Bearing fault identification method based on dynamic generative adversarial network and expert feedback

Abstract

The invention provides a bearing fault identification method based on a dynamic generative adversarial network and expert feedback, and relates to the field of bearing fault diagnosis, and the method comprises the steps: generating a high-fidelity fault vibration signal through employing a condition generator and a triple discriminator generative adversarial network; verifying and generating sample quality through a 1D residual verification network and adding the sample quality into a training set; segmenting the vibration signals passing the test by using layered adaptive sampling, and keeping high-frequency impact characteristics in the vibration signals; a dynamic sparse attention mechanism is adopted to reduce unnecessary attention calculation and improve calculation efficiency, and different types of faults are accurately recognized in combination with a hybrid expert system classifier; and detecting the confidence of the diagnosis result, and triggering a feedback mechanism to regenerate a sample to complete autonomous iterative optimization when the confidence is low. According to the method, a generative adversarial network, a fault diagnosis model and a feedback mechanism are fused, accurate diagnosis of bearing faults is achieved through multi-level data enhancement and screening feedback, the diagnosis precision is continuously improved in continuous iteration, and the method is suitable for solving the problem that a traditional method is poor in performance under data scarcity and noise interference. The innovative closed-loop evolutionary logic of generation-diagnosis-feedback is provided, and the robustness and accuracy of fault recognition are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bearing fault diagnosis, and particularly relates to a bearing fault identification method based on a dynamically generated generative adversarial network and expert feedback. BACKGROUND

[0002] As the core component of rotating machinery, the running state of rolling bearings directly affects the safety and reliability of the equipment. Under complex working conditions such as high speed, heavy load and variable load, bearings often face long-term effects of adverse factors such as vibration impact and poor lubrication, which can easily lead to the occurrence of faults such as wear and fatigue damage. Statistics show that bearing faults account for 40% of the total amount of rotating machinery faults, and are often the main cause of unplanned equipment shutdown. Therefore, timely and accurate diagnosis of bearing faults can not only effectively prevent equipment downtime, but also significantly improve the operational reliability and production efficiency of the equipment. In modern industry, the research and application of bearing fault diagnosis technology has become a key link to ensure stable operation of equipment, prolong service life and reduce maintenance costs. By identifying the abnormal state of bearings early, maintenance measures can be taken in advance to avoid potential major failures and ensure continuous and stable operation of the equipment.

[0003] With the rapid development of deep learning and data-driven methods, various advanced algorithms have been gradually applied to the field of bearing fault diagnosis. However, existing fault diagnosis models still face some key bottlenecks in practical applications. Traditional algorithms such as GAN, SVM and PCA have single modalities when generating samples, making it difficult to cover complex working conditions. LSTM, KNN and random forest have computational redundancy when processing long sequence vibration signals, affecting efficiency. The separation of data generation and diagnosis modules leads to error accumulation, affecting overall diagnostic accuracy. The existence of these problems limits the application effect and performance improvement of the technology. Therefore, in view of these problems, it is necessary to innovate the existing bearing fault diagnosis technology to improve its robustness, accuracy and cross-condition adaptability. SUMMARY

[0004] The present application provides a bearing fault identification method based on a dynamically generated generative adversarial network and expert feedback to solve the problems in the background art.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical solution: a bearing fault identification method based on a dynamically generated generative adversarial network and expert feedback. This method combines a generative adversarial network and a hybrid expert system classifier, and realizes accurate diagnosis of bearing faults through a multi-level data enhancement and screening feedback mechanism. The present application particularly addresses the problems of poor performance of traditional methods under data scarcity and noise interference, and error accumulation caused by module separation, and proposes an innovative "generation-diagnosis-feedback" closed-loop evolution logic, which significantly improves the robustness and accuracy of fault identification. The specific steps include: Step one: adaptive generation of countermeasures, generate high-quality bearing fault samples through conditional generative adversarial network, and use multiple discriminators and Wasserstein optimization to automatically optimize the generator, continuously improve the quality of generated samples, and add qualified samples to the training set until the sample quality is verified by the 1D residual verification network; Step two: fault diagnosis, retain key features through hierarchical sampling, remove unnecessary attention calculation through dynamic attention mechanism, improve computational efficiency, combine hybrid expert system classifier to analyze vibration signals in depth, and accurately identify bearing fault types; Step three: feedback loop, according to the confidence detection of the diagnosis result, trigger the feedback mechanism to regenerate the sample when the confidence is low, realize the self-iteration optimization of the model, and continuously improve the diagnosis performance of the model.

