form an enhanced frequency domain representation. Perform an inverse fast Fourier transform on the enhanced frequency domain representation to reconstruct the time-domain enhanced image; The reconstructed temporal enhanced image is subjected to dynamic range constraints to align the dynamic range of the amplitude spectrum of the reconstructed temporal enhanced image with that of the two-dimensional time-frequency image. This is supplemented by a structure-preserving spatial domain enhancement operation to generate a physically consistent enhanced image.

[0011] Optionally, a two-branch fault diagnosis model is constructed. Two-dimensional time-frequency images are input into the main branch for fault feature modeling, and enhanced images are input into the auxiliary branch. Representation alignment regularization is used to learn latent features semantically consistent with the fault features of the main branch to update the parameters of the auxiliary branch, including: Construct a two-branch fault diagnosis model containing a main branch and an auxiliary branch with the same structure. The encoder and classification head parameters of the main branch are kept frozen, while the encoder and classification head parameters of the auxiliary branch are trainable and initialized in the same way as the main branch. Introduce a shared projection head at the back end of the encoder for both the main and auxiliary branches; Two-dimensional time-frequency images are input into the main branch for fault feature modeling, and enhanced images are input into the auxiliary branch. By sharing the projection head, the encoding features of the two-dimensional time-frequency images by the main branch and the encoding features of the enhanced images by the auxiliary branch are mapped to the same latent space. By maximizing the mutual information of latent features and suppressing redundant representations, representation alignment regularization constraints are constructed. By representing alignment regularization constraints, the auxiliary branch is driven to learn latent features that are semantically consistent with the fault features of the main branch, and classification prediction results are generated based on the latent features. The cross-entropy loss of the classification prediction results of the auxiliary branch on the enhanced image is calculated, and the representation alignment regularization constraint is superimposed as a joint optimization objective to update the encoder and classification head parameters of the auxiliary branch.

[0012] Optionally, parameters of the auxiliary branch are obtained through periodic snapshot sampling and gradually merged into the parameter space of the main branch to form a progressively optimized diagnostic model, which is used to perform motor fault diagnosis tasks, including: During the training of the auxiliary branch, snapshots of the encoder and classification head parameters of the auxiliary branch are taken at preset intervals to generate a set of phased parameter states. The average value of multiple historical parameter snapshots in the stage parameter state set is calculated to generate integrated parameters; The integrated parameters are injected into the parameter space of the main branch, replacing the encoder and classification head parameters of the main branch, thus forming the updated main branch; By periodically synchronizing the parameter snapshots of the auxiliary branches with the parameter space of the main branch, parameter knowledge from different training stages is gradually integrated to form a progressively optimized diagnostic model. The diagnostic model is solidified into a format that the inference engine can load, and then transmitted and deployed to the edge computing unit or cloud monitoring platform of the target electromechanical equipment. In response to the task instruction, the input raw vibration signal is decomposed into multi-band mode based on the spiral cochlear structure to obtain the cochlear time spectrum. Then, with the help of the Mel frequency mapping mechanism, the frequency axis of the cochlear time spectrum is reconstructed in a perception-driven manner to generate a perceptually continuous two-dimensional time-frequency image. The diagnostic results are obtained by performing inference calculations on the two-dimensional time-frequency images through the model deployed on the edge computing unit or cloud monitoring platform of the target electromechanical equipment.

[0013] Secondly, embodiments of the present invention provide a single-source domain generalized motor fault diagnosis system based on cochlear perception enhancement and parameter space integration, comprising: The mode decomposition module is used to perform multi-band mode decomposition based on the spiral cochlear structure on the acquired raw vibration signal to obtain the cochlear time spectrum.

[0014] The frequency axis reconstruction module is used to introduce the Mel frequency mapping mechanism to reconstruct the frequency axis of the cochlear time spectrum in a perception-driven manner, generating a perceptually continuous two-dimensional time-frequency image.

[0015] The image enhancement module is used to perturb the frequency domain amplitude spectrum of a two-dimensional time-frequency image while keeping the phase spectrum unchanged. After time-domain reconstruction and dynamic range normalization, a physically consistent enhanced image is generated.

[0016] The model training module is used to construct a two-branch fault diagnosis model. It inputs two-dimensional time-frequency images into the main branch to model fault features, and inputs enhanced images into the auxiliary branch. It uses representation alignment regularization to learn latent features that are semantically consistent with the fault features of the main branch, so as to update the parameters of the auxiliary branch.

[0017] The model output module is used to obtain the parameters of the auxiliary branch through periodic snapshot sampling and gradually merge them into the parameter space of the main branch to form a progressively optimized diagnostic model for performing motor fault diagnosis tasks.

[0018] Thirdly, embodiments of the present invention provide a single-source domain generalized motor fault diagnosis device based on cochlear perception enhancement and parameter space integration, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the single-source domain generalized motor fault diagnosis method based on cochlear perception enhancement and parameter space integration as described above.

[0019] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the single-source domain generalized motor fault diagnosis method based on cochlear perception enhancement and parameter space integration as described above.

[0020] (III) Beneficial Effects The beneficial effects of this invention are: First, by simulating the spiral cochlear structure of the human auditory system, a cochlear signal decomposition method with multi-scale analysis capabilities is used to enhance the low-frequency representation of the signal and extract the cochlear time spectrum with physical meaning. Furthermore, the Mel frequency mapping mechanism is introduced to reconstruct the frequency axis distribution in a perception-driven manner, transforming the time spectrum into a perceptually continuous two-dimensional time-frequency image with stronger discriminative and generalization capabilities. This significantly enhances the characterization ability of fault features and avoids the feature loss problem caused by high-frequency noise interference in traditional time-frequency transformation methods.

[0021] Secondly, while maintaining the phase spectrum unchanged, a random perturbation is applied to the frequency domain amplitude spectrum, and an enhanced image is generated through inverse Fourier transform and dynamic range normalization. This strategy introduces statistical diversity while ensuring that the enhanced image remains consistent with the original signal in key physical semantics such as temporal impulse characteristics and frequency domain resonant modes, effectively avoiding the damage to feature integrity caused by noise accumulation.

