Part fault diagnosis method based on feature excitation type sample expansion under data scarcity

Abstract

The application discloses a part fault diagnosis method based on feature excitation type sample expansion under data scarcity. In view of the problem that the model is prone to overfitting and has poor generalization performance due to the scarcity of fault samples of the fault diagnosis model of mechanical and electrical product parts, a feature excitation type fault sample generation generative adversarial network is constructed to realize fault sample expansion. In the generator, the prior features filtered by the feature sensitivity and redundancy evaluation criteria are integrated by using a skip-layer feature excitation module, and cross-layer feature maps are fused by using the skip-layer feature excitation module, so that deep features at different levels are fully extracted, and the generation of fault samples is better guided. The feature pre-extractor based on contrast learning is cascaded with the discriminator, the pre-learned features with more obvious category boundary attributes are input into the discriminator, and the training speed is improved. The training set is expanded by using the fault samples generated by the generator, the classification accuracy and generalization performance of the model are improved, and the robust and accurate fault diagnosis of mechanical and electrical product parts under the condition of scarce fault data is realized.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis of mechanical product parts, and relates to a fault diagnosis method for parts based on feature-excited sample amplification under data scarcity. Background Technology

[0002] The health status of critical components in electromechanical equipment directly impacts its safe and stable operation. During operation, these critical components are highly susceptible to wear and surface damage. Fault diagnosis of electromechanical components can significantly reduce maintenance and downtime costs for production lines and equipment. Condition monitoring data for electromechanical components typically originates from the component's health stage, the fault initiation and development stage, and the failure stage. While components remain in a healthy state for most of their lifespan, making it easy to collect a large amount of vibration signal data, continued use of fatigue-damaged components threatens the healthy operation of the equipment. In actual industrial production lines, most component fault information collected is pre-fault information, resulting in a sparse sample size and insufficient labeled training samples. Deep neural networks require sufficient labeled fault data for supervision during fault recognition; insufficient training data can lead to overfitting and poor model generalization performance. Furthermore, scarce fault features are easily overwhelmed by noise, making feature extraction difficult and resulting in poor fault diagnosis performance.

[0003] To address the scarcity of fault data, generative methods can be used to expand it. Traditional generative methods often employ Generative Adversarial Networks (GANs) to generate pseudo-samples, but these networks often suffer from poor convergence, high time costs, pattern collapse, convergence of generated samples, vanishing gradients, and training failures. To address the convergence problem, feature projection and multi-discriminator multi-scale feedback are applied to the discriminator, allowing it to utilize all available information in a balanced manner. Pre-training techniques are applied to upstream tasks, enabling the GAN model to learn more data variations, thereby improving its generalization ability on the test set and accelerating network iteration. Among existing pre-training techniques, contrastive learning, as a self-supervised learning strategy, compares the similarity or difference between different samples, making the distinction between positive and negative samples more obvious in the feature space and improving downstream task performance. Simultaneously, when building the GAN, the pre-trained representation capabilities are used to improve the discriminator, providing better feedback to the generator and improving the quality of generated data and model convergence speed.

[0004] Therefore, it is crucial to invent a diagnostic method that addresses the poor generalization and low accuracy issues caused by the limited number of fault samples in real-world industrial production. This method combines feature pre-learning with collaborative feature generation to achieve an accurate fault diagnosis method based on a limited number of fault samples. Summary of the Invention

[0005] To address the problems mentioned in the background art, this invention proposes a component fault diagnosis method based on feature-excited sample amplification under data scarcity, which enables fault diagnosis of key components of electromechanical products when the number of fault samples is insufficient, and can be widely applied to fault diagnosis of electromechanical product components with scarce samples.

[0006] To achieve the above functions, the technical solution of the present invention is as follows:

[0007] S1: Collect vibration monitoring signals of key components of electromechanical products, and then construct the original vibration sample dataset;

[0008] S2: After extracting and filtering features from the original vibration sample dataset using the prior feature extraction module, a subset of prior features after feature filtering is obtained;

[0009] S3: After training the feature-excited fault sample amplification module using the original vibration sample dataset and the prior feature subset after feature filtering, a well-trained feature-excited fault sample amplification module is obtained, thus obtaining a well-trained generator.

[0010] S4: Input the original vibration samples of each fault type into the trained generator to obtain the amplified vibration samples under the fault type. The original vibration samples and the corresponding amplified vibration samples constitute the final vibration sample dataset.

[0011] S5: After training the fault diagnosis model using the final vibration sample dataset, a trained fault diagnosis model is obtained; the vibration sample to be detected is input into the trained fault diagnosis model, and the corresponding fault diagnosis result is output.

[0012] The vibration monitoring signals of the key components of the electromechanical products include vibration signals in both the horizontal and vertical directions monitored by the accelerometer.

[0013] In S2, the prior feature extraction module is composed of a feature extraction module and a two-stage feature index screening module connected together. The feature extraction module extracts multiple time-domain features and frequency-domain features of the input original vibration sample to obtain the original prior feature set.

[0014] In the two-stage feature index screening module, firstly, Fisher scores are calculated for each feature based on the original prior feature set. Based on the Fisher scores, features are initially screened to obtain features with high fault sensitivity and good discriminative power. Next, Pearson correlation coefficients are calculated for each feature. Based on the Pearson correlation coefficients, the features after initial screening are sorted in descending order, and the feature with the highest discriminative power is retained as the benchmark feature. Then, the correlation between the features after initial screening and the benchmark feature is calculated. Features with a correlation less than the correlation threshold are retained to obtain the features after secondary screening and are recorded as the prior feature subset after feature screening.