[0006] In a preferred embodiment, the step one comprises: The system receives random noise and fault labels and inputs them to the generator to generate virtual samples; The generated samples are evaluated by time domain, frequency domain and time-frequency domain discriminators, and the loss of each discriminator is calculated; Each discriminator calculates the gradient penalty based on the generated sample and evaluates the loss compared to the sample; The losses of multiple discriminators are fused through a weighted fusion strategy; The Wasserstein optimization algorithm is applied to update the generator to regenerate the sample, and the 1D residual verification network is used to verify the sample. Low-quality samples are returned to the generator for regeneration, and high-quality samples are added to the training set.

[0007] In a preferred embodiment, the step two comprises: The original vibration signal is sampled hierarchically, and high-frequency and low-frequency signals are extracted respectively. The high-frequency signal is sampled using a small window, and the low-frequency signal is sampled using a large window. The short-time and long-time characteristics of the vibration signal are captured, which facilitates subsequent processing; Position encoding is added to signals of different frequency bands, so that signals of different frequency bands are effectively aligned with time series, helping the model understand the time correlation of the signal and ensuring effective transmission of time series information; The dynamic attention mechanism is used to calculate the similarity of the signal, and the Top-K most relevant features are selected to suppress noise and increase the attention to fault features; The signal is allocated to different expert networks through the gating mechanism, and is processed specifically for different fault types. Bearing faults include inner ring faults, outer ring faults, rolling element faults, and compound faults; The outputs of each expert network are fused by weighting to generate a final bearing fault diagnosis result, and the outputs of different expert networks are self-adjusted according to the contribution of the final diagnosis result, so that the final diagnosis result can reflect the comprehensive opinions of all expert networks and ensure the diagnosis accuracy.

[0008] In a preferred embodiment, the step three comprises: According to the confidence score of the diagnosis result, the reliability of the result is judged, for high confidence samples, the final diagnosis report is directly output, and the specific fault type and fault position are displayed; For low confidence samples, directional generation instructions are constructed by extracting physical indicators such as impact interval distribution and characteristic frequency intensity, and a feedback loop is triggered to generate physical constraint sample injection increments for training, so as to optimize the gating weight and expert subnetwork, and realize continuous iterative optimization of diagnosis accuracy.

[0009] Compared with the prior art, the present application introduces a generative adversarial network and a hybrid expert system classifier in bearing fault recognition, which has been significantly improved. First, the conditional generative adversarial network module is used to solve the problems of insufficient samples and inability to generate fixed type samples in traditional methods. Through the adversarial training of the conditional generator and the discriminator, diversified false samples are generated, which enhances the robustness and recognition accuracy of the model; hierarchical sampling and dynamic attention mechanism are used to accurately extract key features in the signal, compared with the traditional fixed feature extraction method, the present application can more flexibly and effectively capture time-frequency domain information, and improve the diagnosis accuracy; the expert network and the weighted fusion mechanism are used for special diagnosis according to different fault types, which improves the depth and accuracy of fault recognition; through the confidence detection and feedback mechanism, the present application can trigger sample regeneration under low confidence condition, and continuously optimize the diagnosis result and model performance. Through the innovative multiple mechanisms, the present application overcomes the limitations in the prior art, and significantly improves the accuracy and self-adaptive ability of bearing fault recognition. BRIEF DESCRIPTION OF DRAWINGS

[0010] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art according to these drawings.

