Fault diagnosis method based on manifold density regulation and time-frequency dual-domain self-modulation generation

Abstract

The present application belongs to the technical field of fault diagnosis, and discloses a fault diagnosis method based on manifold density regulation and time-frequency dual-domain self-modulation generation. First, a small number of fault samples are mapped to a low-dimensional manifold space; the low-dimensional manifold space is divided into grid cells, and the local manifold density of each grid cell is calculated; and the local manifold density is used for hierarchical batch sampling and manifold density regulation diffusion noise addition. In the adversarial training stage, the conditional self-modulation generation module generates generated fault samples corresponding to the target fault category according to the noise-added samples, the diffusion time step and the fault label, and the time-frequency dual-domain discrimination module performs sample scoring and multi-layer feature response constraint on the generated fault samples through the time-domain discrimination branch and the frequency-domain discrimination branch. Finally, the generated fault samples and the real standardized samples are fused to form a class-balanced training set, and the fault diagnosis model is used to output the fault diagnosis result. The present application improves the fault diagnosis reliability under the conditions of data imbalance and sample scarcity.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology, and specifically relates to a fault diagnosis method based on manifold density regulation and time-frequency dual-domain self-modulation, which is particularly suitable for fault diagnosis scenarios under conditions of data imbalance and sample scarcity. Background Technology

[0002] Key rotating components of high-speed trains, such as bogie axle box bearings and gearboxes, operate under high-speed, heavy-load, and complex interference environments for extended periods, making them highly susceptible to localized damage. Currently, data-driven fault diagnosis methods based on vibration monitoring data are widely used. However, in actual operating environments, trains operate in a healthy state for most of the time, with extremely low probability of specific types of faults, resulting in a scarcity of actual fault data. This severe data imbalance makes it difficult for conventional data-driven models to fully learn the mapping boundaries of a few types of faults, easily leading to missed fault reports.

[0003] To alleviate the data imbalance problem caused by data scarcity, generative adversarial networks (GANs) or diffusion models are typically used to synthesize virtual samples to expand the training dataset. However, applying such general generative models to the generation and diagnosis of physical vibration signals from high-speed trains still has certain limitations. Specifically, in the forward diffusion stage of data generation, existing diffusion models often employ globally uniform noise intensity or fixed step size scheduling strategies. In the data feature space, extremely scarce fault samples often exhibit sparse and fragile topological manifolds. If deep Gaussian noise of the same intensity as that applied to the majority class normal samples is applied to these fragile samples, their originally weak real fault impact characteristics will be over-smoothed or even destroyed, causing feature collapse during the reverse denoising generation process, thus making it difficult to generate effective high-quality minority class samples.

[0004] Furthermore, existing models still have significant inherent flaws in their adversarial generation mechanisms. Conventional conditional generation models typically fuse conditional information by directly concatenating discrete labels with feature channels. However, when faced with the complex and intense background noise of high-speed trains, this can easily lead to conditional forgetting in the deep network iterations, causing the model to fail to accurately control the underlying statistical distribution of the signal and resulting in blurred boundaries for fault type samples. Conventional generative adversarial networks often rely solely on the global scalar output of the discriminator for macroscopic truth-false games. This single global constraint lacks multi-scale feature calibration and ignores the microscopic physical manifestations of vibration signals in deep latent spaces. Fault signals from rotating components of high-speed trains are essentially periodic transient impacts containing specific characteristic frequencies. If the generator is only constrained by the macroscopic time domain, it is highly susceptible to spectral smoothing artifacts, which disrupt the microscopic waveform texture and frequency band energy distribution, leading to severe distortion of the generated data.

[0005] Meanwhile, in the downstream fault diagnosis stage, the general architecture model also suffers from the defect of destroying physical semantics. Most existing Transformer-based fault diagnosis models directly apply fixed-length hard block strategies from natural language or image processing to process one-dimensional vibration signals. However, vibration signals are composed of periodic impact pulses, and blind fixed-length segmentation can easily truncate a complete physical impact waveform, destroying the underlying physical topology of the signal. This forces the model's self-attention mechanism to search for correlations among fragmented waveform pieces, limiting its ability to accurately capture the evolutionary patterns of cross-cycle faults.

[0006] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art. Summary of the Invention

[0007] The purpose of this invention is to propose a fault diagnosis method based on manifold density regulation and time-frequency dual-domain self-modulation generation. This method achieves the directional generation of minority class fault samples by performing low-dimensional manifold mapping, gridded local manifold density calculation, hierarchical batch sampling, and manifold density regulation diffusion noise addition on minority class fault samples, combined with a conditional self-modulation generation module and a time-frequency dual-domain hierarchical feature matching mechanism. At the same time, it utilizes the fault cycle adaptive block Transformer fault diagnosis model to learn from the class-balanced training set, thereby improving the reliability of rotating machinery fault diagnosis under conditions of data imbalance and sample scarcity.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: The fault diagnosis method based on manifold density modulation and time-frequency dual-domain self-modulation includes the following steps: Step 1. Obtain the original time-series vibration signals of the key components of the rotating machinery to be diagnosed, and perform sliding window truncation and labeling to obtain the original sample set; the original sample set is then standardized by standardization preprocessing to form a standardized sample set; Step 2. Map the minority class fault samples in the standardized samples to a low-dimensional manifold space to obtain the corresponding low-dimensional feature points; divide the low-dimensional manifold space into grid cells and calculate the local manifold density of each grid cell; Step 3. Construct a generative adversarial network that includes a conditional self-modulation generation module and a time-frequency dual-domain discrimination module; The input to the conditional self-modulation generation module is the noisy sample processed by the manifold density-controlled diffusion strategy, the fault label, and the diffusion time step. The output of the conditional self-modulation generation module is the generated fault sample. The manifold density-controlled diffusion strategy adjusts the noise intensity during the forward diffusion process based on the local manifold density; The time-frequency dual-domain discrimination module receives generated fault samples and real standardized fault samples, and extracts time-domain multi-layer feature response maps, frequency-domain multi-layer feature response maps and sample scores through time-domain discrimination branches and frequency-domain discrimination branches, respectively. Among them, sample scores are used to construct discriminator adversarial loss and generator adversarial loss, while time-domain multi-layer feature response maps and frequency-domain multi-layer feature response maps are used to construct time-frequency dual-domain hierarchical feature matching loss; During the training of generative adversarial networks, a hierarchical batch sampling strategy is adopted to adjust the probability of minority class fault samples in the training batch according to the local manifold density. Step 4. Use the trained conditional self-modulation generation module to generate fault samples for the minority fault categories, and then fuse the generated fault samples with the real standardized samples in the standardized sample set to form a class-balanced training set. The fault diagnosis model is trained using the class-balanced training set to obtain the trained fault diagnosis model. The sample to be diagnosed is then input into the trained fault diagnosis model to output the fault diagnosis result.

[0009] Furthermore, based on the aforementioned fault diagnosis method based on manifold density control and time-frequency dual-domain self-modulation, this invention also proposes a computer device, which includes a memory and one or more processors; the memory stores executable code, and when the one or more processors execute the executable code, it is used to implement the steps of the aforementioned fault diagnosis method based on manifold density control and time-frequency dual-domain self-modulation.

[0010] Furthermore, based on the above-mentioned fault diagnosis method based on manifold density control and time-frequency dual-domain self-modulation, this invention also proposes a computer-readable storage medium storing a program thereon. When the program is executed by a processor, it is used to implement the steps of the above-mentioned fault diagnosis method based on manifold density control and time-frequency dual-domain self-modulation.