[0022] Next, a two-branch structure with a main branch and an auxiliary branch working together is designed. The main branch models fault features based on two-dimensional time-frequency images, while the auxiliary branch learns from the augmented images and uses representation alignment regularization to constrain its latent features to be semantically consistent with the main branch. This mechanism explores the generalization features of the augmented data through the auxiliary branch and prevents the auxiliary branch from deviating from the physically real fault modes through regularization constraints, thereby improving diversity while suppressing spurious feature interference.

[0023] Furthermore, by periodically collecting parameter snapshots of auxiliary branches through a parameter space integration mechanism, calculating the historical parameter mean, and then injecting it into the main branch, a progressively optimized model parameter update strategy is formed. This design transforms the parameter evolution process of the exploration branch into a stable knowledge increment, avoiding the generalization problem caused by directly training the main branch, and ultimately constructing a progressive fault diagnosis model that balances anti-interference capability and diagnostic accuracy.

[0024] Therefore, this invention achieves stable knowledge fusion through biosensor-inspired signal representation optimization, physical constraint enhancement to ensure semantic consistency, bi-branch regularization to suppress noise propagation, and parameter integration to form a full-link collaborative mechanism. This significantly solves the problems of feature distortion and generalization performance degradation caused by excessive reliance on random enhancement in traditional single-source domain generalization methods, and demonstrates higher accuracy and cross-domain robustness in fault diagnosis tasks under complex working conditions. Attached Figure Description

[0025] Figure 1 A flowchart illustrating the method provided in an embodiment of the present invention; Figure 2 A schematic diagram of the CoMel spectrum of the method provided in the embodiments of the present invention; Figure 3This is a schematic diagram illustrating the specific process of step S1 of the method provided in this embodiment of the invention; Figure 4 This is a detailed flowchart illustrating step S2 of the method provided in this embodiment of the invention; Figure 5 A schematic diagram of the frequency domain enhancement framework of the method provided in the embodiments of the present invention; Figure 6 This is a detailed flowchart illustrating step S3 of the method provided in this embodiment of the invention; Figure 7 This is a detailed flowchart illustrating step S4 of the method provided in this embodiment of the invention; Figure 8 A schematic diagram of a robust feature learning framework based on a dual-branch structure provided in the embodiments of the present invention; Figure 9 This is a detailed flowchart illustrating step S5 of the method provided in this embodiment of the invention; Figure 10 A schematic diagram of the experimental system for the method provided in the embodiments of the present invention; Figure 11 A schematic diagram of the structure corresponding to multiple motor states in the method provided in the embodiment of the present invention; Figure 12 Visualization results of two-dimensional time-frequency images under five operating conditions (A, C, F, H, O) for the method provided in the embodiments of the present invention; Figure 13 The image representation sample is provided under condition A of the method provided in the embodiment of the present invention. Detailed Implementation

[0026] To better explain and facilitate understanding of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0027] like Figure 1As shown in the embodiment of the present invention, a single-source domain generalized motor fault diagnosis method based on cochlear perception enhancement and parameter space integration includes: performing multi-band mode decomposition based on the spiral cochlear structure on the acquired original vibration signal to obtain the cochlear time spectrum; introducing the Mel frequency mapping mechanism to reconstruct the frequency axis of the cochlear time spectrum in a perception-driven manner to generate a perceptually continuous two-dimensional time-frequency image; applying a perturbation to the frequency domain amplitude spectrum of the two-dimensional time-frequency image while keeping the phase spectrum unchanged, and generating a physically consistent enhanced image after time domain reconstruction and dynamic range normalization; constructing a dual-branch fault diagnosis model, inputting the two-dimensional time-frequency image into the main branch for fault feature modeling, and inputting the enhanced image into the auxiliary branch, using representation alignment regularization to learn latent features that are semantically consistent with the fault features of the main branch to update the parameters of the auxiliary branch; obtaining the parameters of the auxiliary branch through periodic snapshot sampling and gradually fusing them into the parameter space of the main branch to form a progressively optimized fault diagnosis model for performing motor fault diagnosis tasks.

[0028] First, by simulating the spiral cochlear structure of the human auditory system, a cochlear signal decomposition method with multi-scale analysis capabilities is used to enhance the low-frequency representation of the signal and extract the cochlear time spectrum with physical meaning. Furthermore, the Mel frequency mapping mechanism is introduced to reconstruct the frequency axis distribution in a perception-driven manner, transforming the time spectrum into a perceptually continuous two-dimensional time-frequency image with stronger discriminative and generalization capabilities. This significantly enhances the characterization ability of fault features and avoids the feature loss problem caused by high-frequency noise interference in traditional time-frequency transformation methods.

[0029] Secondly, while maintaining the phase spectrum unchanged, a random perturbation is applied to the frequency domain amplitude spectrum, and an enhanced image is generated through inverse Fourier transform and dynamic range normalization. This strategy introduces statistical diversity while ensuring that the enhanced image remains consistent with the original signal in key physical semantics such as temporal impulse characteristics and frequency domain resonant modes, effectively avoiding the damage to feature integrity caused by noise accumulation.

[0030] Next, a two-branch structure with a main branch and an auxiliary branch working together is designed. The main branch models fault features based on two-dimensional time-frequency images, while the auxiliary branch learns from the augmented images and uses representation alignment regularization to constrain its latent features to be semantically consistent with the main branch. This mechanism explores the generalization features of the augmented data through the auxiliary branch and prevents the auxiliary branch from deviating from the physically real fault modes through regularization constraints, thereby improving diversity while suppressing spurious feature interference.

[0031] Furthermore, by periodically collecting parameter snapshots of auxiliary branches through a parameter space integration mechanism, calculating the historical parameter mean, and then injecting it into the main branch, a progressively optimized model parameter update strategy is formed. This design transforms the parameter evolution process of the exploration branch into a stable knowledge increment, avoiding the generalization problem caused by directly training the main branch, and ultimately constructing a progressive fault diagnosis model that balances anti-interference capability and diagnostic accuracy.

[0032] Therefore, this invention achieves stable knowledge fusion through biosensor-inspired signal representation optimization, physical constraint enhancement to ensure semantic consistency, bi-branch regularization to suppress noise propagation, and parameter integration to form a full-link collaborative mechanism. This significantly solves the feature distortion and generalization performance degradation problems caused by excessive reliance on random augmentation in traditional single-source domain generalization methods, demonstrating higher accuracy and cross-domain robustness in fault diagnosis tasks under complex conditions. To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present invention can be understood more clearly and thoroughly, and that the scope of the present invention can be fully conveyed to those skilled in the art.