[0015] The time-domain features include mean, standard deviation, root mean, absolute mean, skewness, kurtosis, variance, maximum value, minimum value, and peak-to-peak value.

[0016] In S3, the feature-excited fault sample amplification module includes a generator G, a feature pre-extractor F, and a discriminator. The prior feature subset after feature screening is used as the input of the generator G. The generator G is connected to the discriminator after passing through the feature pre-extractor F. The input of the generator G also includes random noise and labels. The original vibration sample dataset is used as the input of the feature pre-extractor F.

[0017] The generator G includes an initialization module, an upsampling module, and a skip-level feature activation module (SKEM). Random noise and labels are input into the initialization module. Prior feature subsets after feature filtering are used as inputs to the skip-level feature activation module (SKEM). Each initialization module is connected to the first upsampling module and the first skip-level feature activation module (SKEM). The first upsampling module is connected to the fourth upsampling module via the second and third upsampling modules. The fourth upsampling module is connected to the first skip-level feature activation module (SKEM). The first skip-level feature activation module (SKEM) is connected to the fifth upsampling module. Both the first and fifth upsampling modules are connected to the second skip-level feature activation module (SKEM). The second skip-level feature activation module (SKEM) is connected to the sixth upsampling module. Both the block and the sixth upsampling module are connected to the third skip-layer feature excitation module SKEM. The third skip-layer feature excitation module SKEM is connected to the seventh upsampling module. Both the third and seventh upsampling modules are connected to the fourth skip-layer feature excitation module SKEM. The fourth skip-layer feature excitation module SKEM is connected to the tenth upsampling module after passing through the eighth and ninth upsampling modules. The outputs of the initialization module, the first upsampling module, the second upsampling module, and the third upsampling module are all used as the second input feature map of the corresponding skip-layer feature excitation module SKEM. The outputs of the fourth, fifth, sixth, and seventh upsampling modules are all used as the first input feature map of the corresponding skip-layer feature excitation module SKEM. The output of the tenth upsampling module is denoted as the amplified vibration sample.

[0018] The Skip-Layer Feature Activation Module (SKEM) includes a first adaptive average pooling layer, a convolutional layer, an activation layer, and a linear layer. The second input feature map of the Skip-Layer Feature Activation Module (SKEM) is connected to the second convolutional layer after passing through the first adaptive average pooling layer, the first convolutional layer, and the first activation layer. The prior feature subset after feature filtering is connected to the second linear layer after passing through the first linear layer and the second activation layer. The output of the second convolutional layer is added to the output of the second linear layer and then input to the third activation layer. The output of the Skip-Layer Feature Activation Module (SKEM) is the product of the first input feature map of the Skip-Layer Feature Activation Module (SKEM) and the output of the third activation layer.

[0019] In the feature-driven fault sample amplification module, there are four feature pre-extractors F and four discriminators. Each feature pre-extractor F consists of a feature extraction network f. θ and feature mixing module f mix The input of the feature-stimulated fault sample amplification module is used as the input of the first feature pre-extractor F, which is connected to form a feature extraction network f. θ The output is used as the feature extraction network f of the second graph feature pre-extractor F. θ The input is the second feature pre-extractor F, and the feature extraction network f. θ The output is used as the feature extraction network f of the third graph feature pre-extractor F. θ The input is the feature extraction network f of the third feature pre-extractor F. θ The output is used as the feature extraction network f of the fourth graph feature pre-extractor F. θ The input, the feature mixing module f of the fourth feature pre-extractor F mix The output of the fourth discriminator is used as the input of the feature mixing module f of the fourth feature pre-extractor F. mix The output is also used as the feature mixing module f of the third feature pre-extractor F. mix The input, the feature mixing module f of the third feature pre-extractor F mix The output of the third discriminator is used as the input of the feature mixing module f of the third feature pre-extractor F. mix The output is also used as the feature mixing module f of the second feature pre-extractor F. mix The input, the feature mixing module f of the second feature pre-extractor F mix The output of the second discriminator is used as the input of the feature mixing module f of the second feature pre-extractor F. mix The output is also used as the feature mixing module f of the first feature pre-extractor F. mix The input is the feature mixing module f of the first feature pre-extractor F. mix The output of the first discriminator is used as the input of the second discriminator. The average of the outputs of the four discriminators is then used as the final output of the discriminator.

[0020] The feature extraction network f θ It is the feature extraction network f in the online network of the deep feature pre-learning module based on contrastive learning after pre-training. θ During the training of S3, the feature extraction network f θ The weights remain constant.

[0021] Each discriminator includes N downsampling modules, a global summing pooling layer, an embedding layer, an adaptive average pooling layer, and a third linear layer. The N downsampling modules are cascaded sequentially. The feature map of the input discriminator is used as the input of the first downsampling module, and the label of the input discriminator is used as the input of the embedding layer. The last downsampling module is connected to the third linear layer after passing through the global summing pooling layer. The output of the global summing pooling layer is multiplied element-wise with the output of the embedding layer and then input into the adaptive average pooling layer. The output of the third linear layer is added to the output of the adaptive average pooling layer to output a sub-discriminator vector, which is used as the output of the discriminator.