[0011] The present invention has the following advantages: This invention employs low-dimensional manifold mapping and gridded local manifold density calculation on minority class fault samples. This local manifold density is simultaneously used for hierarchical batch sampling and forward diffusion noise intensity control. This increases the probability of rare fault samples located in low-density grid cells appearing in training batches and corresponds to a smaller noise intensity during diffusion noise addition, thereby reducing the risk of excessive perturbation of weak fault impact envelopes and local waveform textures, and improving the generative model's ability to learn minority class fault features. In the adversarial generation stage, this invention constructs a conditional self-modulation generation module, fusing diffusion time step information and fault category information into a global conditional control vector. This global conditional control vector is used to dynamically predict the scaling and bias parameters of the self-modulation normalization layer, allowing conditional information to continuously participate in the multi-layer feature map distribution adjustment, improving the consistency between generated fault samples and the target fault category. Simultaneously, this invention introduces a time-frequency dual-domain discrimination module and a time-frequency dual-domain hierarchical feature matching loss. The time-domain discrimination branch constrains the waveform texture of the generated samples, while the frequency-domain discrimination branch constrains the frequency band energy distribution of the generated samples, thereby improving the physical fidelity of the generated fault samples. Furthermore, this invention employs a fault cycle adaptive segmented Transformer fault diagnosis model, which determines the data block window size based on the geometric parameters, motion parameters, and sampling frequency of the rotating component. This ensures that the input data block can cover the complete fault impact segment, reducing the damage to the fault impact structure caused by fixed-length segments and improving the reliability of fault diagnosis under conditions of data imbalance and sample scarcity. Attached Figure Description

[0012] Figure 1 This is a flowchart of the fault diagnosis method based on manifold density regulation and time-frequency dual-domain self-modulation of the present invention; Figure 2 This is a schematic diagram of the structure of the conditional self-modulation generation module in an embodiment of the present invention; Figure 3 This is a schematic diagram of the time-frequency dual-domain discrimination module in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of the fault diagnosis module of the fault cycle adaptive block Transformer in this embodiment of the invention; Figure 5 This is a schematic diagram of the confusion matrix of the diagnostic model trained using only an imbalanced training set on an independent test set in an embodiment of the present invention. Figure 6 This is a schematic diagram of the confusion matrix of the fault diagnosis model trained on the same independent test set after data augmentation using the method of the present invention in an embodiment of the present invention. Detailed Implementation

[0013] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: Example 1 To address the problems of data distortion, low conditional modulation accuracy, lack of multi-scale microscopic physical constraints, and low efficiency of fault diagnosis models under extreme imbalance conditions, this invention first acquires the original time-series vibration signals of key components of rotating machinery. These signals are then processed through sliding window truncation, category labeling, and standardization preprocessing to obtain a standardized sample set. Next, low-dimensional manifold mapping and grid cell division are performed on the minority fault samples in the standardized sample set, and the local manifold density of each grid cell is calculated. Based on the local manifold density, a hierarchical batch sampling strategy is constructed to increase the probability of rare fault samples appearing in the training batch. Furthermore, a manifold density-controlled diffusion strategy adaptively adjusts the forward diffusion noise intensity to reduce the risk of excessive perturbation of rare fault features. During the adversarial training phase, a conditional self-modulation generation module is constructed. This module utilizes diffusion time step information and fault category information to form a global conditional control vector and dynamically predicts the scaling and bias parameters of the self-modulation normalization layer to adjust the distribution of generated features. Simultaneously, a time-frequency dual-domain discrimination module and a time-frequency dual-domain hierarchical feature matching loss are introduced to constrain the time-domain waveform texture and frequency-domain energy distribution of the generated fault samples. Finally, the generated fault samples are fused with real standardized samples to form a class-balanced training set, and the fault diagnosis results are output in conjunction with the fault diagnosis model. This invention can improve the reliability of fault diagnosis for complex rotating machinery under conditions of data imbalance and sample scarcity.

[0014] like Figure 1 As shown, the fault diagnosis method based on manifold density modulation and time-frequency dual-domain self-modulation includes the following steps: Step 1: Obtain the original time-series vibration signal of the key component of the rotating machinery to be diagnosed. Perform sliding window truncation and category labeling on the original time-series vibration signal to obtain the original sample set composed of sample fragments and their corresponding category labels.

[0015] The sample fragments in the original sample set are preprocessed to standardize them, resulting in a standardized sample set.

[0016] The key components of the rotating machinery include at least one of bearings, gearboxes, and bogie axle boxes. This embodiment uses bearing vibration signal diagnosis as an example for illustration, and the bearing is a typical key component in rotating machinery.

[0017] The original time-series vibration signal sequence is defined as: ;in, This represents the number of data points in the original time-series vibration signal. These represent the first and second digits of the original time-series vibration signal, respectively. One sampling point.

[0018] The original time-series vibration signal sequence is processed using a sliding window method. The process involves truncating the data to obtain multiple sample segments of equal length. Let the window length be... The sliding step size is Then the first Sample fragments Represented as: .in, .

[0019] This represents the number of sample segments extracted from the original time-series vibration signal, and .

[0020] Based on the device operating status label, for each sample segment Assign corresponding category labels .

[0021] In this embodiment, the category label This includes any one of the following: normal condition, inner race fault, outer race fault, and rolling element fault. The original sample set consists of multiple sample segments of uniform length and their corresponding category labels. .

[0022] For the original sample set Each sample fragment Perform standardization processing to obtain standardized samples. The formula is as follows: .

[0023] in Represents sample fragments Standardized samples after standardization processing.

[0024] Represents sample fragments The mean, Represents sample fragments standard deviation To represent the standardized smoothing coefficient used to prevent the denominator from being zero, The range of values ​​is to Preferred .

[0025] From a standardized sample and their corresponding category labels Composition of standardized sample set: .

[0026] In this embodiment, the window length Take 1024 data points, sliding step size The settings are based on the sampling frequency and signal overlap rate. To simulate the situation where fault samples are scarce in the engineering field, the dataset settings are shown in Table 1.

[0027] Table 1 Imbalanced Datasets

[0028] The training set contains 700 normal class samples, and 20 samples each for the three minority fault categories: inner race fault, outer race fault, and rolling element fault. The test set uses a class-balanced configuration.

[0029] The number of samples for each category in the test set—normal, inner ring fault, outer ring fault, and rolling element fault—was set to 300.

[0030] Step 2. Map the minority fault samples in the standardized samples to a low-dimensional manifold space to obtain the corresponding low-dimensional feature points; divide the low-dimensional manifold space into grid cells and calculate the local manifold density of each grid cell.

[0031] To characterize the scarcity of minority-class fault samples in the feature space, this invention targets a standardized sample set. Low-dimensional manifold mapping and local manifold density calculation are performed on a minority of fault samples.

[0032] Minority failure samples are those with a lower number of samples than normal failure samples. In this embodiment, minority failure samples include inner race failure samples, outer race failure samples, and rolling element failure samples.

[0033] From standardized sample sets By selecting minority class failure samples, a minority class failure sample set is obtained: .in, Indicates the first Standardized samples corresponding to a minority of fault samples express Corresponding fault category labels, This represents the total number of minority fault samples.

[0034] Step 2.1 Use the uniform manifold approximation and projection algorithm to perform dimensionality reduction mapping on the minority class of fault samples.

[0035] With a minority of fault sample sets Standardized samples As a high-dimensional input point, a uniform manifold approximation and projection algorithm is used for dimensionality reduction, reducing the length to [value missing]. Standardized samples Mapped to two-dimensional low-dimensional feature points: .

[0036] in This represents the mapping function between the approximation and projection of a uniform manifold. Indicates the first Two-dimensional low-dimensional feature points corresponding to a minority of fault samples and They represent Coordinate components in the first and second dimensions.

[0037] In this embodiment, the processing steps of the uniform manifold approximation and projection algorithm include the following: Calculate any two standardized samples and Distance between: .

[0038] in, The Euclidean norm is used. In other implementations, the preset distance metric function can also be cosine distance, Manhattan distance, or other distance metric functions that can characterize the degree of difference between time-series samples.

[0039] For each standardized sample According to sample distance Sort in ascending order, starting from the smallest. Select from other standardized samples We obtain the nearest neighbor set by taking samples from each nearest neighbor: .

[0040] in, Standardized samples of Nearest neighbor set Indicates the number of neighbors; This represents the standardized samples sorted in ascending order of sample distance. The Nearest neighbor samples, .

[0041] Construct high-dimensional nearest neighbor weights based on the nearest neighbor set: ; in, Indicated by standardized samples When centered, standardized samples Compared to standardized samples The directed high-dimensional nearest neighbor weights; Show and standardized samples The corresponding local distance scaling parameters.

[0042] Indicated by standardized samples When centered, standardized samples Compared to standardized samples The directed high-dimensional nearest neighbor weights; By swapping the indices in the directed high-dimensional nearest neighbor weight calculation formula and get.