[0033] Specifically, embodiments of the present invention provide a single-source domain generalized motor fault diagnosis method based on cochlear perception enhancement and parameter space integration, including: S1. Perform multi-band mode decomposition based on the spiral cochlear structure on the acquired original vibration signal to obtain the cochlear time spectrum.

[0034] This invention proposes a two-dimensional time-frequency map construction method—CoMel image—that integrates biological sensing mechanisms and perceptual spectrum optimization to improve the signal's expressive power and cross-domain robustness in deep models. This representation method draws on the frequency selectivity of the human auditory system, combining the nonlinear multi-scale modeling advantages of the Cochlear Transform (CT) with the perceptual consistency of Mel frequency mapping. Figure 2 It can construct two-dimensional image representations with clear structure, comprehensive feature perception, and good robustness to non-stationary signals.

[0035] The construction of CoMel images mainly includes two core stages: First, multi-band modal decomposition based on the spiral cochlear structure is achieved through biomimetic modeling; second, the CT spectrum is resampled using the Mel scale to improve the perceptual continuity and structural consistency of the frequency axis distribution. The technical principles and modeling mechanisms of these two parts will be explained below.

[0036] Furthermore, such as Figure 3As shown, step S1 includes: S11. Based on the spatial distribution characteristics of the spiral cochlear structure, a spatially supported Cochlear wavelet filter bank is constructed along a preset angle domain. Each wavelet filter corresponds to a set angular position of the cochlea and has spatial locality and frequency response characteristics for use in multi-band mode decomposition of signals.

[0037] It's important to understand that cochlear transformation is a biomimetic time-frequency analysis method inspired by the structure and dynamics of the spiral cochlea in the human inner ear. Its basic concept is to treat the input signal as a pressure wave propagating within the spiral cochlea, along the angular domain... A set of spatially supported Cochlear wavelet filters is constructed to extract frequency-selective modes from the signal in a biomimetic manner. Each wavelet corresponds to a specific angular position on the spiral cochlea, exhibiting spatial locality and frequency response characteristics. Its center frequency is determined by the cochlear frequency-position mapping function. (1) This mapping function reflects the cochlear system's high resolution of low-frequency signals and its sparse perception of high-frequency signals, enabling the cochlear transformation method to exhibit stronger weak feature extraction capabilities in the low-frequency band.

[0038] S12. For the signal components in the acquired original vibration signal that are outside the cochlear sensing frequency range, generate an analytical signal through Hilbert transform, and multiply it with the modulated carrier to shift the high-frequency components to the cochlear sensing frequency range, generating a modulated propagation wave after frequency shift.

[0039] To accommodate the cochlear structure's sensing frequency range [10.39 Hz, 14.57 kHz], signal components exceeding this frequency band need to be converted to a perceptible range using frequency modulation techniques. That is, for an input signal x(t) whose frequency components exceed the frequency range covered by CT, where t represents a time point, an analytical expression is first obtained through Hilbert transform: (2) in, The Hilbert transform operator is then used in conjunction with the modulated carrier. Multiply, where f m The frequency is 1000Hz, which is an empirical parameter value, and j is the imaginary unit. The generated modulated propagation wave p(t) is: (3) S13. Using a Cochlear wavelet filter bank, the modulated propagating wave is projected onto the spatially supported Cochlear wavelet modes, and the response coefficients of each mode are calculated by inner product, where different modes represent the local frequency response at different angular positions of the cochlea.

[0040] S14. Perform Hilbert analysis on the response coefficients to obtain the instantaneous amplitude and phase information of each mode, and generate the cochlear temporal spectrum with color intensity proportional to amplitude based on amplitude intensity mapping.

[0041] Next, the modulated propagating wave p(t) is projected onto k spatially supported Cochlear wavelet modes. Different modalities represent the local frequency response at different angular positions of the cochlea. The CT response coefficients for each modality... Calculated using the inner product form: (4) Further Hilbert analysis can then be used to obtain the instantaneous amplitude A of each mode. k (t) and phase : (5) This allows for the construction of the original cochlear time-to-time spectrum (CT-TFR), whose color intensity and amplitude A are correlated. k (t) is directly proportional, as follows: (6) S2. Introduce the Mel frequency mapping mechanism to reconstruct the frequency axis of the cochlear time spectrum in a perception-driven manner, and generate a perceptually continuous two-dimensional time-frequency image.

[0042] It is important to emphasize that although CT possesses strong modal modeling capabilities, the uneven spectral distribution in its output images results in discontinuous local textures. This discontinuity can lead to feature misidentification and decreased generalization ability in deep learning models. To address these issues, this invention introduces a Mel frequency mapping mechanism to reconstruct the frequency axis of CT-TFR in a perceptually driven manner, resulting in a more perceptually continuous and structurally consistent two-dimensional time-frequency image, termed a CoMel image (Cochlear-Mel Spectrogram).

[0043] Furthermore, such as Figure 4 As shown, step S2 includes: S21. Construct a frequency mapping relationship based on Mel-sensing characteristics, perform resolution expansion on linear frequencies below a preset threshold frequency, and perform redundancy compression on linear frequencies above the threshold frequency to generate a nonlinear sensing frequency scale aligned with the range of human hearing. This frequency mapping relationship based on Mel-sensing characteristics is implemented by defining a nonlinear transformation function as follows: (7) in, For linear frequency, It is a unit of sensing frequency. Its core design principle is: using a preset boundary frequency as a critical point, resolution expansion is implemented for low-frequency components below this value to enhance the significance of weak signals; redundancy compression is performed on high-frequency regions above the boundary frequency to suppress interference information in non-sensitive frequency bands and improve representation robustness.

[0044] S22. Based on the nonlinear sensing frequency scale, a set of overlapping triangular bandpass filter banks is obtained by uniformly dividing the corresponding Mel frequency axis. Each triangular bandpass filter covers the corresponding sensing frequency range, and the boundaries of adjacent triangular bandpass filters overlap smoothly to achieve continuous sensing coverage of the frequency axis.