[0022] The feature mixing module f mix It consists of cross-channel blending and cross-scale blending. Cross-channel blending interacts and blends the various channels of the feature map. Cross-scale blending is calculated in a top-down order. After upsampling the small-sized feature map above, it is added to the feature map to complete the cross-scale blending.

[0023] In S5, the fault diagnosis model is based on a multi-scale convolutional residual network, specifically consisting of an input layer, a fifth convolutional layer, a first batch of normalization layers, a SELU nonlinear activation function layer, n layers of multi-scale convolutional residual blocks with different kernel sizes, a flattening layer, and an output layer connected in sequence; the multi-scale convolutional residual block includes a first channel-wise convolutional layer, a second batch of normalization layers, a first SELU nonlinear activation function layer, a second channel-wise convolutional layer, a third batch of normalization layers, and a second SELU nonlinear activation function layer;

[0024] The input of the multi-scale convolutional residual block is used as the input of the first channel-wise convolutional layer. The first channel-wise convolutional layer is connected to the third batch of normalization layers after passing through the second batch of normalization layers, the first SELU nonlinear activation function layer, and the second channel-wise convolutional layer. The output of the third batch of normalization layers is added to the input of the multi-scale convolutional residual block and then input into the second SELU nonlinear activation function layer. The output of the second SELU nonlinear activation function layer is used as the output of the multi-scale convolutional residual block.

[0025] This invention addresses the poor generalization performance of data-driven fault diagnosis models for key components of electromechanical products due to the scarcity of fault samples. It employs a sliding window to divide a single, collected vibration signal segment into multiple, non-overlapping vibration signal samples, and then sequentially performs random data augmentation, standardization, feature extraction, and feature fusion on these samples. A multi-domain feature set is constructed, and feature indicators are selected using feature sensitivity and redundancy evaluation criteria. During generative adversarial network training, expert prior features are used to stimulate the generator, guiding the generation of fault samples. Contrastive learning is performed using the consistency latent features of the samples to complete fault feature pre-learning, providing preliminary extracted fault features for subsequent fault data augmentation. The pre-learned deep features help the discriminator identify the authenticity of fault samples, and finally, the generator is used to augment the component fault data. The augmented samples are used for downstream diagnostic tasks, improving the model's classification accuracy and generalization performance, achieving robust and accurate component fault diagnosis under scarce sample conditions.

[0026] Compared with existing technologies, the present invention has the following advantages:

[0027] Compared to traditional methods, the proposed method constructs a feature-incentivized fault sample generative adversarial network (GAN). Prior features are incorporated into the generator to guide fault sample generation. A pre-trained network is used to construct a feature pre-extraction module based on contrastive learning. The extracted pre-learned features, with more pronounced class boundary attributes, are input into the discriminator. Compared to directly inputting fault samples into the discriminator, the use of multi-layer pre-learned features reduces the learning burden on the discriminator, resulting in faster training speeds compared to traditional GANs. A skip-layer feature excitation module is designed in the generator, enabling the reuse of features from different levels through the fusion of cross-layer feature maps. Embedding prior features utilizes the expert prior knowledge contained within them to provide appropriate excitation during the network's sample generation process, better guiding fault sample generation. During training, both the generator and discriminator receive sample label information. After training, fault samples generated by the generator, with the fault category that can be specified, are effectively and specifically addressed to solve the problem of fault sample scarcity, improving the accuracy and generalization ability of downstream classification tasks.

[0028] This invention addresses the scarcity of fault samples in industrial production lines by using the deep features of a limited set of fault samples to stimulate a generative network, generating a sufficient number of artificial samples containing effective fault features. Contrastive learning is then employed to improve the speed of artificial fault sample generation. The expanded sample set effectively and specifically alleviates the problem of poor accuracy and generalization ability in data-driven diagnostic models caused by the scarcity of fault samples, demonstrating both theoretical innovation and practicality. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating the method of the present invention.

[0030] Figure 2 This is the overall model structure of the present invention.

[0031] Figure 3 This is a structural diagram of the generator in the feature-driven fault sample amplification network module proposed in this invention.

[0032] Figure 4 This is a structural diagram showing the collaborative operation of the feature pre-extractor and discriminator in the feature-excited fault sample amplification network module proposed in this invention.

[0033] Figure 5 This is a comparison chart of real data and generated data after fault sample amplification in an embodiment of the present invention; where (a) is a real inner ring fault sample, (b) is a generated inner ring fault sample, (c) is a real cage fault sample, (d) is a generated cage fault sample, (e) is a real outer ring fault sample, and (f) is a generated outer ring fault sample. Detailed Implementation

[0034] The present invention will be further described below with reference to the accompanying drawings and specific embodiments:

[0035] like Figure 1 As shown, specific embodiments of the present invention are as follows:

[0036] S1: Collect vibration monitoring signals from key components of electromechanical products to construct an original vibration sample dataset. Specifically, install accelerometers near the key components and collect vibration monitoring signals under normal and fault conditions. Store, integrate, and label the vibration monitoring signal data for fault types, and divide the data into training and test sample sets. Specifically, initially summarize and integrate the sensor monitoring data, and label the fault type of the component to obtain fault type labels. For each collected vibration signal segment, use a sliding window to divide it into multiple non-overlapping samples, perform standardized preprocessing, and obtain training and test sample sets.

[0037] Vibration monitoring signals for key components of electromechanical products include vibration signals in both horizontal and vertical directions monitored by accelerometers.