[0043] Symmetric nearest neighbor weights are obtained by symmetricizing the directed high-dimensional nearest neighbor weights: .

[0044] in, Standardized samples and Symmetric nearest neighbor weights between them.

[0045] Calculate two-dimensional feature points in a two-dimensional low-dimensional space. and Low-dimensional similarity between them: .

[0046] in, Representing two-dimensional low-dimensional feature points and Low-dimensional similarity, and These are parameters used to control the shape of the low-dimensional similarity curve; to maintain the high-dimensional nearest neighbor weights. Compared with low-dimensional similarity With consistency as the goal, construct an optimization objective function: .

[0047] By minimizing the optimization objective function Iteratively update each two-dimensional low-dimensional feature point This ensures that the neighborhood relationships between the two-dimensional low-dimensional feature points maintain the local neighborhood relationships of the minority class of fault samples in the high-dimensional space.

[0048] In this embodiment, the number of neighbors in the uniform manifold approximation and projection algorithms The minimum distance is set to 0.1, and the target dimension is set to 2. After dimensionality reduction, all two-dimensional low-dimensional feature points form a low-dimensional manifold space. .

[0049] in, It is used to characterize the local neighborhood distribution relationship of minority class fault samples in the reduced-dimensional feature space.

[0050] Step 2.2 Mesh the low-dimensional manifold space and calculate the local manifold density.

[0051] For low-dimensional manifolds The mesh is generated to obtain a set of mesh cells: .

[0052] in Represents a set of grid cells. Represents any grid cell, Indicates the number of grid cells; for any grid cell Statistical analysis of cells falling into the grid Number of minority fault samples ,in: .

[0053] in Indicates the indicator function; when two-dimensional low-dimensional feature points Falling into the grid cell hour, ,otherwise Calculate the mesh element according to the following formula. Local manifold density: .

[0054] in Represents grid cells Local manifold density, This represents the density smoothing coefficient in the calculation of local manifold density.

[0055] In this embodiment, The range of values ​​is to Preferred .

[0056] grid cells Local manifold density Assign the value to the cell that falls into the grid. For minority class fault samples, the local manifold density corresponding to each minority class fault sample is obtained.

[0057] When the number of minority fault samples in a certain grid cell When the density is low, the local manifold density of the grid cell is low. A lower value indicates that the fault feature region corresponding to the grid cell is relatively scarce in the training dataset.

[0058] As a common control variable for subsequent hierarchical batch sampling strategies and manifold density-controlled diffusion strategies: On the one hand, it is used to increase the probability of minority fault samples in low-density grid cells entering the training batch; on the other hand, it is used to adjust the noise intensity in the forward diffusion process to reduce the risk of excessive perturbation of scarce fault features in the noise addition process.

[0059] Step 3. Construct and train a generative adversarial network that includes a conditional self-modulation generation module and a time-frequency dual-domain discrimination module.

[0060] The input to the conditional self-modulation generation module is the noisy sample processed by the manifold density-controlled diffusion strategy, the fault label, and the diffusion time step. The output of the conditional self-modulation generation module is the generated fault sample.

[0061] The manifold density-controlled diffusion strategy adjusts the noise intensity during the forward diffusion process based on the local manifold density.

[0062] The time-frequency dual-domain discrimination module receives generated fault samples and real standardized fault samples, and extracts time-domain multi-layer feature response maps, frequency-domain multi-layer feature response maps and sample scores through time-domain discrimination branches and frequency-domain discrimination branches, respectively.

[0063] Among them, sample scores are used to construct discriminator adversarial loss and generator adversarial loss, while time-domain multi-layer feature response maps and frequency-domain multi-layer feature response maps are used to construct time-frequency dual-domain hierarchical feature matching loss.

[0064] Furthermore, during the training of the generative adversarial network, this invention employs a hierarchical batch sampling strategy, which adjusts the probability of extracting minority class fault samples in the training batch based on the local manifold density.

[0065] In this embodiment, the generative adversarial network includes a conditional self-modulation generation module. Time-frequency dual-domain discrimination module .in, This indicates a conditional self-modulation generation module. This represents the network parameters of the conditional self-modulation generation module; This indicates the time-frequency dual-domain discrimination module. This represents the network parameters of the time-frequency dual-domain discrimination module.

[0066] During the training of the generative adversarial network, based on the minority class fault sample set obtained in step 2... Grid cell set and local manifold density The training batch is constructed using a hierarchical batch sampling strategy. Then, manifold density modulation diffusion noise is applied to the minority fault samples in the training batch. The input conditional self-modulation generation module generates fault samples. Finally, the adversarial loss and time-frequency dual-domain hierarchical feature matching loss are calculated through the time-frequency dual-domain discrimination module.

[0067] Step 3.1 Construct training batches using a hierarchical batch sampling strategy.

[0068] Based on the mesh cell set obtained in step 2 and the number of minority fault samples in each grid cell Determine the set of non-empty grid cells with a sample size greater than zero: .

[0069] in, This represents a non-empty set of grid cells containing at least one minority class fault sample. express Any non-empty grid cell in, Indicates falling into a grid cell The number of minority fault samples.

[0070] For any non-empty grid cell Based on its local manifold density Calculate the sampling probability of this grid cell: .

[0071] in, Indicates from non-empty grid cells The sampling probability of extracting minority class fault samples from the sample. Represents a set of non-empty grid cells Any grid cell used for summation traversal. Indicates non-empty grid cells The local manifold density.

[0072] Represents the sampling probability smoothing coefficient. The range of values ​​is to Preferred .

[0073] When constructing each training batch, first follow the sampling probability. from Select a target grid cell, and then randomly extract training samples from the minority class fault samples within the target grid cell until the minority class fault samples required for the current training batch are obtained.

[0074] Local manifold density The smaller the sampling probability, the lower the probability. The larger the value, the higher the probability of minority class fault samples located in low-density grid cells appearing in the training batch of the generative adversarial network.

[0075] Step 3.2. Implement the manifold density-controlled diffusion strategy.

[0076] Standardized minority class fault samples are selected from the current training batch as initial diffusion samples, denoted as . ,in, This represents the standardized minority class of fault samples that participate in forward diffusion noise addition. Indicates the first A standardized minority of fault samples.

[0077] Based on the low-dimensional mapping relationship in step 2: Determine the initial sample for diffusion. Corresponding two-dimensional low-dimensional feature points And determine the two-dimensional low-dimensional feature points. The associated grid cell is denoted as .

[0078] in Indicates the initial sample of diffusion Corresponding two-dimensional low-dimensional feature points The grid cell into which it falls.

[0079] Let the diffusion time step The preset noise variance is .

[0080] in, , Indicates the maximum diffusion time step; based on the initial diffusion sample Belonging grid cell Local manifold density For the preset noise variance After correction, the noise variance after density adjustment is obtained: .

[0081] in, Indicates the initial sample of diffusion At the diffusion time step The noise variance after local manifold density modulation Represents the protection coefficient of scarce features and .

[0082] In this embodiment, the preferred value is 0.5. Represents the density scaling factor and In this embodiment, the value is set according to the mean or median of the local manifold density distribution.

[0083] The retention coefficient after density adjustment is defined as: .

[0084] in Indicates the initial sample of diffusion At the diffusion time step The signal retention coefficient after density modulation.

[0085] And define from the first diffusion time step to the second... The cumulative retention coefficient for each diffusion time step is: .

[0086] in Indicates the initial sample of diffusion In the Signal retention coefficient after density modulation at each diffusion time step This represents the cumulative index of the diffusion time steps.

[0087] Initial sample for diffusion From the first diffusion time step to the second The cumulative signal retention coefficient for each diffusion time step.

[0088] According to the cumulative retention coefficient The diffusion time step is obtained. The following noisy samples: .

[0089] in, Indicates the initial sample of diffusion At the diffusion time step The generated noisy samples, Indicates Gaussian noise, and . Indicates the initial sample of diffusion An identity matrix matching the dimensions. The manifold density-controlled diffusion strategy is based on the initial diffusion sample. Belonging grid cell Local manifold density Adjusting the preset noise variance .