[0011] Figure 1 The flowchart of the bearing fault recognition method based on dynamic generative adversarial network and expert feedback of the present application;

[0012] Figure 2 The structural schematic diagram of the adaptive generative adversarial module of the present application;

[0013] Figure 3The figure is a structural schematic diagram of the fault diagnosis module of the application. DETAILED DESCRIPTION

[0014] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be described in detail and completely below with reference to the drawings in the present application. Obviously, the described embodiments are some embodiments of the present application, but not all the embodiments. They should not be understood as limiting the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application. In the description of the present application, it should be understood that the terms used are only for the purpose of description, and should not be understood as indicating or implying relative importance.

[0015] The present application provides a bearing fault identification method based on dynamic generative adversarial network and expert feedback. Figure 1 Figure 2 Figure 3 The present application provides a bearing fault identification method based on dynamic generative adversarial network and expert feedback.

[0016] Referring to Figure 1 The present application provides a bearing fault identification method based on dynamic generative adversarial network and expert feedback. S1, an adaptive generative adversarial module generates a synthetic vibration signal; S2, the fault diagnosis module performs hierarchical slicing, analyzes the sliced data to realize fault classification; S3, the confidence of the classification result is evaluated, the result with high confidence is confirmed and output, and the result with substandard confidence is transmitted to the feedback loop for cyclic processing.

[0017] Referring to Figure 2 The step S1 in the bearing fault identification method based on dynamic generative adversarial network and expert feedback comprises: S11, synthesizing a vibration signal of a specific fault according to a label; S12, using a triple discriminator and a conditional generator to constitute an adversarial training, providing an adaptive gradient penalty unit for each discriminator, jointly verifying the authenticity of the sample, and optimizing the generator through a loss function; S13, using a 1D residual verification network containing a spectrum normalization layer to screen the synthetic vibration signal.

[0018] ​​As described in step S11 above, the random noise z (128-dimensional Gaussian distribution vector) is first input into the embedding layer together with the fault label c (one-hot encoding), where the label c is mapped to a 128-dimensional semantic vector through an embedding matrix. The two are spliced in the feature dimension to form a joint vector, at which time the noise provides generation diversity and the label injects directional constraints of fault type characteristics. After the vector is input into the fully connected layer and activated by batch normalization and ReLU activation function, it is projected into a 3584-dimensional feature tensor, and then reshaped into a two-dimensional feature map. Progressive upsampling is performed through five layers of transpose convolution: the first layer uses a 5*1 convolution kernel and a step of 2 to expand the 7-point sequence to 14 points and reduce the channel to 256; each layer is followed by batch normalization and ReLU activation, which gradually expands the time sequence length while preserving the fault characteristics. After five levels of upsampling, the feature map evolves into a 16-channel * 224-length. Finally, a 1*1 convolution is used to fuse the multi-channel features into a single-channel waveform, and a Tanh activation function is used to compress the amplitude to [-1, 1] to simulate the real vibration range. To match the standard 2048-point length, the generated signal is expanded by zero-padding at the beginning and end. In this process, the generator extracts and reconstructs multiple fault characteristics through the local connection characteristics of the transpose convolution: the shallow network captures the characteristic frequency of the bearing fault; the middle layer constructs the equidistant pulse of the impact fault, focusing on the time sequence characteristics; the deep layer refines the transient impact waveform form, focusing on the decay oscillation of the fault. The fault label information controls the feature generation direction through conditional batch normalization parameters, such as the inner ring fault label that strengthens the frequency conversion harmonic component, and the composite fault label that excites the multi-band coupling effect. The final output synthetic signal contains complete physical characteristics: in the time domain, it presents fault-specific impact sequences; in the frequency domain, it exhibits characteristic frequencies and their harmonics; and in the time-frequency domain, it displays the decay characteristics of impact energy. These features are verified by multiple domain discriminators, forming physically interpretable bearing fault vibration signals.