[0045] In implementation, Mel mapping performs spectral resampling through a set of triangular bandpass filters (Mel Filter Banks). Each triangular bandpass filter H... m (n) has the following form: (8) in, Let f(m) be the response value of the m-th Mel triangular bandpass filter at frequency index n, where n is the frequency index of the power spectrum, and f(m) is the center frequency of the m-th filter, which is uniformly divided by the Mel frequency. In this embodiment, the filter is uniformly divided into 60 filters, and adjacent filters overlap to form a smooth and continuous frequency coverage.

[0046] S23. Input the power spectrum component of the cochlear time spectrum into the bandpass filter bank, and perform weighted calculation of the spectrum energy through each filter to extract the perceptual alignment energy value of the corresponding Mel frequency band.

[0047] S24. The perceptual alignment energy values ​​of each Mel frequency band are spliced ​​and stacked along the time dimension to generate a two-dimensional time-frequency image that integrates the cochlear modal decomposition characteristics and Mel perceptual consistency.

[0048] Finally, after weighted calculation using all Mel filters, the resulting Mel spectrum is: (9) in, S ( l ) is the first l The energy of one Mel frequency band, H m (f) represents the weight of the m-th Mel filter at frequency f, determined by historical data. These energies are stacked over time to form a two-dimensional time-frequency image. This image combines the advantages of cochlear transformation modal structure modeling with the perceptual scale consistency of Mel mapping, retains the nonlinear decomposition capability of CT for low-frequency modes, and significantly improves image texture continuity, which is beneficial for perceptual modeling of neural networks. Figure 12As shown, the two-dimensional time-frequency image exhibits continuous texture features similar to ocean waves, which intuitively reflects the time-frequency evolution characteristics of the signal and is an ideal input representation for subsequent modeling.

[0049] S3. Perturb the frequency domain amplitude spectrum of the two-dimensional time-frequency image while keeping the phase spectrum unchanged. After time-domain reconstruction and dynamic range normalization, a physically consistent enhanced image is generated.

[0050] To improve the generalization performance of the model under unknown conditions, this invention proposes a frequency-domain perturbation-driven structure-preserving 2D image enhancement strategy, such as... Figure 5 As shown. This strategy introduces small perturbations in the frequency distribution of the image by applying controlled perturbations in the frequency domain, while maintaining the image structure and perceptual semantics, thereby generating diverse and physically consistent pseudo-domain samples. This strategy is based on the Fast Fourier Transform, and the specific process is as follows: Furthermore, such as Figure 6 As shown, step S3 includes: S31. Perform a fast Fourier transform on the two-dimensional time-frequency image to separate the amplitude spectrum and phase spectrum in the frequency domain.

[0051] First, let the input image I(x,y) be of size M×N, and its two-dimensional discrete Fourier transform is defined as follows: (10) Here, F(u,y) represents the frequency domain, where u and v are frequency coordinates, and j is the imaginary unit. The frequency domain representation can be further decomposed into an amplitude spectrum. Phase spectrum Combinations: (11) (12) (13) In the formula, and They represent The real and imaginary parts.

[0052] S32. Keeping the phase spectrum unchanged, apply a perturbation of controllable intensity to the amplitude spectrum to generate the perturbed amplitude spectrum. Combine the perturbed amplitude spectrum with the phase spectrum to form the enhanced frequency domain representation.

[0053] The phase spectrum determines the structure and geometry of an image at the perceptual level, while the amplitude spectrum mainly affects the texture and contrast of the image. Therefore, this invention keeps the phase spectrum unchanged during image enhancement and only applies additive perturbation to the amplitude spectrum to achieve statistical fine-tuning of the image texture distribution without destroying its structural properties.

[0054] The specific perturbation strategy is defined as follows: (14) in, It is the enhanced amplitude spectrum. It is a proportional factor that controls the intensity of the disturbance. From uniform distribution The random variables extracted are shown. The scaling factor is typically chosen within the range [0, 0.5] to balance perturbation diversity and semantic consistency, and is generally set to 0.4. The perturbed amplitude spectrum is then analyzed. Compared with the original phase spectrum Recombining them yields the enhanced frequency domain representation. : (15) S33. Perform an inverse fast Fourier transform on the enhanced frequency domain representation to reconstruct a time-domain enhanced image.

[0055] Then, the reconstructed image is obtained by applying the two-dimensional inverse fast Fourier transform. : (16) S34. Apply dynamic range constraints to the reconstructed time-domain enhanced image, align the dynamic range of the amplitude spectrum of the reconstructed time-domain enhanced image with that of the two-dimensional time-frequency image, and supplement with structure-preserving spatial domain enhancement operations to generate a physically consistent enhanced image.

[0056] To prevent artifacts or amplitude explosion introduced by frequency perturbations, this embodiment of the invention performs normalization processing on the amplitude spectrum and reconstructed image after frequency domain perturbation, ensuring that the dynamic range of the enhanced image is consistent with the original. Figure 1 This facilitates stable model reception and training. Furthermore, the invention employs a small number of structure-preserving spatial domain enhancement operations during training, such as contrast and brightness adjustments, to further expand the sample space and improve the diversity and robustness of the training set.

[0057] S4. Construct a two-branch fault diagnosis model. Input the two-dimensional time-frequency image into the main branch to model the fault features, and input the enhanced image into the auxiliary branch. Use representation alignment regularization to learn latent features that are semantically consistent with the fault features of the main branch to update the parameters of the auxiliary branch.

[0058] In single-source domain generalized fault diagnosis tasks based on enhanced images, although the proposed enhancement strategies effectively improve the diversity of training data and maintain semantic consistency, the perturbations they introduce may still lead to pseudo-features inconsistent with the original sample structure, resulting in feature distortion and instability in generalization performance. To alleviate this problem, this invention proposes a robust feature learning framework based on a bi-branch structure—Brain (Bi-branch Regularization with Averaged Integration). Through the constructed bi-branch structure, representation alignment mechanism, and parameter averaging integration strategy, enhanced features are robustly fused, significantly improving the model's generalization ability.

[0059] Furthermore, such as Figure 7 As shown, step S4 includes: S41. Construct a two-branch fault diagnosis model containing a main branch and an auxiliary branch with the same structure. The encoder and classification head parameters of the main branch are kept frozen, while the encoder and classification head parameters of the auxiliary branch are trainable and initialized in the same way as the main branch.