[0090] When the initial sample is diffused When located in low-density grid cells, Smaller, corrected noise variance The corresponding reduction results in a smaller noise intensity for minority fault samples in low-density grid cells during forward diffusion, thereby preserving their temporal impact envelope and local waveform texture and reducing the risk of excessive perturbation of the temporal impact envelope and local waveform texture.

[0091] Step 3.3 Generate fault samples using the conditional self-modulation generation module.

[0092] like Figure 2 As shown, the conditional self-modulation generation module It includes a time step embedding branch, a fault label embedding branch, a conditional fusion layer, multiple one-dimensional convolutional generation blocks, multiple self-modulation normalization layers, and an output convolutional layer.

[0093] The conditional self-modulation generation module receives the density-modulated noisy samples. diffusion time step and fault labels and output generated fault samples. .

[0094] in Represents a real standardized fault sample Having the same fault category label The generation of fault samples.

[0095] Time step embedding branch is used to diffuse time steps Mapped to time step embedding vectors: ;in, Indicates the time step embedding function, Indicates diffusion time step The corresponding time step embedding vector; Fault tag embedding branch is used to embed fault tags Mapped to fault label embedding vectors: .in, This indicates the fault label embedding function. Indicates fault label The corresponding fault label embedding vector.

[0096] Will and Vector concatenation is performed, and a conditional fusion layer is introduced to obtain the global conditional control vector: .in, This represents the global conditional control vector. This represents a multilayer perceptron. This indicates vector concatenation.

[0097] Global Conditional Control Vector Used to characterize diffusion time step information and target fault category information in the current generation process.

[0098] Among them, the diffusion time step information is used to characterize the noise disturbance stage of the generation process, and the target fault category information is used to limit the fault type to which the generated fault sample belongs.

[0099] The local manifold density controls the generation process through a self-modulation generation module based on the input conditions of the density-controlled noisy sample.

[0100] Multiple one-dimensional convolutional generation blocks are arranged in series along the one-dimensional time-series waveform generation direction. Each one-dimensional convolutional generation block includes a one-dimensional convolutional layer and a nonlinear activation layer, and a self-modulation normalization layer is connected after each one-dimensional convolutional generation block.

[0101] Let the first The feature map output by each one-dimensional convolutional generator block is Then the first Each self-modulation normalization layer controls the vector according to global conditions. Prediction scaling parameters and bias parameters: .

[0102] in Indicates the first The parameter prediction network corresponding to each self-modulation normalization layer Indicates the first Scaling parameters of the output of the self-modulation normalization layer. Indicates the first The bias parameters of the output of the self-modulated normalization layer.

[0103] Using scaling parameters and bias parameters , for the Feature maps output by a one-dimensional convolutional generator block Perform self-modulation normalization processing to obtain the modulated feature map. The formula is as follows: ; in Representation of feature map The mean, Representation of feature map standard deviation This represents element-wise multiplication. This represents the modulation smoothing coefficient used to prevent the denominator from being zero during the self-modulation normalization process.

[0104] After processing through multiple layers of one-dimensional convolutional generating blocks and self-modulation normalization layers, the output convolutional layer will output the deep features from the last self-modulation normalization layer. Mapping to a one-dimensional time-series waveform yields the generated fault samples: .

[0105] in, This indicates the convolution kernel parameters of the output convolutional layer. This represents a one-dimensional convolution operation. This represents the bias vector of the output convolutional layer. This indicates the number of layers in the one-dimensional convolutional generator block and the self-modulation normalization layer.

[0106] Step 3.4. Use the time-frequency dual-domain discrimination module to perform discrimination and feature matching constraints.

[0107] like Figure 3 As shown, the time-frequency dual-domain discrimination module It includes a time-domain discrimination branch and a frequency-domain discrimination branch. The time-frequency dual-domain discrimination module receives real standardized fault samples. and generating fault samples Output sample scores and multi-layer feature response maps respectively.

[0108] in To obtain samples from a minority of fault classes The real standardized fault samples selected from them The output of the conditional self-modulation generation module is a sample of the real standardized fault. Having the same fault category label The generation of fault samples.

[0109] The temporal discriminant branch includes a one-dimensional convolutional unit, a temporal multi-layer feature response map extraction unit, and a temporal sample scoring unit; One-dimensional convolutional unit receives and The temporal waveform is obtained, and temporal convolution features are extracted.

[0110] The temporal multi-layer feature response map extraction unit outputs temporal multi-layer feature response maps from different network layers of temporal convolutional features; the temporal sample scoring unit outputs temporal sample scores based on the temporal multi-layer feature response maps.

[0111] Time-domain discriminant branch for real standardized fault samples The processing result is expressed as follows: , The time-domain discriminant branch generates fault samples. The processing result is expressed as follows: , .

[0112] in Represents real standardized fault samples In the time-domain discrimination branch The temporal characteristic response map output by the layer. Indicates the generation of fault samples In the time-domain discrimination branch Temporal characteristic response map of the layer output.

[0113] This indicates the number of network layers used for feature matching in the temporal discriminant branch.

[0114] and These represent the time-domain sample scores of real standardized fault samples and generated fault samples, respectively.

[0115] The frequency domain discrimination branch includes a Fast Fourier Transform (FFT) unit, an amplitude spectrum calculation unit, a frequency domain multilayer feature response map extraction unit, and a frequency domain sample scoring unit; the FFT unit respectively performs... and Perform frequency domain transformation.

[0116] The amplitude spectrum calculation unit calculates the frequency domain amplitude spectrum based on the frequency domain transformation results.

[0117] The frequency domain multi-layer feature response map extraction unit outputs frequency domain multi-layer feature response maps from different network layers of the frequency domain amplitude spectrum.

[0118] The frequency domain sample scoring unit outputs the frequency domain sample score based on the frequency domain multi-layer feature response map.

[0119] Real standardized fault samples and generating fault samples The frequency domain amplitude spectra are expressed as follows: , .

[0120] in, Represents the Fast Fourier Transform. Represents real standardized fault samples The frequency domain amplitude spectrum, Indicates the generation of fault samples The frequency domain amplitude spectrum.

[0121] Frequency domain discriminant branch for real standardized fault samples The processing result is expressed as follows: , ; Frequency domain discriminant branch generates fault samples The processing result is expressed as follows: , .

[0122] in Represents real standardized fault samples In the frequency domain discrimination branch The frequency domain characteristic response map of the layer output. Indicates the generation of fault samples In the frequency domain discrimination branch Frequency domain characteristic response diagram of the layer output.

[0123] This indicates the number of network layers used for feature matching in the frequency domain discriminant branch.

[0124] and These represent the frequency domain sample scores of real standardized fault samples and generated fault samples, respectively.

[0125] The time-frequency dual-domain discrimination module obtains a comprehensive sample score based on the time-domain sample score and the frequency-domain sample score: ; in The comprehensive sample score representing the true standardized fault samples. This represents the overall sample score for generating faulty samples. and Let represent the weight coefficients of the time-domain sample scores and the frequency-domain sample scores, respectively, and satisfy the following: , as well as .

[0126] In this embodiment, the time-domain sample scoring and the frequency-domain sample scoring use the same weight, and the weight coefficient of the time-domain sample scoring is... Weighting coefficients for frequency domain sample scores All are set to 0.5, that is: ; The comprehensive sample score is used to form the true and false discrimination results between real standardized fault samples and generated fault samples, and is used to construct the discriminator adversarial loss and the generator adversarial loss; The time-domain multi-layer feature response map and the frequency-domain multi-layer feature response map are used to calculate the time-frequency dual-domain hierarchical feature matching loss; During training, the discriminator adversarial loss is used to update the network parameters of the time-frequency dual-domain discriminant module; The generator adversarial loss and the time-frequency dual-domain hierarchical feature matching loss are used together to update the network parameters of the conditional self-modulation generation module to constrain the generation of fault samples. It closely approximates real standardized fault samples in terms of time-domain waveform texture and frequency-domain energy distribution.

[0127] Step 3.5 Calculate the loss function and update the network parameters.

[0128] Suppose a training batch contains Each real standardized fault sample and its corresponding generated fault sample represent:

[0129] ;in, and Having the same fault category label .

[0130] In this embodiment, the comprehensive sample score is used to construct the discriminator adversarial loss and the generator adversarial loss.