[0019] As described in step S12 above, in the adversarial training process of the generation module, the conditional generator and the triple discriminators including the time domain discriminator, the frequency domain discriminator, and the time-frequency domain discriminator achieve signal optimization through dynamic game. In the early stage of training, the generator receives random noise z and fault label c, and generates a preliminary vibration signal through a seven-layer transpose convolution network. At this time, the triple discriminators operate in parallel: the time domain discriminator uses a 5-layer 1D CNN structure to analyze the pulse interval and envelope form of the waveform, the frequency domain discriminator performs FFT on the signal after applying a Hanning window, detects the amplitude distribution of the characteristic frequency through a 3-layer convolution network, and the time-frequency domain discriminator uses Morlet wavelet transform to generate a scale-time matrix, and verifies the decay characteristics of impact energy through a 2D CNN. The three-channel loss is weighted and fused to drive the generator optimization, and the loss weighted fusion satisfies: , , , , where SNR is the signal-to-noise ratio of the input signal. Each discriminator is equipped with an adaptive gradient penalty unit, whose penalty coefficient is dynamically adjusted according to the Wasserstein distance of the current batch of samples. During training, the three-way discriminant loss is fused into the total loss through attention weighting, driving the generator to perform four-stage optimization. The time-domain optimization stage reduces the pulse interval error, using the RMSProp adaptive learning rate optimization algorithm with a momentum term. The frequency-domain refinement stage introduces spectral normalization constraints to ensure that the feature frequency position deviation is less than 5% of the theoretical value. The time-frequency joint optimization stage enhances the shock attenuation characteristics through wavelet scale correlation loss. The fine-tuning stage uses dynamic learning rate decay, with an initial learning rate of 5e-5, which is reduced by 20% every 20 rounds. Each iteration includes 5 discriminator updates and 1 generator update. The 1D ResNet of the validation module is equipped with 8 residual blocks for real-time evaluation of the generation quality. When the sample signal-to-noise ratio is greater than 30 dB and the kurtosis is greater than 3.5, the sample is included in the training set. Otherwise, the noise injection mechanism is triggered to regenerate the sample, forming a closed-loop optimization system.

[0020] As described in step S13 above, in the generation module, the 1D residual verification network with spectral normalization layer realizes strict screening of the synthesized vibration signal through multi-scale feature analysis. The network uses an 8-residual-block cascade architecture, with each residual block containing two layers of 1D convolution, a spectral normalization layer, and a skip connection. The input synthesized signal first passes through the first convolution layer for coarse-grained feature extraction, and then gradually extracts features through the residual blocks: first, focusing on macro-pulse periodicity, analyzing the matching degree of pulse interval variance and theoretical fault frequency; second, analyzing the meso-waveform structure, calculating the local kurtosis index to verify the shock characteristics; finally, combining global average pooling to evaluate the micro-noise distribution. The spectral normalization layer ensures feature stability by constraining the Lipschitz constant of each weight matrix, while using an improved power iteration method to accelerate computation. During the verification process, the network output quality score is obtained by weighting the three indicators: the time-domain pulse correlation weight is 40%, the frequency-domain feature frequency energy ratio weight is 30%, and the time-frequency domain wavelet energy aggregation degree weight is 30%. When the network output quality score is greater than 0.7, the signal is judged to be qualified. Non-compliant samples automatically trigger optimization, first applying a small amount of Gaussian disturbance to the noise vector to expand the solution space exploration capability, rigidly rotating the fault label semantic vector along the characteristic hyperplane to fine-tune the fault mode expression, reconstructing the deep convolution kernel size and compensating the channel number to enhance the local waveform modeling accuracy, and finally implementing iterative optimization through conditional normalization parameter correction, convolution kernel gradient constraint, and feature fusion layer momentum optimization. This strategy significantly improves the pulse timing consistency, accurately controls the feature frequency shift, enhances the shock waveform steepness, and stabilizes the time-frequency energy decay behavior, until the comprehensive quality indicators exceed the preset threshold, forming an adaptive optimization pathway under closed-loop physical constraints.