[0060] like Figure 8 The illustrated framework comprises two structurally identical but functionally distinct model branches: a frozen main branch (the supervising model F) and an auxiliary branch (the trainable exploratory model E). The supervising model F is used to stabilize the representation of the original structural information, which is generated by the frozen encoder. With frozen classification head Composition, that is ,in It is a function composition symbol representing module concatenation. The exploration model E and the guidance model F share the same architecture, i.e. H e To explore the encoder of the model, C e To explore the model's classification head, the parameters of model E are explored at the start of training. Initialized with the same parameters as the guiding model F ,Right now .

[0061] S42. A shared projection head is introduced at the back end of the encoders of the main branch and the auxiliary branch. The shared projection head is a parameter-sharing nonlinear transformation module integrated into the back end of the dual-branch model encoder. Its function is to map the encoder output features of the main branch and the auxiliary branch to the same latent space to achieve cross-branch representation alignment.

[0062] S43. Input the two-dimensional time-frequency image into the main branch to model fault features, and input the enhanced image into the auxiliary branch. Through the shared projection head, map the encoding features of the two-dimensional time-frequency image of the main branch and the encoding features of the enhanced image of the auxiliary branch to the same latent space. By maximizing the mutual information of the latent features and suppressing redundant representations, construct the representation alignment regularization constraint.

[0063] To constrain enhanced sample features while maintaining structural consistency, Brain introduces a shared projection head. The implemented regularization-driven representation alignment mechanism. Specifically, it inputs the original sample x into the guiding model, and enhances the sample... Input the exploration model and construct the Barlow Twins (BT) representation alignment regularization term: (17) Where z is the baseline feature vector, which is generated from the original sample via H f Encoding and R projection yield , To enhance the feature vector, the enhanced sample is processed by H e Encoding and R projection yield: M is z and The cross-correlation matrix represented, M ii M represents the diagonal elements of the cross-correlation matrix. ij The off-diagonal elements of the cross-correlation matrix are used in this loss function, which aims to maximize the mutual information between the representations of the guiding and exploratory models, thereby improving semantic consistency and reducing redundant representations. The first term promotes feature consistency (the diagonal term approaches 1), the second term suppresses redundancy (the off-diagonal term approaches 0), and λ is a balance coefficient set to 0.05. Essentially, this loss maximizes the lower bound of the mutual information between the original and augmented sample encoding representations, thus mitigating the feature distortion caused by the augmentation process.

[0064] S44. By representing the alignment regularization constraint, the auxiliary branch is driven to learn latent features that are semantically consistent with the fault features of the main branch, and classification prediction results are generated based on the latent features.

[0065] S45. Calculate the cross-entropy loss of the classification prediction results of the auxiliary branch on the enhanced image, and superimpose the representation alignment regularization constraint as a joint optimization objective to update the encoder and classification head parameters of the auxiliary branch.

[0066] Specifically, the total loss function of the exploration model E is composed of the cross-entropy classification loss and the representation alignment regularization term: (18) in, It is the cross-entropy classification loss function. To characterize the alignment regularization term, x is the real sample. To enhance the samples, y represents the sample label, and ω controls the regularization strength of the character alignment for the overall training process, set to 2.0.

[0067] S5. The parameters of the auxiliary branch are obtained by periodic snapshot sampling and gradually merged into the parameter space of the main branch to form a progressively optimized diagnostic model, which is used to perform motor fault diagnosis tasks.

[0068] Furthermore, such as Figure 9 As shown, step S5 includes: S51. During the training of the auxiliary branch, the encoder and classification head parameters of the auxiliary branch are sampled at a preset period to generate a set of phased parameter states.

[0069] S52. Calculate the mean of multiple historical parameter snapshots in the stage parameter state set to generate integrated parameters.

[0070] S53. Inject the integrated parameters into the parameter space of the main branch, replace the encoder and classification head parameters of the main branch, and form the updated main branch.

[0071] S54. By periodically synchronizing the parameter snapshots of the auxiliary branches with the parameter space of the main branch, the parameter knowledge of different training stages is gradually integrated to form a progressively optimized fault diagnosis model.

[0072] To prevent the accumulation of spurious features introduced by augmentation during training, the guiding model F does not directly participate in the training process of augmented samples. Instead, it adopts a parameter-space averaging ensemble approach, periodically synchronizing with the exploration model to form a progressive ensemble knowledge accumulation strategy. (19) Where T is the integration cycle. This represents the parameter snapshots of the exploratory model E after the i-th T-th training round. Or To explore the parameter state of the model at the iT-th iteration during training, parameter space ensemble calculates the mean of these periodic historical weights to form a stable guide for model updates. This strategy is essentially equivalent to constructing a low-cost, continuously evolving model ensemble in the parameter space, effectively mitigating generalization performance fluctuations during the mid-training phase by integrating training knowledge from different stages.

[0073] Furthermore, after forming a progressively optimized fault diagnosis model, the process also includes: solidifying the diagnostic model into a format that the inference engine can load (such as ONNX, TensorRT, CoreML, etc.), stripping away training-specific operators to reduce computational overhead, and transmitting and deploying it to the edge computing unit or cloud monitoring platform of the target electromechanical equipment; responding to task instructions, performing standardized operations on the input image consistent with the training phase, such as performing multi-band modal decomposition based on the spiral cochlear structure on the input raw vibration signal to obtain the cochlear time spectrum, and using the Mel frequency mapping mechanism to reconstruct the frequency axis of the cochlear time spectrum in a perception-driven manner to generate a perceptually continuous two-dimensional time-frequency image; and performing inference operations on the two-dimensional time-frequency image through the model deployed on the edge computing unit or cloud monitoring platform of the target electromechanical equipment to obtain the diagnostic results.

[0074] In one specific embodiment, the experimental system used in the present invention is as follows: Figure 10 As shown, the experimental system includes an induction motor, an adjustable load device, a signal acquisition card, and a multi-channel vibration sensor system. The rated parameters of the three-phase induction motor are listed in Table 1; the models and performance specifications of the vibration sensors are listed in Table 2. The experimental design covers seven motor states, including: Healthy Motor (HM), Broken Bar Fault (BBF), Rotor Bend Fault (RBF), Bearing Fault (BF), Unbalance Fault (UF), Single-Phase Short Circuit Fault (SSCF), and Eccentric Fault (EF).