[0131] For example, the discriminator adversarial loss can be expressed as: ; The generator adversarial loss can be expressed as: .

[0132] in This represents the Sigmoid function. This indicates that the discriminator is susceptible to loss. This indicates that the generator is resisting loss.

[0133] Time-frequency dual-domain hierarchical feature matching loss is used to measure true standardized fault samples. With generating fault samples The difference between the multilayer feature response maps of the time-domain discriminant branch and the frequency-domain discriminant branch is calculated using the following formula: .

[0134] in This represents the time-frequency dual-domain hierarchical feature matching loss; The time-domain discriminant branch is represented by the first... Hierarchical weights of the layer feature response map; The frequency domain discriminant branch is represented by the first... Hierarchical weights of the layer feature response map; express Norm.

[0135] Hierarchical weights satisfy: , , , .

[0136] The generator loss function of the conditional self-modulation generation module is defined as: .in, This represents the generator loss of the conditional self-modulation generation module. The weight coefficients of the time-frequency dual-domain hierarchical feature matching loss are represented by... .

[0137] During network training, the time-frequency dual-domain discrimination module Network parameters Counteracting loss based on the discriminator Update; Conditional self-modulation generation module Network parameters Based on generator loss Update.

[0138] Time-frequency dual-domain hierarchical feature matching loss Used to update the conditional self-modulation generation module The network parameters are used to constrain the generation of fault samples. The time-domain waveform texture and frequency-domain energy distribution closely approximate real-world standardized fault samples. .

[0139] Discriminator parameters are updated using discriminator-adversarial loss. .

[0140] Step 4. Use the trained conditional self-modulation generation module to generate fault samples for the minority fault categories, and then fuse the generated fault samples with the real standardized samples in the standardized sample set to form a class-balanced training set. A fault diagnosis model is trained using a class-balanced training set to obtain a trained fault diagnosis model. The sample to be diagnosed is then input into the trained fault diagnosis model to output the fault diagnosis result.

[0141] The fault diagnosis model of this invention adopts a fault diagnosis model based on fault cycle adaptive block Transformer.

[0142] In the conditional self-modulation generation module After training is completed, for the minority of fault categories with insufficient sample size in the training set, the trained conditional self-modulation generation module is used to generate corresponding fault samples.

[0143] Set up a standardized training set The training set used for model training is: .in, This represents the true standardized samples in the training set. Indicates the corresponding category label, This indicates the number of training samples.

[0144] Let the set of categories be: .in, This indicates the number of fault categories. In this embodiment, the categories include normal, inner ring fault, outer ring fault, and rolling element fault. .

[0145] Set training categories The corresponding number of true standardized samples is And let the target number after class balance be: .

[0146] For any minority of fault categories The number of samples that need to be generated is: .

[0147] in, Indicates category The number of fault samples that need to be generated needs to be increased. This is achieved using the trained conditional automodulation generation module. For a few fault categories generate By generating fault samples, a generated sample set is obtained: .

[0148] in For category The corresponding number One generated fault sample, Assign a fault category label to the generated fault sample.

[0149] The generated sample sets corresponding to all minority fault categories are fused with the real standardized samples in the training set to form a class-balanced training set. .

[0150] ;in, This represents the set of minority fault categories.

[0151] In this embodiment, the number of normal category samples in the training set is 700, and the number of true standardized samples for the three minority fault categories—inner ring fault, outer ring fault, and rolling element fault—is 20 each.

[0152] Therefore, 680 generated fault samples were generated for each of the inner ring fault, outer ring fault, and rolling element fault, and then fused with the real standardized samples in the training set, so that the number of training samples for each category of normal, inner ring fault, outer ring fault, and rolling element fault was 700.

[0153] like Figure 4 As shown, the fault diagnosis model adopts the fault cycle adaptive block Transformer fault diagnosis model.

[0154] The fault cycle adaptive segmentation Transformer fault diagnosis model includes a fault cycle estimation unit, an adaptive segmentation unit, a block embedding layer, a position encoding layer, a Transformer encoder, and a fault classification layer.

[0155] The fault cycle estimation unit calculates the theoretical fault characteristic frequency corresponding to different candidate fault types based on the geometric and motion parameters of the rotating component under test, and determines the reference fault characteristic frequency for adaptive segmentation based on the theoretical fault characteristic frequency.

[0156] Taking rolling bearings as an example, the geometric parameters include the number of rolling elements. , rolling element diameter Bearing pitch diameter and contact angle The motion parameters include rotational speed. Then the bearing rotation frequency Represented as: .

[0157] in, This indicates rotational speed, measured in revolutions per minute. This indicates the bearing rotational frequency, measured in Hertz (Hz). Theoretical characteristic frequency for outer ring faults. Inner ring fault theoretical characteristic frequency Theoretical characteristic frequencies of rolling element failure They are represented as follows: ; ; .

[0158] The candidate fault characteristic frequency set is composed of the theoretical characteristic frequencies of outer ring faults, inner ring faults, and rolling element faults. .

[0159] From the set of candidate fault feature frequencies The minimum value is selected from the values ​​and used as the baseline fault characteristic frequency for adaptive block partitioning. .in, This represents the baseline fault characteristic frequency used to determine the data block window size.

[0160] Selecting the minimum value among the candidate fault characteristic frequencies can make the data block window cover the longest-lasting fault impact cycle among the candidate fault types, reducing the risk of complete fault impact segments being truncated.

[0161] Determine the reference fault characteristic frequency Then, the adaptive block unit calculates the reference fault characteristic frequency. and sampling frequency Calculate the data block window size using the following formula: ;in, Indicates the size of the data block window. Indicates the sampling frequency. Indicates the reference fault characteristic frequency, This represents the periodic tolerance coefficient, and , This indicates rounding up to the nearest integer.

[0162] The periodic tolerance coefficient Used to control the length of the fault impact cycle contained in each data block.

[0163] when At the same time, each data block covers at least one baseline fault cycle; when background noise increases or speed fluctuations increase, the efficiency is improved. To increase the redundancy of fault cycles included in the data blocks, The value range is 1 to 2, preferably 1.2 to 1.5.

[0164] Adaptive block units balance the training set by category The standardized samples in the data are divided into multiple lengths. The data block. Let the input fault diagnosis model's first... The sample is Then its data block sequence is represented as: .

[0165] in, Indicates the first The sequence of data blocks corresponding to each standardized sample Indicates the first The first standardized sample One data block, Indicates the first The number of data blocks obtained by dividing a standardized sample.

[0166] The block embedding layer will embed each data block Mapped to block embedding vectors: .in, Indicates the first The block embedding vector corresponding to each data block This represents the weight matrix of the block embedding layer. Represents the bias vector of the block embedding layer; The positional encoding layer is the embedding vector for each block. Add position encoding vector The resulting block embedding vector with added position encoding is: .

[0167] in, This represents the block embedding vector after adding position encoding. Indicates the first The location-coded vectors corresponding to each data block; the block embedding vector sequence is composed of the block embedding vectors after adding the location codes: .

[0168] The Transformer encoder embeds a sequence of block vectors. Perform self-attention calculations to extract feature correlations across fault impact cycles and output a global feature vector: .

[0169] in, Indicates the first The global feature vector corresponding to each standardized sample.

[0170] The fault classification layer receives the global feature vector. And output the fault category probability vector: ;in, Represents the probability vector of fault categories. This represents the weight matrix of the fault classification layer. This represents the bias vector of the fault classification layer.

[0171] Based on the fault category probability vector The maximum probability component in the model determines the final fault diagnosis result: .

[0172] in Indicates the first The final fault diagnosis results for each sample Indicates the first Each sample belongs to category The predicted probability. The final fault diagnosis result is any one of normal, inner ring fault, outer ring fault, and rolling element fault.

[0173] The fault cycle adaptive block Transformer fault diagnosis model utilizes a class-balanced training set. Once training is complete, the sample to be diagnosed is input into the trained fault diagnosis model, and the corresponding fault diagnosis result will be output.