[0021] As described in S11, S12, and S13 above: The generation module realizes the synthesis and optimization of vibration signals through a conditional generative adversarial network. First, Gaussian noise and fault labels are fused into joint features through the embedding layer, projected and reshaped through the fully connected layer, and then decoupled through five layers of transposed convolution layers. The shallow layer establishes the pulse interval corresponding to the characteristic frequency, the middle layer regulates the impact timing, and the deep layer optimizes the attenuation characteristics. The generated signal is jointly verified by a triple discriminator. The time domain discriminator detects the pulse interval error, the frequency domain discriminator analyzes the characteristic frequency signal-to-noise ratio, and the time-frequency discriminator evaluates the wavelet energy distribution. A 1D residual verification network equipped with spectral normalization scores the signal quality. Unqualified samples trigger the generator parameters to be fine-tuned and regenerated, thus achieving a closed loop from noise to physically interpretable signals.

[0022] See Figure 3 , step S2 in the bearing fault identification method based on dynamic generative adversarial network and expert feedback includes: S21, layered adaptive sampling technology, decomposes the original vibration signal into high-frequency transient impact components and low-frequency harmonic components, respectively extracting microsecond-level pulse details and frequency-shifting harmonic structure; S22, dynamic sparse attention mechanism, performs Top-K selective enhancement on time-domain impact positions and frequency-domain feature frequencies, suppressing noise interference and improving the signal-to-noise ratio of fault features; S23, hybrid expert system classifier, builds multiple expert networks, automatically assigns weights to the four types of expert networks (inner race, outer race, rolling element, and composite) based on the physical properties of the fault characteristics, and performs weighted fusion to output the final fault type and confidence level.

[0023] As described in step S21 above, in the layered adaptive sampling process, the raw vibration signal first undergoes frequency band separation, breaking the input data into two parallel paths: a high-frequency path intercepts the terminal segments of the signal and extracts microsecond-level transient impact components through bandpass filtering, accurately capturing the transient waveform distortion characteristics caused by the fault. The low-frequency path uses sliding window mean downsampling to compress the signal scale while preserving macro-periodic characteristics such as rotation harmonics. The high-frequency data is then fed into a learnable position encoding layer to dynamically annotate the precise temporal location of the impact event. The low-frequency data is then fed into a fixed-pattern sinusoidal position encoding layer to establish long-range phase correlations. Both signals are then fed into a dynamic sparse attention layer, which selects key correlated points at each moment based on feature similarity calculations, performs local weight enhancement, and performs global noise suppression. Finally, the dual-path features are fused at the concatenation layer to form a decoupled representation that combines local impact details with global harmonic structure, providing physically meaningful feature input for subsequent fault diagnosis.