[0075] Corresponding structure diagram as follows Figure 11 As shown, the specific construction details are as follows: BBF ( Figure 11 a): Four rotor bars are randomly cut from a total of 28 rotor bars in the motor rotor to simulate broken-bar rotor winding damage; RBF ( Figure 11 b): Apply a symmetrical bending amount of approximately 0.4 mm at both ends of the rotor (measured using a dial indicator) to simulate nonlinear deformation of the shaft; BF ( Figure 11 c): Installing the inner and outer ring faulty rolling bearings at the fan end and drive end respectively introduces rolling element defects; UF ( Figure 11 d): Centrifugal off-center loading is induced by installing a 20g unbalanced mass block (circled in red) at a specific location on the rotor; SSCF ( Figure 11 e): Simulate a single-phase operation scenario by disconnecting one phase of the three-phase winding and selecting activation by a control switch; EF ( Figure 11f): A jack-type eccentric adjustment device is used to offset the rotor shaft center, simulating a static eccentricity fault of the rotor.

[0076] The experiment included 5 load levels (0, 2.25, 4.5, 6.75, 9 N·m) and 3 speed levels (1000 rpm, 1250 rpm, 1500 rpm), totaling 15 typical operating conditions, detailed in Table 3. The signal sampling frequency is based on f. s =5000Hz, each signal segment lasts for 1 second, and each sample contains 5000 sampling points. 200 samples are collected for each motor state and operating condition combination to construct a high-resolution vibration signal dataset covering multiple faults, multiple loads, and multiple rates, laying the foundation for subsequent diagnosis and domain generalization.

[0077] Table 1 Motor Parameters

[0078] Table 2 Sensor Parameters

[0079] Table 3. Detailed information on the motor dataset

[0080] Next, to verify the advantages of the proposed CoMel image in cross-condition generalization, a signal generalization capability experiment was designed. In each round of the experiment, one condition was selected as the source domain, and the model was trained using the data of that condition. The remaining 14 conditions were used as the target domain to evaluate the diagnostic performance of the model under unknown conditions.

[0081] Figure 12 The CoMel image visualization results for the five operating conditions (A, C, F, H, O) shown in Table 3 are presented. For comprehensive comparison, this embodiment of the invention selects five mainstream signal-to-image methods as a reference: Short-Time Fourier Transform (STFT), Gram Angular Field (GAF), Symmetric Point Plot (SDP), Wavelet Packet Transform (WVT), and Continuous Wavelet Transform (CWT). Figure 13This paper presents sample images representing conditions A (1000 r / min, 0 N·m), providing a visual basis for subsequent performance comparisons. The diagnostic model used here employs the classic VGG-16 convolutional neural network. Training parameters were set as follows: 50 training epochs, Adam optimizer, learning rate 0.0001, Dropout rate 0.5, and batch size 32. To further demonstrate the advantages of CoMel images, models based on time-series signals and Transformers were constructed for comparison. Considering the instability during one-dimensional signal training, the number of training epochs was increased to 100, and the model structure was set to 4 attention heads and 2 encoding layers, with the remaining hyperparameters consistent with the image model. Experiments were conducted using conditions A (1000 r / min, 0 N·m), H (1250 r / min, 4.5 N·m), and O (1500 r / min, 9 N·m) as source domains, with each setting repeated 5 times to reduce the impact of randomness.

[0082] In the single-source-domain generalization experiment, the method based on CoMel image representation in this invention exhibits significant advantages: when using condition A as the source domain, the model achieves an average accuracy of 75.76% across 14 target conditions, far exceeding that of STFT images (67.04%) and time-series signals (45.34%), verifying the importance of two-dimensional time-frequency features for stable generalization capabilities. Especially under complex changing conditions, CoMel images demonstrate robust performance; for example, when using conditions H and O as the source domain, the average accuracy reaches 76.29% and 68.00%, respectively, representing a 5%-10% improvement over suboptimal methods.

[0083] Compared with mainstream methods, the proposed method achieves average target domain accuracies of 93.60%, 95.38%, and 84.87% in the three source domains (A, H, and O), respectively, outperforming the best comparison methods HmmSenet (86.37%), ALT (91.72%), and HmmSenet (78.40%) by 7.23%, 3.66%, and 6.47%, respectively. Experimental results confirm that the proposed method significantly outperforms existing adversarial learning (such as SDCGAN and ALT) and temporal augmentation strategies (such as HmmSenet) in generalizing to domain shifts caused by changes in rotational speed and load. Furthermore, it maintains stability in complex domain difference scenarios. For example, the RandConv method suffers performance degradation due to random perturbations, and UCL-SDG falls below the ERM baseline by 8.58% in some experiments. The CoMel method, however, achieves robust diagnosis across operating conditions through perceptually aligned feature modeling.

[0084] Furthermore, this invention provides a single-source domain generalized motor fault diagnosis system based on cochlear perception enhancement and parameter space integration, comprising: a mode decomposition module for performing multi-band mode decomposition on the acquired original vibration signal based on the spiral cochlear structure to obtain the cochlear time spectrum; a frequency axis reconstruction module for introducing the Mel frequency mapping mechanism to reconstruct the frequency axis of the cochlear time spectrum in a perception-driven manner, generating a perceptually continuous two-dimensional time-frequency image; an image enhancement module for applying perturbation to the frequency domain amplitude spectrum of the two-dimensional time-frequency image while keeping the phase spectrum unchanged, and generating a physically consistent enhanced image after time domain reconstruction and dynamic range normalization; a model training module for constructing a two-branch fault diagnosis model, inputting the two-dimensional time-frequency image into the main branch for fault feature modeling, and inputting the enhanced image into the auxiliary branch, using representation alignment regularization to learn latent features semantically consistent with the fault features of the main branch to update the parameters of the auxiliary branch; and a model output module for obtaining the parameters of the auxiliary branch through periodic snapshot sampling and gradually fusing them into the parameter space of the main branch to form a progressively optimized fault diagnosis model for performing motor fault diagnosis tasks.