[0174] This invention addresses intelligent diagnostic scenarios for rotating machinery where a few types of fault samples are scarce. It identifies the region where scarce samples are located through local manifold density, protects weak fault impact characteristics through local manifold density sensing sampling and local manifold density regulation diffusion, generates target fault category samples through a conditional self-modulation generation module, and uses a time-frequency dual-domain discrimination module to constrain the time-domain waveform texture and frequency-domain energy distribution of the generated fault samples. Finally, it accurately outputs the fault diagnosis result, i.e., the corresponding fault category, through a fault cycle adaptive block Transformer diagnostic model.

[0175] To verify the effectiveness of the method of the present invention, two sets of fault diagnosis models were set up for comparison.

[0176] The first set of fault diagnosis models uses the original imbalanced training set obtained in step 1. Training is performed; the second set of fault diagnosis models uses the class-balanced training set formed in step 4 after supplementing with generated fault samples. Conduct training.

[0177] Two sets of fault diagnosis models were evaluated using the same network structure, training parameters, and independent test sets to compare the diagnostic performance differences before and after the generation of fault samples. This embodiment uses overall accuracy, precision, recall, and F1 score as evaluation metrics. Overall accuracy evaluates the overall classification correctness of all test samples, while precision, recall, and F1 score evaluate the recognition performance of each fault category. For any category, the evaluation metrics are expressed as: ; ; ; .

[0178] in, This indicates the number of samples in the test set that were correctly predicted. This represents the total number of samples in the test set. This represents the number of samples that were correctly predicted as belonging to the current class. This indicates the number of samples from other classes that were incorrectly predicted as the current class. This indicates the number of samples in the current category that were incorrectly predicted as other categories.

[0179] The two fault diagnosis models were evaluated on an independent test set. Detailed comparison results are shown in Table 2, and the corresponding confusion matrices are as follows: Figure 5 and Figure 6 As shown. Among them, Figure 5 This indicates that only the original imbalanced training set is used. The confusion matrix of the trained fault diagnosis model. Figure 6 This indicates the use of a class-balanced training set. The confusion matrix of the trained fault diagnosis model.

[0180] Table 2. Comparison of diagnostic performance under different training set conditions

[0181] From Table 2 and Figure 5 It can be seen that using only the original imbalanced training set The trained fault diagnosis model has relatively insufficient ability to identify minority class fault samples. Although the overall accuracy of the model reaches 0.919, the recall rate for the inner ring fault category is only 0.75, indicating that the model is prone to misclassifying some fault categories when the number of minority class fault samples is limited. For the rolling element fault category, the recall rate is 0.93, indicating that some fault samples are still misclassified.

[0182] From Table 2 and Figure 6 It can be seen that the method of this invention generates fault samples and constructs a class-balanced training set. Subsequently, the overall accuracy of the fault diagnosis model improved to 0.988. Specifically, the recall rate for the inner race fault category increased from 0.75 to 0.95, and the recall rate for the rolling element fault category increased to 1.00. The F1 score for each fault category reached 0.98 or higher. These results demonstrate that, under conditions of scarce minority class fault samples, the fault samples generated by the method of this invention can supplement the minority class information in the original imbalanced training set, thus helping to improve the fault diagnosis model's ability to identify minority class faults.

[0183] This invention first performs uniform manifold approximation and projection mapping on minority class fault samples to obtain two-dimensional low-dimensional feature points, which then constitute a low-dimensional manifold space. Subsequently, local manifold density is calculated through grid partitioning and used for hierarchical batch sampling and manifold density-controlled diffusion noise addition. During the adversarial training phase, a conditional self-modulation generation module generates generated fault samples corresponding to the target fault category based on the noise-added samples, diffusion time steps, and fault labels. A time-frequency dual-domain discrimination module performs sample scoring and multi-layer feature response constraints on the generated fault samples through time-domain and frequency-domain discrimination branches. Finally, the generated fault samples are fused with real standardized samples to form a class-balanced training set, and a fault diagnosis result is output using a fault cycle adaptive block Transformer fault diagnosis model. Therefore, it can improve the training sample distribution of the fault diagnosis model under conditions of data imbalance and scarcity of minority class fault samples, thereby improving the reliability of rotating machinery fault diagnosis.

[0184] Example 2 This embodiment 2 describes a computer device including a memory and one or more processors. Executable code is stored in the memory. When the processor executes the executable code, it implements the steps of the fault diagnosis method based on manifold density modulation and time-frequency dual-domain self-modulation in embodiment 1 above.

[0185] In this embodiment, the computer device can be any device or apparatus with data processing capabilities, and will not be described in detail here.

[0186] Example 3 This embodiment 3 describes a computer-readable storage medium storing a program that, when executed by a processor, is used to implement the steps of the fault diagnosis method based on manifold density regulation and time-frequency dual-domain self-modulation in embodiment 1.

[0187] The computer-readable storage medium can be an internal storage unit of any device or apparatus with data processing capabilities, such as a hard disk or memory, or an external storage device of any device with data processing capabilities, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc.

[0188] Of course, the above description is only a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. It should be noted that any equivalent substitutions or obvious modifications made by those skilled in the art under the guidance of this specification fall within the scope of this specification and should be protected by the present invention.

Claims

1. A fault diagnosis method based on manifold density regulation and time-frequency dual-domain self-modulation, characterized in that, Includes the following steps: Step 1. Obtain the original time-series vibration signals of the key components of the rotating machinery to be diagnosed, and perform sliding window truncation and labeling to obtain the original sample set; the original sample set is then standardized by standardization preprocessing to form a standardized sample set; Step 2. Map the minority class fault samples in the standardized samples to a low-dimensional manifold space to obtain the corresponding low-dimensional feature points; divide the low-dimensional manifold space into grid cells and calculate the local manifold density of each grid cell; Step 3. Construct a generative adversarial network that includes a conditional self-modulation generation module and a time-frequency dual-domain discrimination module; The input to the conditional self-modulation generation module is the noisy sample processed by the manifold density-controlled diffusion strategy, the fault label, and the diffusion time step. The output of the conditional self-modulation generation module is the generated fault sample. The manifold density-controlled diffusion strategy adjusts the noise intensity during the forward diffusion process based on the local manifold density; The time-frequency dual-domain discrimination module receives generated fault samples and real standardized fault samples, and extracts time-domain multi-layer feature response maps, frequency-domain multi-layer feature response maps and sample scores through time-domain discrimination branches and frequency-domain discrimination branches, respectively. During the training of the generative adversarial network, a hierarchical batch sampling strategy is adopted, which adjusts the probability of minority class fault samples in the training batch according to the local manifold density. Step 4. Use the trained conditional self-modulation generation module to generate fault samples for the minority fault categories, and fuse the generated fault samples with the real standardized samples in the standardized sample set to form a class-balanced training set. A fault diagnosis model is trained using a class-balanced training set to obtain a trained fault diagnosis model. The sample to be diagnosed is then input into the trained fault diagnosis model to output the fault diagnosis result.

2. The method according to claim 1, wherein, Step 1 specifically involves: Define the original time-series vibration signal sequence for: ;in This represents the number of data points in the original vibration signal. , … These represent the first and second digits of the original time-series vibration signal, respectively. One sampling point; The original time-series vibration signal sequence was processed using a sliding window method. The process involves truncating the sample to obtain multiple fragments of equal length; let the window length be... The sliding step size is Then the first Sample fragments Represented as: ; wherein ; , This indicates the number of sample segments obtained from the original time-series vibration signal; Based on the device operating status label, for each sample segment Assign corresponding category labels Category tags This includes any one of the following: normal condition, inner ring fault, outer ring fault, and rolling element fault; The original sample set consists of multiple sample fragments of the same length and their corresponding class labels: ; For the original sample set Each sample fragment The standardization process is performed using the following formula: ; in Represents sample fragments Standardized samples after standardization processing; Represents sample fragments The mean, Represents sample fragments standard deviation To represent the standardized smoothing coefficient used to prevent the denominator from being zero; From a standardized sample and their corresponding category labels Forming a standardized sample set : .

3. The fault diagnosis method based on manifold density regulation and time-frequency dual-domain self-modulation as described in claim 1, characterized in that, Step 2 specifically involves: Step 2.