[0024] As described in step S22 above, the dynamic sparse attention mechanism realizes fault feature enhancement by constructing a physics-inspired feature selection architecture, and its processing flow includes three layers of operations. At the input layer, the feature vector is first linearly projected into three independent subspaces: the query vector captures the physical state of the current time series position, such as the impact phase or frequency amplitude, the key vector encodes historical feature information, and the value vector stores the feature entity content. The similarity calculation layer establishes a feature association matrix based on the scaled dot product theory. The element values ​​of this matrix reflect the coupling strength between the time domain impact event and the frequency domain harmonic component. The core Top-K screening layer introduces prior knowledge of fault physics: for time domain features, a local sliding window constraint is set to retain only significant correlation points within the period before and after the current impact event to ensure that the transient impact waveform obtains local full connection enhancement; for frequency domain features, the period restriction is relaxed to allow the fundamental frequency and harmonic components to establish long-range correlation across periods. During the calculation, only k positions are retained to calculate the attention weight, and the similarity matrix between the query matrix Q and the key matrix K is calculated, and the rows are screened. The first k largest values ​​in each row of the matrix, the k value satisfies: , where L is the sequence length, and the dynamic sparse attention calculation satisfies: , where K is the key matrix and k is the number of filters; the screening process strictly adheres to the Shannon sampling theorem to ensure the integrity of key characteristic frequency components. The weight allocation layer normalizes the filtered sparse matrix and implements differentiated weighting based on the physical characteristics of the fault features: transient impulse waveforms are assigned peak enhancement weights to improve the time-domain kurtosis index, and characteristic frequency components are assigned energy enhancement coefficients to increase the frequency-domain spectral kurtosis. Noise components are naturally attenuated during the weighting process because the associated paths are cut off. In the final output feature tensor, the improved signal-to-noise ratio of the fault features is due to a triple mechanism: the time-domain impulse waveform is phase-synchronized through local window weighting, the frequency-domain characteristic frequencies achieve energy accumulation through cross-period correlation, and background noise is effectively suppressed due to the severing of sparse connections. The theoretical basis of this mechanism is that bearing fault features have structured sparsity in the time-frequency domain, the time-domain impulses are periodically sparsely distributed, and the frequency-domain characteristic frequencies are line spectrum components. Dynamic sparse attention essentially constructs a sparse connection graph that matches the physical characteristics of the fault through learnable association screening, so that limited computing resources can be focused on the fault-sensitive feature area, forming an optimized feature expression at the output end of the attention layer with concentrated impact event energy, prominent characteristic frequency components, and suppressed noise floor, providing physically interpretable input features for subsequent diagnostic networks.

[0025] As described in step S23 above, the hybrid expert system classifier realizes accurate fault diagnosis through physical feature driven adaptive weight assignment, the core of which is to build a hierarchical decision architecture. The input dual-path feature vector (containing high-frequency impact details and low-frequency harmonic structure) is first compressed into a global feature descriptor through a global average pooling layer, which preserves key physical properties such as time-domain impact density distribution and frequency-domain feature energy ratio. The feature descriptor is input into a two-layer fully connected gating network: the first layer reconstructs the feature space through nonlinear transformation to extract fault mode sensitive features; the second layer maps to a four-dimensional weight space, and after normalization processing, generates the activation weights of the inner ring, outer ring, rolling body, and composite four types of expert networks, which strictly follow the fault physical characteristics. When the input feature presents the unique rotating frequency harmonic group of the inner ring fault, the inner ring expert network obtains the dominant weight; if the equally spaced impact sequence of the outer ring fault is detected, the outer ring expert network weight significantly increases. The four expert networks process the original feature vector in parallel: the inner ring expert network uses a narrowband harmonic analysis layer to capture the rotating frequency and its multiple components; the outer ring expert network analyzes the impact interval variance through a time series correlation layer; the rolling body expert network builds a modulation sideband detection network to identify the characteristic frequency sideband; the hybrid expert system classifier deploys a cross-attention mechanism to analyze the multi-fault coupling effect. After each expert network outputs a fault probability vector, the gating weight performs weighted fusion, and the fusion process is essentially an optimal linear combination of the physical feature space, for example, when the impact interval variance and rotating frequency harmonic energy are both significant, the outer ring and inner ring expert network outputs are synergistically affected in proportion to the weight. The final output comprehensive fault probability vector synchronously generates a confidence index, which integrates the maximum probability value and the probability distribution entropy dual information, when the probability distribution presents a unimodal high-entropy characteristic, the confidence increases, and if the multimodal distribution is triggered, a feedback loop is triggered. This mechanism accurately matches the physical features and the strengths of the expert networks to improve the diagnosis accuracy in the bearing composite fault scenario.