[0038] This embodiment uses the XJTU-SY publicly available bearing dataset to verify the effectiveness of the fault diagnosis method for scarce sample parts generated considering fault feature excitation. The test bearing model is LDK UER204. The sampling frequency in the experiment was 25.6 kHz, the sampling duration was 1.28 s, and the data length of the vibration signal in each sampling file was 32768. As shown in Table 1, healthy and faulty data files were selected, and a sliding window of length 2048 was used to divide the vibration signals into non-overlapping segments to obtain healthy and faulty samples.

[0039] Table 1 shows the usage of XJTU-SY data.

[0040]

[0041] S2: As Figure 2 As shown, after extracting and filtering features from the original vibration sample dataset using the prior feature extraction module, a subset of prior features after feature filtering is obtained. The prior feature extraction module makes the subset features as sensitive to the fault category as possible and minimizes the redundancy between features.

[0042] The prior feature extraction module consists of a feature extraction module and a two-stage feature index screening module. The feature extraction module extracts multiple time-domain and frequency-domain features from the input original vibration sample to obtain the original prior feature set. The time-domain features include mean, standard deviation, root square amplitude, absolute mean, skewness, kurtosis, variance, maximum value, minimum value, and peak-to-peak value, totaling 10 time-domain prior features. Thirteen frequency-domain prior features are extracted through Fast Fourier Transform (FFT) to reflect changes in vibration energy, the concentration and dispersion of the spectrum, and the position of the main frequency band, as shown in Table 2. Here, y(k) is the Fourier spectrum of the time series x(n), k = 1, 2, 3...K, where K is the number of spectral lines in the Fourier spectrum, and f... k This is the frequency value corresponding to the k-th spectral line. FF1 reflects the change in vibrational energy in the frequency domain, FF2~FF4, FF6 and FF... 10 ~FF 13 Describes the degree of concentration and dispersion of the spectrum; FF5 and FF7-FF9 represent the changes in the position of the main frequency band.

[0043] Table 2 shows the characteristic indicators in the frequency domain.

[0044]

[0045] In the two-stage feature index screening module, Fisher scores are first calculated for each feature based on the original prior feature set. Features are then preliminarily screened based on Fisher scores to obtain features with high fault sensitivity and good discriminative power. Next, Pearson correlation coefficients are calculated for each feature. Pearson correlation coefficients are used to further reduce the redundancy between features. Based on the Pearson correlation coefficients, the features after the preliminarily screened feature set are sorted in descending order. The feature with the highest discriminative power is retained as the benchmark feature. The correlation between the features after the preliminarily screened feature set and the benchmark feature is then calculated. Features with a correlation less than the correlation threshold are retained to obtain the features after the second screening and are recorded as the prior feature subset after feature screening.

[0046] Fisher scores include between-class divergence. Intra-class divergence And Fisher's fractionsk The calculation formula is as follows:

[0047]

[0048]

[0049]

[0050] Where C is the number of fault categories, and N i It is the number of fault samples of type i. For the k-th feature of the j-th sample in the i-th type of fault samples, Let u be the mean of all samples in the i-th type of fault sample for the k-th feature. k Let be the mean of the k-th feature for all samples across all classes, where k = 1, 2, ..., K, and K is the number of prior feature indices. The inter-class divergence and intra-class divergence of each feature are summed; the ratio of these two values ​​reflects the average Fisher score, which is used as the threshold η for selecting sensitive features. fisher The calculation formula is:

[0051]

[0052] Based on the dataset information in this embodiment, the threshold η of the Fisher score is calculated. fisher It is 4.28.

[0053] Pearson correlation coefficient is used to further reduce redundancy between features. The Pearson correlation coefficient r between feature indices a and b is... a,b Calculate according to the following formula:

[0054]

[0055] In the formula, N is the total number of samples, 1≤p,q≤C, 1≤s≤N p , 1≤t≤N q r a,b The larger the value of r, the stronger the correlation between the two types of features. If r is greater than the set threshold δ, pearson If the threshold δ is not met, then one of the two features will be removed. Set a threshold δ for the Pearson correlation coefficient. pearson The value is 0.9. The features are sorted from high to low in terms of discriminative power. The feature with the highest discriminative power is retained. Then, the correlation between this feature and other features is determined to be less than the correlation threshold. Features with a correlation threshold are selected to form a feature subset.

[0056] S3: After training the feature-excited fault sample amplification module using the original vibration sample dataset and the prior feature subset after feature filtering, a well-trained feature-excited fault sample amplification module is obtained, thus obtaining a well-trained generator.

[0057] The feature-driven fault sample augmentation module includes a generator G, a feature pre-extractor F, and a discriminator. A subset of prior features after feature filtering serves as input to generator G. Generator G is connected to the discriminator after passing through feature pre-extractor F. The input to generator G also includes random noise and labels. The original vibration sample dataset serves as input to feature pre-extractor F. Prior features are embedded into the generator of the fault sample augmentation module, and the deep features extracted by the feature pre-extractor are used to enhance the discriminator's ability, thereby co-stimulating the fault sample generation network.

[0058] like Figure 3 As shown, the generator G learns the distribution of real fault data, and the generator processes random noise z. noise Multiple upsampling operations are performed until the output feature map matches the size of the real sample. This includes an initialization module, an upsampling module, and a skip-layer feature activation module (SKEM). The initialization and upsampling modules are the same as those in traditional GAN ​​network model generators. Compared to traditional skip-layer methods, the skip-layer feature activation module (SKEM) can perform skip connections across a wider range of resolutions, reducing computational burden and accelerating gradient shrinkage.