1. From the standardized sample set By selecting minority class failure samples, a minority class failure sample set is obtained: ； in This represents a minority class of fault sample set. Representing the A few minority of fault samples, express Corresponding fault category labels, This represents the total number of minority class fault samples; Minority failure samples are failure category samples whose sample size is lower than that of normal category samples, including at least one of inner ring failure, outer ring failure and rolling element failure samples; Uniform manifold approximation and projection algorithms are used for minority class fault sample sets. Standardized samples Perform dimensionality reduction mapping to standardize samples Mapped to two-dimensional low-dimensional feature points The formula is as follows: ； in This represents the mapping function between the uniform manifold approximation and projection algorithms. Indicates the first Two-dimensional low-dimensional feature points corresponding to a minority of fault samples and They represent Coordinate components in the first and second dimensions; Low dimensional manifold space consisting of all two-dimensional low dimensional feature points : ； wherein, for characterizing the local neighborhood distribution relationship of the minority class fault samples in the dimension reduction feature space; Step 2.

2. For the low-dimensional manifold space The mesh is generated to obtain a set of mesh cells: ； in Represents any grid cell, Indicates the number of grid cells; for any grid cell Statistical analysis of cells falling into the grid Number of minority fault samples ,in: ； in Indicates the indicator function; when two-dimensional low-dimensional feature points Falling into the grid cell hour, ,otherwise Calculate the mesh element according to the following formula. Local manifold density: ； in Represents grid cells Local manifold density, This represents the density smoothing coefficient in the calculation of local manifold density.

4. The fault diagnosis method based on manifold density regulation and time-frequency dual-domain self-modulation as described in claim 3, characterized in that, In step 2.1, the processing flow of the uniform manifold approximation and projection algorithm is as follows: With a minority of fault sample sets Standardized samples As a high-dimensional input point, calculate any two standardized samples and Distance between: ; in Indicates the preset distance metric function. Standardized samples and The sample distance between them; For each standardized sample According to sample distance Sort in ascending order, starting from the smallest. Select from other standardized samples We obtain the nearest neighbor set by taking samples from each nearest neighbor. ； in, Standardized samples of Nearest neighbor set Indicates the number of neighbors; This represents the standardized samples sorted in ascending order of sample distance. The Nearest neighbor samples, ; Construct high-dimensional nearest neighbor weights based on the nearest neighbor set: ； in, Indicated by standardized samples When centered, standardized samples Compared to standardized samples The directed high-dimensional nearest neighbor weights; Show and standardized samples The corresponding local distance scaling parameters; Indicated by standardized samples When centered, standardized samples Compared to standardized samples The directed high-dimensional nearest neighbor weights; By swapping the indices in the directed high-dimensional nearest neighbor weight calculation formula and get; The directed high-dimensional nearest neighbor weights are symmetricized to obtain symmetric nearest neighbor weights: ； in, Standardized samples and Symmetric nearest neighbor weights between them; Each standardized sample Mapped to two-dimensional low-dimensional feature points: ； in This represents the mapping function between the approximation and projection of a uniform manifold. Indicates the first Two-dimensional low-dimensional feature points corresponding to a minority of fault samples and Representing two-dimensional low-dimensional feature points respectively Coordinate components in the first and second dimensions; In low-dimensional space, compute low-dimensional similarity between two-dimensional feature points and ​ ； in, Representing two-dimensional low-dimensional feature points and Low-dimensional similarity, and These are parameters used to control the shape of the low-dimensional similarity curve; to maintain the high-dimensional nearest neighbor weights. Compared with low-dimensional similarity With consistency as the goal, construct an optimization objective function: ； By minimizing the optimization objective function Iteratively update each two-dimensional low-dimensional feature point This ensures that the neighborhood relationships between low-dimensional feature points in the two dimensions maintain the local neighborhood relationships of minority class fault samples in the high-dimensional space. Output all two-dimensional low-dimensional feature points to form a low-dimensional manifold space: .

5. The fault diagnosis method based on manifold density regulation and time-frequency dual-domain self-modulation as described in claim 4, characterized in that, In step 3, the manifold density-controlled diffusion strategy is specifically as follows: From a minority of fault sample sets Standardized minority class fault samples were selected as the initial diffusion samples, denoted as . ;in For standardized minority class fault samples participating in forward diffusion noise addition, For the first A standardized minority class of fault samples; Based on the low-dimensional mapping relationship in step 2 Determine the initial sample for diffusion Corresponding two-dimensional low-dimensional feature points And determine the two-dimensional low-dimensional feature points. The associated grid cell is denoted as ; Let the diffusion time step be denoted by ; in, , Indicates the maximum diffusion time step; based on the initial diffusion sample Belonging grid cell Local manifold density For the preset noise variance After correction, the noise variance after density adjustment is obtained: ； in Indicates the initial sample of diffusion At the diffusion time step The noise variance after local manifold density modulation; Indicates the initial sample of diffusion Belonging grid cell Local manifold density, Represents the protection coefficient of scarce features and , Represents the density scaling factor and ; In the forward diffusion process, the retention coefficient after density modulation is defined as: ； wherein, represents the diffusion initial sample the density regulated signal preservation factor after the diffusion time step under. And define from the first diffusion time step to the second... The cumulative retention coefficient for each diffusion time step is: ； in Indicates the initial sample of diffusion In the Signal retention coefficient after density modulation at each diffusion time step Indicates the cumulative multiplication index of the diffusion time steps; Initial sample for diffusion From the first diffusion time step to the second The cumulative signal retention coefficient for each diffusion time step; Based on cumulative retention factor The diffusion time step is obtained. The following noisy samples: ； in, Indicates the initial sample of diffusion At the diffusion time step The generated noisy samples, Indicates Gaussian noise, and ; Indicates the initial sample of diffusion An identity matrix that matches the dimensions.

6. The fault diagnosis method based on manifold density regulation and time-frequency dual-domain self-modulation as described in claim 1, characterized in that, In step 3, the conditional self-modulation generation module includes a time step embedding branch, a fault label embedding branch, a conditional fusion layer, multiple one-dimensional convolutional generation blocks, multiple self-modulation normalization layers, and an output convolutional layer. The conditional self-modulation generation module receives the noisy sample processed by the manifold density-controlled diffusion strategy. diffusion time step and fault labels and output generated fault samples. ; Time step embedding branch is used to diffuse time steps Mapped to time step embedding vectors: ;in, Indicates the time step embedding function, Indicates diffusion time step The corresponding time step embedding vector; Fault tag embedding branch is used to embed fault tags Mapped to fault label embedding vectors: ;in, This indicates the function for embedding fault tags. Indicates fault label The corresponding fault label embedding vector; Will and Vector concatenation is performed, and a conditional fusion layer is introduced to obtain the global conditional control vector: ;in, This represents the global conditional control vector. This represents a multilayer perceptron. This indicates vector concatenation; Multiple one-dimensional convolutional generation blocks are arranged in series along the one-dimensional time-series waveform generation direction. Each one-dimensional convolutional generation block includes a one-dimensional convolutional layer and a nonlinear activation layer, and a self-modulation normalization layer is connected after each one-dimensional convolutional generation block. Let the first The feature map output by each one-dimensional convolutional generator block is Then the first Each self-modulation normalization layer controls the vector according to the global condition. Prediction scaling parameters and bias parameters: ; in Indicates the first The parameter prediction network corresponding to each self-modulation normalization layer Indicates the first Scaling parameters of the output of the self-modulation normalization layer. Indicates the first The bias parameters of the output of the self-modulated normalization layer; Using scaling parameters and bias parameters , for the Feature maps output by a one-dimensional convolutional generator block Perform self-modulation normalization processing to obtain the modulated feature map. The formula is as follows: ； in, Representation of feature map The mean, Representation of feature map standard deviation This represents element-wise multiplication. This represents the modulation smoothing coefficient used to prevent the denominator from being zero during the self-modulation normalization process; After processing through multiple layers of one-dimensional convolutional generating blocks and self-modulation normalization layers, the output convolutional layer will output the deep features from the last self-modulation normalization layer. Mapping to a one-dimensional time-series waveform yields the generated fault samples corresponding to the target fault category: ； in This indicates the generation of fault samples. This indicates the convolution kernel parameters of the output convolutional layer. This represents a one-dimensional convolution operation. This represents the bias vector of the output convolutional layer. This indicates the number of layers in the one-dimensional convolutional generation block kernel self-modulation normalization layer.