[0026] As described in S21, S22 and S23 above, in the fault diagnosis process, the hierarchical adaptive sampling first decouples the original vibration signal into high-frequency transient impact components and low-frequency harmonic components, preserves the microsecond-level pulse details through band-pass filtering, and extracts the rotating frequency harmonic structure by downsampling. The dual-path signals are embedded with position coding to establish time-domain precise positioning and frequency-domain long-range correlation. The decoupled features are input into a dynamic sparse attention layer, which builds a selective reinforcement mechanism based on physical properties. The time-domain path uses local window constraints to reinforce impact event peak correlation, and the frequency-domain path establishes a cross-cycle model to accumulate feature frequency energy. Through Top-K screening, background noise is suppressed, and the optimized features are output with steepened impact waveforms and concentrated feature frequency energy. The optimized dual-path features are spliced into a gating network, which analyzes impact density and harmonic energy ratio, generates weights for the inner ring (narrowband harmonic analysis), outer ring (time series correlation detection), rolling body (sideband analysis), and composite (coupling effect modeling) four expert networks, and outputs fault probability through weighted fusion.

[0027] Referring to Figure 1 and Figure 3 , the step S3 in the bearing fault identification method based on the dynamic generative adversarial network and expert feedback, in the classification result confidence evaluation stage, the system implements comprehensive judgment by analyzing the fault probability distribution characteristics output by the gating network, first extracts the probability peak value of the dominant fault category as the certainty reference, and quantifies the discrete degree of the probability distribution based on the information entropy theory to evaluate the diagnostic ambiguity, and the two are fused to generate a comprehensive confidence score. When the score exceeds the preset threshold, the system confirms the high confidence result and outputs the diagnosis report; for the samples that do not meet the standard, a feedback loop guided by physical characteristics is started, the time domain impact interval distribution characteristics and the frequency domain feature frequency energy ratio of the original signal are extracted, combined with the most uncertain fault category in the probability distribution, a directional generation instruction containing the target impact timing mode and the feature frequency intensity requirement is constructed. The conditional generator receives the instruction and embeds the uncertainty mode characteristics in the noise injection link, accurately controls the theoretical period alignment of the impact event through the convolution kernel phase modulation layer, and constrains the time domain error and frequency domain offset of the output signal by using the multi-domain physical discriminator. After the new synthesized sample is confirmed to be qualified by the quality score layer of the verification network, the incremental training process is injected to fine-tune the gating network weight distribution logic of the uncertainty mode, and the sensitive feature extraction capability of the corresponding fault expert network is specifically enhanced, and finally the model reprocesses the original sample through lightweight global optimization until the confidence requirement is met.

[0028] The bearing fault identification method based on the dynamic generative adversarial network and expert feedback proposed in the application constructs a closed-loop intelligent system, the generation module first synthesizes a vibration signal with time domain pulse sequence, frequency domain feature frequency and time frequency attenuation characteristics according to the fault label, and inputs the vibration signal into the diagnosis module after being verified by the multi-domain discriminator; the diagnosis end decouples the high-frequency impact and low-frequency harmonic features through hierarchical adaptive sampling, strengthens the fault sensitive area by means of dynamic sparse attention, and drives each expert network to make a weighted decision by the gating network; the confidence evaluation layer analyzes the certainty and ambiguity of the diagnosis result, extracts physical indicators such as impact interval distribution and feature frequency intensity for low-confidence samples, and constructs a directional generation instruction to trigger the feedback loop; The generator synthesizes the physically constrained samples to inject incremental training, and specifically optimizes the gating weight and the expert network. The modules cooperate to form a "generation-diagnosis-feedback" self-evolution channel, which continuously improves the diagnosis accuracy in continuous iteration.

[0029] The above merely provides a specific implementation of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A bearing fault identification method based on dynamic generative adversarial network and expert feedback, characterized in that: include: Adaptive Generative Adversarial Module, consisting of a conditional generator, a triple discriminator, and a 1D residual verification network, receives a noise vector and a fault label and outputs a specific synthetic vibration signal; The fault diagnosis module includes a hierarchical adaptive sampling module, a dynamic sparse attention mechanism, and a hybrid expert system classifier. It slices the input signal, performs feature analysis, and outputs the fault category. The feedback loop feeds the low-confidence samples and their fault labels output by the fault diagnosis module back to the adaptive generative adversarial module to resynthesize enhanced data.