[0059] Random noise and labels are input together into the initialization module. The prior feature subsets after feature filtering are all used as inputs to the skip-layer feature activation module (SKEM). Each initialization module is connected to the first upsampling module and the first skip-layer feature activation module (SKEM), respectively. The first upsampling module, after passing through the second and third upsampling modules, is connected to the fourth upsampling module. The fourth upsampling module is connected to the first skip-layer feature activation module (SKEM), which is then connected to the fifth upsampling module. Both the first and fifth upsampling modules are connected to the second skip-layer feature activation module (SKEM), which is connected to the sixth upsampling module. Both the second and sixth upsampling modules are connected to the third skip-layer feature activation module (SKEM). M is connected, the third jump-layer feature excitation module SKEM is connected to the seventh upsampling module, both the third and seventh upsampling modules are connected to the fourth jump-layer feature excitation module SKEM, the fourth jump-layer feature excitation module SKEM is connected to the tenth upsampling module after passing through the eighth and ninth upsampling modules. The outputs of the initialization module, the first upsampling module, the second upsampling module, and the third upsampling module are all used as the second input feature map b of the corresponding jump-layer feature excitation module SKEM, and the outputs of the fourth, fifth, sixth, and seventh upsampling modules are all used as the first input feature map a of the corresponding jump-layer feature excitation module SKEM. The output of the tenth upsampling module is denoted as the amplified vibration sample, which is the input of the feature pre-extractor F.

[0060] The Skip-Layer Feature Activation Module (SKEM) includes a first adaptive average pooling layer, a convolutional layer, an activation layer, and a linear layer. The second input feature map b of the Skip-Layer Feature Activation Module (SKEM) is connected to the second convolutional layer after passing through the first adaptive average pooling layer, the first convolutional layer, and the first activation layer. The prior feature subset after feature filtering is connected to the second linear layer after passing through the first linear layer and the second activation layer. The output of the second convolutional layer is added to the output of the second linear layer and then input to the third activation layer. The output of the Skip-Layer Feature Activation Module (SKEM) is the product of the first input feature map a of the Skip-Layer Feature Activation Module (SKEM) and the output of the third activation layer.

[0061] In the initialization module, noise and label vectors are concatenated and used as input. Transposed convolution, batch normalization, and SiLU activation layer operations are then performed sequentially, respectively increasing the network size, accelerating network training, and enhancing the network's nonlinear fitting ability. The upsampling module consists of nearest neighbor interpolation, convolutional layers, classification conditional batch normalization layers, and SiLU activation layers. In the skip-layer feature activation module, the prior feature subset after feature selection is embedded into the generator network through linear layers, SiLU activation layers, and another linear layer. Simultaneously, the network feature map is processed through a first adaptive average pooling layer, convolutional layers, SiLU activation layers, and another convolutional layer, then added to the processed prior features. After passing through a Sigmoid activation layer, it is multiplied element-wise with the original input feature map, achieving the fusion of cross-layer feature maps. The generator's output module consists of a single-layer convolutional network. The generator G receives random noise z. noise Label c and prior features f prior The generator produces fake faulty samples. The generator aims to produce samples that are as realistic as possible, and its optimization objective is:

[0062]

[0063] in, Let z follow the probability distribution P z (z) is the mathematical expectation, where D() is the discriminator model function, F() is the feature pre-extractor model function, and G() is the generator model function.

[0064] like Figure 4 As shown, in the feature-driven fault sample amplification module, there are four feature pre-extractors F and four discriminators. Each feature pre-extractor F consists of a feature extraction network f. θ and feature mixing module f mix The input of the feature-stimulated fault sample amplification module is used as the input of the first feature pre-extractor F, which is connected to form a feature extraction network f. θ The output is used as the feature extraction network f of the second graph feature pre-extractor F. θ The input is the second feature pre-extractor F, and the feature extraction network f.θ The output is used as the feature extraction network f of the third graph feature pre-extractor F. θ The input is the feature extraction network f of the third feature pre-extractor F. θ The output is used as the feature extraction network f of the fourth graph feature pre-extractor F. θ The input, the feature mixing module f of the fourth feature pre-extractor F mix The output of the fourth discriminator is used as the input of the feature mixing module f of the fourth feature pre-extractor F. mix No input feature map. The feature mixing module f of the fourth feature pre-extractor F. mix The output is also used as the feature mixing module f of the third feature pre-extractor F. mix The input, the feature mixing module f of the third feature pre-extractor F mix The output of the third discriminator is used as the input of the feature mixing module f of the third feature pre-extractor F. mix The output is also used as the feature mixing module f of the second feature pre-extractor F. mix The input, the feature mixing module f of the second feature pre-extractor F mix The output of the second discriminator is used as the input of the feature mixing module f of the second feature pre-extractor F. mix The output is also used as the feature mixing module f of the first feature pre-extractor F. mix The input is the feature mixing module f of the first feature pre-extractor F. mix The output of the first discriminator is used as the input of the second discriminator. The average of the outputs of the four discriminators is then used as the final output of the discriminator.

[0065] The optimization objective of the discriminator is:

[0066]

[0067] in, This indicates that x follows a probability distribution P. data The expected value under condition (x), where x represents the input matrix of the discriminator.