7. The method according to claim 3, wherein, In step 3, the hierarchical batch sampling strategy is specifically as follows: Based on the mesh cell set obtained in step 2 and the number of minority fault samples in each grid cell Determine the set of non-empty grid cells with a sample size greater than zero: ; in This represents a non-empty set of grid cells containing at least one minority class fault sample. express Any non-empty grid cell in the array; For any non-empty grid cell Based on its local manifold density Calculate the sampling probability of this grid cell: ； in Indicates from non-empty grid cells The sampling probability of extracting minority class fault samples from the sample. express Any grid cell used for summation traversal, Indicates non-empty grid cells Local manifold density, Represents the sampling probability smoothing coefficient; When constructing each training batch, first follow the sampling probability. from Select a target grid cell, and then randomly extract training samples from the minority class fault samples within the target grid cell until the minority class fault samples required for the current training batch are obtained.

8. The method according to claim 1, wherein, In step 3, the time-frequency dual-domain discrimination module is used to receive real standardized fault samples. and generating fault samples It includes a time-domain discrimination branch, a frequency-domain discrimination branch, and a feature matching loss calculation unit; The temporal discriminant branch includes a one-dimensional convolutional unit, a temporal multi-layer feature response map extraction unit, and a temporal sample scoring unit; One-dimensional convolutional unit receives and The temporal waveform is obtained, and temporal convolution features are extracted; The temporal multi-layer feature response map extraction unit outputs temporal multi-layer feature response maps from different network layers of temporal convolutional features; the temporal sample scoring unit outputs temporal sample scores based on the temporal multi-layer feature response maps; The time domain discrimination branch processes the real normalized fault samples The processing result is represented as: , ; Time-domain discriminant branch generates fault samples The processing result is expressed as follows: , ; in Represents real standardized fault samples In the time-domain discrimination branch The temporal characteristic response map output by the layer. Indicates the generation of fault samples In the time-domain discrimination branch Temporal characteristic response map of the layer output; This indicates the number of network layers used for feature matching in the temporal discriminant branch; and These represent the time-domain sample scores of real standardized fault samples and generated fault samples, respectively. The frequency domain discrimination branch includes a Fast Fourier Transform (FFT) unit, an amplitude spectrum calculation unit, a frequency domain multilayer feature response map extraction unit, and a frequency domain sample scoring unit; the FFT unit respectively performs... and Perform frequency domain transformation; The amplitude spectrum calculation unit calculates the frequency domain amplitude spectrum based on the frequency domain transformation result; the frequency domain multi-layer feature response map extraction unit outputs the frequency domain multi-layer feature response map from different network layers of the frequency domain amplitude spectrum; The frequency domain sample scoring unit outputs a frequency domain sample score based on the frequency domain multi-layer feature response map. Real normalized fault samples and generating fault samples The frequency domain amplitude spectrum of the real normalized fault samples is represented as: , ;in, Represents the Fast Fourier Transform. Represents real standardized fault samples The frequency domain amplitude spectrum, Indicates the generation of fault samples The frequency domain amplitude spectrum; Frequency domain discriminant branch for real standardized fault samples The processing result is expressed as follows: , ; Frequency domain discriminant branch generates fault samples The processing result is expressed as follows: , ; in Represents real standardized fault samples In the frequency domain discrimination branch The frequency domain characteristic response map of the layer output. Indicates the generation of fault samples In the frequency domain discrimination branch Frequency domain characteristic response map of the layer output; This indicates the number of network layers used for feature matching in the frequency domain discrimination branch; and These represent the frequency domain sample scores of real standardized fault samples and generated fault samples, respectively. The time-frequency dual-domain discrimination module obtains a comprehensive sample score based on the time-domain sample score and the frequency-domain sample score: ； in The comprehensive sample score representing the true standardized fault samples. This represents the overall sample score for generating faulty samples. and Let represent the weight coefficients of the time-domain sample scores and the frequency-domain sample scores, respectively, and satisfy the following: ； The comprehensive sample score is used to form the true and false discrimination results between real standardized fault samples and generated fault samples, and is used to construct the discriminator adversarial loss and the generator adversarial loss; The time-domain multi-layer feature response map and the frequency-domain multi-layer feature response map are used to calculate the time-frequency dual-domain hierarchical feature matching loss; During training, the discriminator adversarial loss is used to update the network parameters of the time-frequency dual-domain discriminator module; the generator adversarial loss and the time-frequency dual-domain hierarchical feature matching loss are used together to update the network parameters of the conditional self-modulation generator module.

9. The method according to claim 8, wherein, In step 3, the time-frequency dual-domain hierarchical feature matching loss is used to measure the true standardized fault samples. With generating fault samples Differences in multilayer feature response maps between the time-domain and frequency-domain discriminant branches; Suppose a training batch contains There are 1 real standardized fault samples and their corresponding generated fault samples, where the real standardized fault samples and generated fault samples represent: ;in, Indicates the first A real, standardized fault sample. Indicates the first One fault sample is generated, and and Having the same fault category label ; Time-frequency dual-domain hierarchical feature matching loss Defined as: ； in Indicates the number of sample pairs in the training batch; The time-domain discriminant branch is represented by the first... Hierarchical weights of the layer feature response map; The frequency domain discriminant branch is represented by the first... Hierarchical weights of the layer feature response map; express Norm; Hierarchical weights satisfy: , , , ; Time-frequency dual-domain hierarchical feature matching loss updating the network parameters of the conditional self-modulation generation module; Generator loss function of conditional self-modulation generation module Defined as: ； in This represents the generator adversarial loss formed by the comprehensive sample score output by the time-frequency dual-domain discrimination module. The weight coefficients of the time-frequency dual-domain hierarchical feature matching loss are represented by... .

10. The fault diagnosis method based on manifold density regulation and time-frequency dual-domain self-modulation as described in claim 1, characterized in that, In step 4, the processing flow of the fault diagnosis model is as follows: A fault diagnosis model with adaptive block segmentation based on fault cycle is constructed, which includes a fault cycle estimation unit, an adaptive block segmentation unit, a block embedding layer, a position encoding layer, a Transformer encoder, and a fault classification layer. The fault cycle estimation unit calculates the theoretical fault characteristic frequencies corresponding to different candidate fault types based on the geometric and motion parameters of the tested rotating part, and forms a set of candidate fault characteristic frequencies from the theoretical fault characteristic frequencies. Based on the candidate fault characteristic frequency set, determine the baseline fault characteristic frequency for adaptive block division. ; Adaptive segmentation unit based on reference fault characteristic frequency and sampling frequency Calculate the data block window size using the following formula: ;in, Indicates the size of the data block window. Indicates the sampling frequency. Indicates the reference fault characteristic frequency, This represents the periodic tolerance coefficient, and , Indicates rounding up; Periodic tolerance coefficient Used to control the length of the fault impact cycle contained in each data block; The adaptive block unit divides the standardized samples in the class-balanced training set into multiple blocks of length 1. From the data blocks, we obtain a data block sequence: ; in, Indicates the first The sequence of data blocks corresponding to each standardized sample Indicates the first The first standardized sample One data block, Indicates the first The number of data blocks obtained by dividing a standardized sample; The block embedding layer will embed each data block Mapped to block embedding vectors: ;in Indicates the first The block embedding vector corresponding to each data block This represents the weight matrix of the block embedding layer. Represents the bias vector of the block embedding layer; The positional encoding layer is the embedding vector for each block. Add position encoding vector This yields the block embedding vector after adding position encoding. : ;in For the first The location encoding vector corresponding to each data block; A block embedding vector sequence is formed by adding position-encoded block embedding vectors. : ; The Transformer encoder embeds a sequence of block vectors. Perform self-attention calculations to extract feature correlations across fault impact cycles and output a global feature vector: ; wherein, represents the global feature vector corresponding to the i-th standardized sample; represents the global feature vector corresponding to the i-th standardized sample; The failure classification layer receives the global feature vector and outputs a failure class probability vector : ； in This represents the weight matrix of the fault classification layer. Represents the bias vector of the fault classification layer; according to Output any one of the following categories as the fault diagnosis result: normal, inner ring fault, outer ring fault, and rolling element fault.

Images