[0085] Furthermore, embodiments of the present invention provide a single-source domain generalized motor fault diagnosis device based on cochlear perception enhancement and parameter space integration, characterized in that it includes: at least one processor; and a memory communicatively connected to at least one processor; wherein the memory stores instructions executable by at least one processor, the instructions being executed by at least one processor to enable at least one processor to execute the single-source domain generalized motor fault diagnosis method based on cochlear perception enhancement and parameter space integration as described above.

[0086] Meanwhile, embodiments of the present invention provide a computer-readable storage medium storing computer-executable instructions. When the executable instructions are executed by a processor, they implement the single-source domain generalized motor fault diagnosis method based on cochlear perception enhancement and parameter space integration as described above.

[0087] In summary, to address the single-source domain generalization challenge in motor fault diagnosis, this invention provides a single-source domain generalized motor fault diagnosis method, system, device, and medium based on cochlear perception enhancement and parameter space integration. Within this framework, this invention creatively proposes a CoMel-BRAIN diagnostic scheme that integrates cochlear perception enhancement and a dual-branch regularized parameter averaging integration framework. By combining the modal decomposition advantages of cochlear transformation with the perceptual spectrum optimization of Mel mapping to construct a two-dimensional time-frequency image (CoMel image), this invention effectively enhances the expressive power of low-frequency fault features. Furthermore, a physically consistent frequency domain perturbation image enhancement strategy is proposed. By perturbing the amplitude spectrum while maintaining phase spectrum invariance, physically consistent pseudo-domain samples are generated, effectively simulating domain shift under real-world operating conditions. Combining a dual-branch architecture of a frozen main branch (guided model) and a trainable auxiliary branch (exploratory model), and through a dual-branch learning mechanism of Barlow Twins representation alignment and parameter space averaging integration, the pseudo-feature interference caused by enhanced samples is effectively mitigated, significantly improving the model's generalization stability and diagnostic accuracy. Finally, a periodic parameter space averaging ensemble mechanism is employed to gradually integrate the exploratory knowledge of the auxiliary branches, effectively suppressing spurious feature interference and significantly improving the model's generalization stability and diagnostic accuracy under unknown operating conditions. The experimental results show that the proposed method significantly outperforms existing state-of-the-art algorithms on an induction motor dataset covering 15 complex operating conditions, demonstrating excellent cross-condition generalization ability and robustness. Therefore, this invention, by addressing both biosensory modeling and robust learning mechanisms, provides a new approach and effective means for the cross-domain extension of motor fault diagnosis.

[0088] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0089] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions.

[0090] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the claims should be interpreted to include both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0091] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, then this invention should also include these modifications and variations.

Claims

1. A single-source domain generalization motor fault diagnosis method based on cochlea perception enhancement and parameter space integration, characterized by, The method comprises the following steps: performing multi-band modal decomposition on the obtained original vibration signal based on the cochlea structure to obtain a cochlea time-frequency spectrum; introducing a Mel frequency mapping mechanism to reconstruct the cochlea time-frequency spectrum in the frequency axis in a perceptually driven manner to generate a perceptually continuous two-dimensional time-frequency image; applying perturbation to the frequency domain amplitude spectrum of the two-dimensional time-frequency image and keeping the phase spectrum unchanged, performing time domain reconstruction and dynamic range normalization to generate an enhanced image with physical consistency; constructing a double-branch fault diagnosis model, inputting the two-dimensional time-frequency image into a main branch to model fault features, and inputting the enhanced image into an auxiliary branch, learning latent features that are consistent with the fault feature semantics of the main branch by using representation alignment regularization, and updating the parameters of the auxiliary branch; obtaining the parameters of the auxiliary branch through periodic snapshot sampling and gradually fusing them into the parameter space of the main branch to form a gradually optimized diagnosis model for performing the fault diagnosis task of the motor.

2. The single-source domain generalization motor fault diagnosis method based on cochlea perception enhancement and parameter space integration of claim 1, wherein, The method comprises the following steps: based on the spatial distribution characteristics of the cochlea structure, constructing a cochlear wavelet filter bank along a preset angle domain, wherein each wavelet filter corresponds to a preset angle position of the cochlea and has spatial locality and frequency response characteristics for multi-band modal decomposition of the signal; for the signal components in the obtained original vibration signal that are outside the frequency range of cochlear perception, generating an analytical signal through Hilbert transform and multiplying it with a modulated carrier to generate a modulated propagating wave after frequency modulation migration; projecting the modulated propagating wave onto the cochlear wavelet mode supported by the cochlear wavelet filter bank through inner product calculation to obtain the response coefficients of each mode, wherein different modes represent the local frequency response of different angle positions of the cochlea; performing Hilbert analysis on the response coefficients to obtain the instantaneous amplitude and phase information of each mode, and generating a cochlea time-frequency spectrum with color intensity proportional to amplitude based on amplitude and intensity mapping.

3. The single-source domain generalization motor fault diagnosis method based on cochlea perception enhancement and parameter space integration of claim 1, wherein, The method comprises the following steps: constructing a frequency mapping relationship based on Mel perception characteristics, expanding the resolution of linear frequencies below a preset boundary frequency and compressing the redundancy of linear frequencies above the boundary frequency to generate a nonlinear perceptual frequency scale aligned with the human hearing range; based on the nonlinear perceptual frequency scale, uniformly dividing the corresponding Mel frequency axis to obtain a set of overlapping triangular bandpass filters, each triangular bandpass filter covering a corresponding perceptual frequency interval, and the boundaries of adjacent triangular bandpass filters overlapping to achieve perceptual continuity coverage of the frequency axis; inputting the power spectrum components of the cochlea time-frequency spectrum into the bandpass filter bank to calculate the weighted spectrum energy through each filter to extract the perceptually aligned energy values of the corresponding Mel bands; stacking the perceptually aligned energy values of each Mel band in the time dimension to generate a two-dimensional time-frequency image that integrates the cochlea modal decomposition characteristics and the Mel perceptual consistency.

4. The single-source domain generalization motor fault diagnosis method based on cochlea perception enhancement and parameter space integration of claim 3, wherein, Each triangular bandpass filter has the following form: ； wherein, is the response value of the mth triangular bandpass filter at frequency index n, n is the frequency index of the power spectrum, and f(m) is the center frequency of the mth filter, which is uniformly divided by the Mel frequency.