[0068] Feature extraction network f θ It is the feature extraction network f in the online network of the deep feature pre-learning module based on contrastive learning after pre-training. θ During the training of S3, the feature extraction network f θ The weights remain fixed. Specifically: the deep feature pre-learning module based on contrastive learning consists of the online network z. online and target network z target Composition. The online network consists of a feature extraction network f. θ Projection network g θ The target network consists of three neural networks: the prediction network q, the feature extraction network f, and the target network q.ξ and projection network g ξ Composition. Feature extraction network f of online network. θ and the feature extraction network f of the target network ξ Both adopt the ResNet18 network architecture, and the projection network g of the online network. θ The projection network g of the target network ξ Both the prediction network q and the prediction network use fully connected networks.

[0069] For each fault sample, two data augmentation methods are randomly selected from five methods—sequence flipping, sequence replacement, sequence scaling, fragment splicing, and sequence noise addition—to construct a pair of contrastive learning positive samples (v and v′). First, feature representations are obtained through a feature extraction network, and then these representations are projected onto the network. In each training step, the target network parameters ξ are updated according to ξ←τξ+(1-τ)θ, where τ is the decay rate and θ is the online network parameter. Finally, the online network outputs the vector z. online The prediction network q outputs the target network vector z. target Make a prediction. Calculate the loss using negative cosine similarity:

[0070] z online (v)=g θ (f θ (v))

[0071] z target (v')=g ξ (f ξ (v'))

[0072]

[0073] Where ||·||2 represents the L2 norm. The loss obtained by inputting the first positive sample v into the online network module and the second positive sample v′ into the target network module is: Symmetrically, the loss obtained by inputting the second positive sample v′ into the online network module and the first positive sample v into the target network module is: Comparative learning loss The sum of the two,

[0074] The Adam W optimizer is used in the training process of the deep feature pre-learning module based on contrastive learning. The hyperparameter values ​​during training are shown in Table 3. The decay rate τ is updated in each training step according to the following formula, where τ base Let m be the base decay rate, m be the current training step, and M be the maximum training step.

[0075]

[0076] Table 3 shows the training hyperparameters of the feature pre-extraction module based on contrastive learning.

[0077]

[0078] like Figure 4 As shown, the feature mixing module f mix It consists of cross-channel blending (CCM) and cross-scale blending (CSM). Cross-channel blending interacts and blends the various channels of the feature map; cross-scale blending is calculated in a top-down order, upsampling the smaller feature map above and adding it to the feature map to complete the cross-scale blending. The third feature map is used as the input to the third convolutional layer. The output of the third convolutional layer is added to the fourth feature map after linear interpolation and then input into the fourth convolutional layer. The output of the fourth convolutional layer is used as the final output.

[0079] Each discriminator consists of N downsampling modules, a global summing pooling layer, an embedding layer, an adaptive average pooling layer, and a third linear layer. The N downsampling modules are cascaded sequentially. The feature map input to the discriminator is used as the input to the first downsampling module, and the label input to the discriminator is used as the input to the embedding layer. The last downsampling module is connected to the third linear layer after passing through the global summing pooling layer. The output of the global summing pooling layer is multiplied element-wise with the output of the embedding layer and then input into the adaptive average pooling layer. The output of the third linear layer is added to the output of the adaptive average pooling layer to output a sub-discriminator vector, which is used as the output of the discriminator.

[0080] The AdamW optimizer used in the training process of S3 has the following training parameter settings as shown in Table 4.

[0081] Before training begins, f in the deep feature pre-learning module θ The weight parameters are initialized using the pre-trained contrastive learning-based deep feature pre-learning module, and then the feature mixing network f is initialized. mix Kaiming initialization is performed, followed by freezing of the network parameters of the feature pre-extraction module based on contrastive learning. When training the feature-driven fault sample augmentation network module, the feature pre-extraction module does not participate in optimization; it only learns the network parameters of the generator and discriminator. The optimization objectives of the generator and discriminator are combined to obtain the training joint optimization objective:

[0082]

[0083] Table 4 shows the hyperparameters of the feature-driven fault sample amplification network module.

[0084]

[0085] S4: Input the original vibration samples for each fault type into the trained generator to obtain amplified vibration samples for that fault type, thus completing the fault sample amplification task. The final vibration sample dataset consists of the original vibration samples and the corresponding amplified vibration samples.

[0086] S5: After training the fault diagnosis model using the final vibration sample dataset, a trained fault diagnosis model is obtained; the vibration sample to be detected (i.e., the test dataset) is input into the trained fault diagnosis model, and the corresponding fault diagnosis result is output.

[0087] In specific implementation, the fault diagnosis model is based on a multi-scale convolutional residual network, which consists of an input layer, a fifth convolutional layer, a first batch of normalization layers, a SELU nonlinear activation function layer, n layers of multi-scale convolutional residual blocks with different kernel sizes, a flattening layer, and an output layer connected in sequence. The multi-scale convolutional residual block includes a first channel-wise convolutional layer, a second batch of normalization layers, a first SELU nonlinear activation function layer, a second channel-wise convolutional layer, a third batch of normalization layers, and a second SELU nonlinear activation function layer.

[0088] The input of the multi-scale convolutional residual block is used as the input of the first channel-wise convolutional layer. The first channel-wise convolutional layer is then connected to the third batch of normalization layers after passing through the second batch of normalization layers, the first SELU nonlinear activation function layer, and the second channel-wise convolutional layer. The output of the third batch of normalization layers is added to the input of the multi-scale convolutional residual block and then input into the second SELU nonlinear activation function layer. The output of the second SELU nonlinear activation function layer is used as the output of the multi-scale convolutional residual block, thereby enabling the residual block to effectively mine the deep features of the part fault signal.