2. The bearing fault identification method based on dynamic generative adversarial network and expert feedback according to claim 1 is characterized in that: The adaptive generation adversarial module includes: The condition generator is a convolutional neural network that embeds fault labels and synthesizes vibration signals of specific faults based on the labels; The triple discriminator performs adversarial training with the conditional generator, driving generator optimization through a loss function. It includes a time domain discriminator, a frequency domain discriminator, and a time-frequency domain discriminator. Each discriminator is equipped with an adaptive gradient penalty unit to jointly verify the authenticity of the sample. A 1D residual verification network with a spectral normalization layer is used to deeply verify the authenticity of the generated signal features.

3. The bearing fault identification method based on dynamic generative adversarial network and expert feedback according to claim 1 is characterized in that: The fault diagnosis module comprises: Layered adaptive sampling module to process vibration signals in non-uniform blocks; Dynamic sparse attention mechanism, which processes block signals for feature extraction and dynamically selects key segments for attention calculation; The hybrid expert system classifier constructs a multi-expert gated network, automatically assigns weights to the four expert networks of inner race, outer race, rolling element, and composite according to the physical properties of the fault characteristics, and performs weighted fusion to output the final fault type and confidence level.

4. The bearing fault identification method based on dynamic generative adversarial network and expert feedback according to claim 2 is characterized by: The time domain discriminator is used to analyze the waveform pulse interval and envelope shape, focusing on the amplitude and period of the vibration signal; The frequency domain discriminator applies a Hanning window to the signal and then performs FFT transformation to detect the amplitude distribution of the characteristic frequency; The time-frequency domain discriminator uses Morlet wavelet transform to generate a scale-time matrix to verify the attenuation characteristics of the impact energy; The three-channel loss weighted fusion drives the generator optimization, and the loss weighted fusion satisfies: ,coefficient , , satisfy: , where SNR is the signal-to-noise ratio of the input signal.

5. The bearing fault identification method based on dynamic generative adversarial network and expert feedback according to claim 3 is characterized by: Layered adaptive sampling: dynamically divides slices based on signal envelope entropy, using a small window with high-density sampling for high-frequency bands and a large window with low-density sampling for low-frequency bands; Dynamic sparse attention mechanism performs Top-K selective enhancement on time-domain impact positions and frequency-domain feature frequencies, suppressing noise interference and improving the signal-to-noise ratio of fault features; Hybrid expert system classifier, expert networks E1, E2, E3, and E4 correspond to fault types: inner race fault, outer race fault, rolling element fault, and composite fault, respectively. The gating network dynamically assigns expert network weights based on input features, sets differentiated activation thresholds for different fault types, and performs weighted fusion to output the final fault type and confidence level.

6. The bearing fault identification method based on dynamic generative adversarial network and expert feedback according to claim 1 is characterized in that: A closed-loop optimization method, comprising: The generation module synthesizes a physical constraint vibration signal with time domain pulse sequence, frequency domain characteristic frequency and time-frequency attenuation characteristics based on the fault label, and inputs it into the diagnosis module after collaborative verification by the multi-domain discriminator; The diagnostic end uses layered adaptive sampling to decouple high-frequency impact and low-frequency harmonic features, uses dynamic sparse attention to enhance fault-sensitive areas, and uses a gating network to drive multi-expert weighted decision-making. The confidence assessment layer analyzes the certainty and ambiguity of the diagnostic results, extracts physical indicators such as the impact interval distribution and characteristic frequency intensity of low-confidence samples to construct directional generation instructions, triggers the feedback loop to generate physical constraint samples for injection into incremental training, and optimizes the gating weights and expert subnetworks in a targeted manner to achieve continuous iterative optimization of diagnostic accuracy.

Images