5. The single-source domain generalization motor fault diagnosis method based on cochlea perception enhancement and parameter space integration of claim 1, wherein, Applying a perturbation to a frequency domain amplitude spectrum of a two-dimensional time-frequency image and keeping a phase spectrum unchanged, performing time domain reconstruction and dynamic range normalization, a physically consistent enhanced image is generated, including: Performing fast Fourier transform on the two-dimensional time-frequency image to separate the amplitude spectrum and the phase spectrum represented in the frequency domain; Keeping the phase spectrum unchanged, applying a controllable intensity perturbation to the amplitude spectrum to generate a perturbed amplitude spectrum, combining the perturbed amplitude spectrum with the phase spectrum to form an enhanced frequency domain representation; Performing inverse fast Fourier transform on the enhanced frequency domain representation to reconstruct a time domain enhanced image; Performing dynamic range constraint on the reconstructed time domain enhanced image to align the reconstructed time domain enhanced image with the dynamic range of the amplitude spectrum of the two-dimensional time-frequency image, and supplementing with a spatial domain enhancement operation that maintains the structure to generate a physically consistent enhanced image.

6. The single-source domain generalization motor fault diagnosis method based on cochlea perception enhancement and parameter space integration of claim 1, wherein, A double-branch fault diagnosis model is constructed, the two-dimensional time-frequency image is input into the main branch to model the fault features, and the enhanced image is input into the auxiliary branch, and the latent features consistent with the fault feature semantics of the main branch are learned by using the representation alignment regularization to update the parameters of the auxiliary branch, including: A double-branch fault diagnosis model containing a main branch and an auxiliary branch with the same structure is constructed, wherein the parameters of the encoder and the classification head of the main branch are kept frozen, and the parameters of the encoder and the classification head of the auxiliary branch are trainable and initialized to be consistent with the main branch; A shared projection head is introduced at the back end of the encoders of the main branch and the auxiliary branch; The two-dimensional time-frequency image is input into the main branch to model the fault features, and the enhanced image is input into the auxiliary branch, and the encoding features of the two-dimensional time-frequency image by the main branch and the encoding features of the enhanced image by the auxiliary branch are mapped to the same latent space through the shared projection head, and the representation alignment regularization constraint is constructed by maximizing the mutual information of the latent features and suppressing redundant representations; Through the representation alignment regularization constraint, the auxiliary branch is driven to learn latent features consistent with the fault feature semantics of the main branch, and a classification prediction result is generated based on the latent features; The cross-entropy loss of the classification prediction result of the auxiliary branch on the enhanced image is calculated, and the representation alignment regularization constraint is stacked as a joint optimization target to update the encoder and classification head parameters of the auxiliary branch.

7. The single-source domain generalization motor fault diagnosis method based on cochlea perceptual enhancement and parameter space integration of any one of claims 1-6, wherein, The parameters of the auxiliary branch are obtained by periodic snapshot sampling and gradually fused into the parameter space of the main branch to form a gradually optimized diagnosis model for performing motor fault diagnosis tasks, including: During the training process of the auxiliary branch, the encoder and classification head parameters of the auxiliary branch are periodically sampled to generate a set of stage parameter states; The mean value of multiple historical parameter snapshots in the set of stage parameter states is calculated to generate an integrated parameter; The integrated parameter is injected into the parameter space of the main branch to replace the encoder and classification head parameters of the main branch, forming an updated main branch; The parameter snapshots of the auxiliary branch are periodically synchronized with the parameter space of the main branch, and the parameter knowledge of different training stages is gradually fused to form a gradually optimized diagnosis model; The diagnosis model is solidified into a format loadable by an inference engine, transmitted and deployed to an edge computing unit of a target mechatronic device or a cloud monitoring platform; In response to the task instruction, the input original vibration signal is subjected to spiral cochlea structure-based multi-band modal decomposition to obtain a cochlea time-frequency spectrum, and the cochlea time-frequency spectrum is subjected to frequency axis reconstruction in a perception-driven manner by means of a Mel frequency mapping mechanism to generate a perceptually continuous two-dimensional time-frequency image. The two-dimensional time-frequency image is subjected to inference operation by a model deployed on an edge computing unit or a cloud monitoring platform of the target electromechanical equipment to obtain a diagnosis result.

8. A single-source domain generalization motor fault diagnosis system based on cochlea perceptual enhancement and parameter space integration, characterized in that, The method comprises: a modal decomposition module configured to perform spiral cochlea structure-based multi-band modal decomposition on the acquired original vibration signal to obtain a cochlea time-frequency spectrum; a frequency axis reconstruction module configured to introduce a Mel frequency mapping mechanism to perform frequency axis reconstruction on the cochlea time-frequency spectrum in a perception-driven manner to generate a perceptually continuous two-dimensional time-frequency image; an image enhancement module configured to apply perturbation to a frequency domain amplitude spectrum of the two-dimensional time-frequency image while keeping a phase spectrum unchanged, and perform time domain reconstruction and dynamic range normalization to generate a physically consistent enhanced image; a model training module configured to construct a double-branch fault diagnosis model, input the two-dimensional time-frequency image into a main branch to model fault features, and input the enhanced image into an auxiliary branch, learn latent features consistent with fault feature semantics of the main branch by representation alignment regularization, and update parameters of the auxiliary branch; a model output module configured to acquire the parameters of the auxiliary branch by periodic snapshot sampling, and gradually fuse the parameters into a parameter space of the main branch to form a gradually optimized diagnosis model for performing a motor fault diagnosis task.

9. A single-source domain generalization machine fault diagnosis device based on cochlea perception enhancement and parameter space integration, characterized in that, The method comprises: at least one processor; and a memory in communication with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the single-source domain generalization motor fault diagnosis method based on cochlear perception enhancement and parameter space integration according to any one of claims 1-7.

10. A computer-readable storage medium having stored thereon computer- executable instructions, the computer-executable instructions comprising instructions for: receiving a request for a resource; determining whether the request is for a resource that is subject to a policy; and if the request is for a resource that is subject to a policy, then determining whether the request is from a client that is subject to the policy. Executable instructions are executed by the processor to implement the single-source domain generalization motor fault diagnosis method based on cochlear perception enhancement and parameter space integration according to any one of claims 1-7.

Images