[0089] The optimizer used in the training process of the multi-scale convolutional residual network is the Adam optimizer. In this embodiment, the learning rate is set to 5×10. -5 The training loss function uses the cross-entropy function.

[0090] In this embodiment, a feature-driven fault sample augmentation network is used to generate original fault samples, expanding the number of fault samples. The relative Pearson correlation coefficient is used as the evaluation index for the fault sample generation effect. The larger the relative Pearson correlation coefficient, the higher the similarity between the generated signal and the real signal, and the better the fault sample generation effect. In this embodiment, the relative Pearson correlation coefficients of the three types of fault samples are shown in Table 5:

[0091] Table 5 shows the relative Pearson correlation coefficients of the generated fault samples.

[0092]

[0093] The trained model was tested using 20% ​​of the original dataset as the test set. For the three types of part failures, three experiments were conducted using both unexpanded and expanded samples on the same device and with the same classifier model. The computational accuracy of the classification task on the test set was recorded, and the mean and standard deviation of the above indicators were calculated as the final evaluation parameters. In this embodiment, the average and standard deviation of the classification accuracy for expanded and unexpanded samples are shown in Table 6.

[0094] Table 6 shows the fault diagnosis results.

[0095]

[0096]

[0097] As shown in Table 6, the classification accuracy of the expanded samples is significantly improved compared to the unexpanded samples in the classification experiment. Furthermore, the standard deviation of the experimental results is less than 0.015, indicating minimal fluctuation and reliable results. These experimental results demonstrate that this method is feasible and effective for diagnosing part faults using scarce samples.

[0098] This invention addresses the poor generalization performance of data-driven component fault diagnosis models due to the scarcity of fault samples. It constructs a multi-domain feature set and uses feature sensitivity and redundancy evaluation criteria to screen feature indicators, removing redundant features to improve the quality of diagnostic samples. A feature-incentivized fault sample generative adversarial network (GAN) is built, incorporating prior features into the generator to guide fault sample generation. A pre-trained network is used to construct a feature pre-extraction module based on contrastive learning. The extracted pre-learned features with clearer class boundary attributes are input into the discriminator. Compared to directly inputting fault samples into the discriminator, the use of multi-layer pre-learned features reduces the discriminator's learning burden, resulting in faster training speeds compared to traditional GANs. A skip-layer feature excitation module is designed in the generator, enabling the reuse of features from different levels through the fusion of cross-layer feature maps. Embedding prior features and utilizing the expert prior knowledge they contain provides appropriate excitation during the network's sample generation process, better guiding the generation of fault samples. During training, both the model generator and the discriminator receive sample label information. After training, the generator generates fault samples with specified fault categories, effectively and specifically solving the problem of scarce fault samples and improving the accuracy and generalization ability of downstream classification tasks.

[0099] The above embodiments represent an optimization result of the bearing dataset used in this embodiment of the present invention, but the specific implementation of the present invention is not limited to the above embodiments. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention; at the same time, for those skilled in the art, there will be changes in specific implementation methods and application scope based on the ideas of the present invention. In summary, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for part fault diagnosis based on feature excitation type sample expansion under data scarcity, characterized in that, Includes the following steps: S1: Collect vibration monitoring signals of key components of electromechanical products, and then construct the original vibration sample dataset; S2: After extracting and filtering features from the original vibration sample dataset using the prior feature extraction module, a subset of prior features after feature filtering is obtained; In S2, the prior feature extraction module is composed of a feature extraction module and a two-stage feature index screening module connected together. The feature extraction module extracts multiple time-domain features and frequency-domain features of the input original vibration sample to obtain the original prior feature set. In the two-stage feature index screening module, firstly, Fisher scores are calculated among the features based on the original prior feature set. Based on the Fisher scores, the features are initially screened to obtain features with high fault sensitivity and good discriminative power. Next, Pearson correlation coefficients are calculated among the features. Based on the Pearson correlation coefficients, the features after initial screening are sorted in descending order, and the feature with the highest discriminative power is retained as the benchmark feature. Then, the correlation between the features after initial screening and the benchmark feature is calculated. Features with a correlation less than the correlation threshold are retained to obtain the features after secondary screening and are recorded as the prior feature subset after feature screening. S3: After training the feature-excited fault sample amplification module using the original vibration sample dataset and the prior feature subset after feature filtering, a well-trained feature-excited fault sample amplification module is obtained, thus obtaining a well-trained generator. In S3, the feature-excited fault sample amplification module includes a generator G, a feature pre-extractor F, and a discriminator; the prior feature subset after feature screening is used as the input of the generator G, and the generator G is connected to the discriminator after passing through the feature pre-extractor F. The input of the generator G also includes random noise and labels, and the original vibration sample dataset is used as the input of the feature pre-extractor F. The generator G includes an initialization module, an upsampling module, and a skip-level feature activation module (SKEM). Random noise and labels are input into the initialization module. Prior feature subsets after feature filtering are used as inputs to the skip-level feature activation module (SKEM). Each initialization module is connected to the first upsampling module and the first skip-level feature activation module (SKEM). The first upsampling module is connected to the fourth upsampling module via the second and third upsampling modules. The fourth upsampling module is connected to the first skip-level feature activation module (SKEM). The first skip-level feature activation module (SKEM) is connected to the fifth upsampling module. Both the first and fifth upsampling modules are connected to the second skip-level feature activation module (SKEM). The second skip-level feature activation module (SKEM) is connected to the sixth upsampling module. Both the block and the sixth upsampling module are connected to the third jump-layer feature excitation module SKEM. The third jump-layer feature excitation module SKEM is connected to the seventh upsampling module. Both the third and seventh upsampling modules are connected to the fourth jump-layer feature excitation module SKEM. The fourth jump-layer feature excitation module SKEM is connected to the tenth upsampling module after passing through the eighth and ninth upsampling modules. The outputs of the initialization module, the first upsampling module, the second upsampling module, and the third upsampling module are all used as the second input feature map of the corresponding jump-layer feature excitation module SKEM. The outputs of the fourth, fifth, sixth, and seventh upsampling modules are all used as the first input feature map of the corresponding jump-layer feature excitation module SKEM. The output of the tenth upsampling module is denoted as the amplified vibration sample. S4: Input the original vibration samples of each fault type into the trained generator to obtain the amplified vibration samples under the fault type. The original vibration samples and the corresponding amplified vibration samples constitute the final vibration sample dataset. S5: After training the fault diagnosis model using the final vibration sample dataset, a trained fault diagnosis model is obtained; the vibration sample to be detected is input into the trained fault diagnosis model, and the corresponding fault diagnosis result is output.

2. The method for part fault diagnosis based on feature-excited sample amplification under data scarcity as described in claim 1, characterized in that, The vibration monitoring signals of the key components of the electromechanical products include vibration signals in both the horizontal and vertical directions monitored by the accelerometer.

3. The method for part fault diagnosis based on feature-excited sample amplification under data scarcity as described in claim 1, characterized in that, The time-domain features include mean, standard deviation, root mean, absolute mean, skewness, kurtosis, variance, maximum value, minimum value, and peak-to-peak value.

4. The method for part fault diagnosis based on feature-excited sample amplification under data scarcity as described in claim 1, characterized in that, The Skip-Layer Feature Activation Module (SKEM) includes a first adaptive average pooling layer, a convolutional layer, an activation layer, and a linear layer. The second input feature map of the Skip-Layer Feature Activation Module (SKEM) is connected to the second convolutional layer after passing through the first adaptive average pooling layer, the first convolutional layer, and the first activation layer. The prior feature subset after feature filtering is connected to the second linear layer after passing through the first linear layer and the second activation layer. The output of the second convolutional layer is added to the output of the second linear layer and then input to the third activation layer. The output of the Skip-Layer Feature Activation Module (SKEM) is the product of the first input feature map of the Skip-Layer Feature Activation Module (SKEM) and the output of the third activation layer.

5. The method for part fault diagnosis based on feature-excited sample amplification under data scarcity as described in claim 1, characterized in that, In the feature-driven fault sample amplification module, there are four feature pre-extractors F and four discriminators. Each feature pre-extractor F is composed of a feature extraction network. and feature blending module The input of the feature-stimulated fault sample amplification module is used as the input of the first feature pre-extractor F, which is a connected feature extraction network. The output is used as the feature extraction network of the second graph feature pre-extractor F. The input, the feature extraction network of the second feature pre-extractor F The output is used as the feature extraction network of the third graph feature pre-extractor F. The input, the feature extraction network of the third feature pre-extractor F The output is used as the feature extraction network of the fourth graph feature pre-extractor F. The input, the feature mixing module of the fourth feature pre-extractor F The output of the fourth discriminator is used as the input of the feature mixing module of the fourth feature pre-extractor F. The output is also used as the feature mixing module of the third feature pre-extractor F. The input, the feature mixing module of the third feature pre-extractor F The output of the third discriminator is used as the input of the feature mixing module of the third feature pre-extractor F. The output is also used as the feature mixing module of the second feature pre-extractor F. The input, the feature mixing module of the second feature pre-extractor F The output of the second discriminator is used as the input of the feature mixing module of the second feature pre-extractor F. The output is also used as the feature mixing module of the first feature pre-extractor F. The input to the feature mixing module of the first feature pre-extractor F. The output of the first discriminator is used as the input of the second discriminator. The average of the outputs of the four discriminators is then used as the final output of the discriminator.

6. The method for part fault diagnosis based on feature-excited sample amplification under data scarcity as described in claim 5, characterized in that, The feature extraction network It is the feature extraction network in the online network of the deep feature pre-learning module based on contrastive learning after pre-training. During the training of S3, the feature extraction network The weights remain constant.

7. The method for part fault diagnosis based on feature-excited sample amplification under data scarcity as described in claim 5, characterized in that, Each discriminator includes N downsampling modules, a global summing pooling layer, an embedding layer, an adaptive average pooling layer, and a third linear layer. The N downsampling modules are cascaded sequentially. The feature map of the input discriminator is used as the input of the first downsampling module, and the label of the input discriminator is used as the input of the embedding layer. The last downsampling module is connected to the third linear layer after passing through the global summing pooling layer. The output of the global summing pooling layer is multiplied element-wise with the output of the embedding layer and then input into the adaptive average pooling layer. The output of the third linear layer is added to the output of the adaptive average pooling layer to output a sub-discriminator vector, which is used as the output of the discriminator.

Images