Single-source-domain generalization fault diagnosis method based on deep learning

Abstract

The invention discloses a single-source-domain generalization fault diagnosis method based on deep learning, and belongs to the field of mechanical fault diagnosis. The method comprises the following steps: extracting features through an encoder to obtain class-related features and class-independent features; enhancing the class-related features and the class-independent features through a feature-level blending enhancement module; processing the enhancement class related features and the enhancement class unrelated features through a decoder; performing local clustering on samples in the training set; and training the diagnosis model, and inputting the test set data into the trained diagnosis model to realize fault diagnosis. According to the method, class-related features and class-independent features are extracted through a gradient orthogonality and redundancy suppression constraint encoder, appropriate enhanced data is selected to expand a data set after feature enhancement and similarity screening, and finally distinguishing between classes is enhanced through local clustering, and a model is trained to improve the diagnosis accuracy.

Description

Technical Field

[0001] The present invention relates to mechanical fault diagnosis technology, and in particular to a single-source domain generalized fault diagnosis method based on deep learning. Background Art

[0002] With the surge in industrial monitoring data and the rapid advancement of artificial intelligence (AI), research on data-centric fault diagnosis has experienced explosive growth. Deep learning has demonstrated remarkable success in fault diagnosis, automatically extracting deep feature information and rapidly responding to machine health conditions. Thanks to deep learning's ability to automatically extract end-to-end features, intelligent fault diagnosis solutions have evolved from traditional signal processing and model-based approaches to modern deep learning-based methods.

[0003] In real-world scenarios, the data available for model training often comes from different operating conditions of the same machine or different machines of the same type, i.e., cross-operating condition and cross-machine scenarios. Due to the diverse operating conditions and physical characteristics of machines, cross-domain fault diagnosis is necessary. Multi-source data contain similar fault signatures because they come from machines with the same operating principles and failure mechanisms. By leveraging knowledge from related datasets, highly accurate and robust diagnostic models can be achieved. The most important issue in diagnostic model construction is to acquire domain-invariant knowledge while eliminating domain-specific information. Therefore, domain generalization-based fault diagnosis (DGFD) has been widely studied. DGFD learning methods train deep models on multiple source domain datasets and then directly apply them to the target diagnostic task.

[0004] Despite significant progress in deep learning-based domain generalization fault diagnosis for vibration signals, it still faces a series of challenges. Most diagnostic models based on domain generalization are constructed based on information from multiple domains. Their excellent performance comes at the expense of massive amounts of training data. In real-world industrial scenarios, directly adopting these multi-domain approaches may be impractical given the high cost and safety risks of data collection. Summary of the Invention

[0005] Purpose of the invention: In response to the above problems, the purpose of the present invention is to provide a single-source domain generalized fault diagnosis method based on deep learning. By introducing orthogonal operations and redundancy reduction in the encoder, extracting class-related features and class-irrelevant features, selecting appropriate enhanced data to expand the data set, and finally enhancing the distinction between categories through local clustering, the model is trained to improve the diagnostic accuracy.

[0006] Technical solution: The single-source domain generalized fault diagnosis method based on deep learning of the present invention includes the following steps: Step 1: Collect a single-source domain ball bearing vibration signal from the bearing seat closest to the ball bearing, divide it into vibration signal sample segments, and then preprocess it. The preprocessed vibration signal sample segments are divided into a training set containing single-source domain data and a test set containing multi-source domain data; Step 2: Extract features from the data in the training set through the encoder and constrain them to obtain class-related features and class-independent features; Step 3: Enhance the class-related features and class-irrelevant features through a feature-level co-mixing enhancement module to generate enhanced class-related features and enhanced class-irrelevant features; Step 4: Process the enhancement-related features and the enhancement-irrelevant features through a decoder to generate an enhanced signal; Step 5: Calculate the cosine similarity between the enhanced signal and the original ball bearing vibration signal, and select the enhanced signal based on the cosine similarity; Step 6: Merge the original ball bearing vibration signal and the filtered enhanced signal to construct a joint data space, and perform local clustering on the samples in the joint data space; In step 7, the clustered joint data is input into the classifier to obtain the fault type, and the total loss value is calculated. The parameters of the diagnostic model are updated in reverse to obtain a trained diagnostic model. The test set data is input into the trained diagnostic model to realize the diagnosis of the fault. The diagnostic model includes an encoder and a classifier.

[0007] Furthermore, step 2 includes: The vibration signal sample input to the encoder is recorded as , class-related features are obtained through the encoder Class-independent features , and then constrain class-related features through feature decoupling loss Class-independent features ; The feature decoupling loss includes orthogonal loss and redundancy suppression loss , the formulas are: , , Where, and They are class-related loss and class-independent loss relative to class-related features. Class-independent features The gradient, represents the L2 norm, and They are class-related and class-independent losses respectively; is a square matrix with element 1, is the identity matrix, used to mask the diagonal elements, is a class-related feature matrix, each row represents a class-related feature vector of a sample, is a class-independent matrix; represents the Hadamard product, represents the Frobenius norm, is the characteristic dimension; The feature decoupling loss expression is: .

[0008] Furthermore, step 3 includes: Class-related features Class-independent features Input into the feature-level synergistic enhancement module for synergistic data enhancement to generate class-related enhanced features Class-independent enhancement features , the formulas are: , , Where, represents the enhancement coefficient, Indicates that the random index The class-related features after arrangement, Indicates that the random index Class-independent features after permutation.

[0009] Furthermore, step 4 includes: Enhance the class-related features Class-independent enhancement features The input is sent to the decoder, and the two feature vectors are merged into a joint feature along the feature dimension. Then the dimension is gradually reduced through the linear layer, convolutional layer and ReLU, and finally becomes the enhanced signal z mapped to the original signal dimension.

[0010] Furthermore, step 5 includes: Perform fast Fourier transform on the enhanced signal z and the original mechanical vibration signal to obtain an analytical signal, set the negative frequency part of the obtained analytical signal to zero, and retain the positive frequency part; Then perform inverse fast Fourier transform on the spectrum of the analytical signal to obtain the complex form of the analytical signal , according to the properties of Hilbert transform, the analytical signal Expressed as: , Where, is the original ball bearing vibration signal, is an imaginary unit, is the Hilbert transform result of the original ball bearing vibration signal; Calculate the envelope value of the original ball bearing vibration signal using the formula: , Calculate the cosine similarity based on the envelope value. The formula is: , Where, and represent the envelope value of the original ball bearing vibration signal and the envelope value of the enhanced signal, respectively. is the dot product of two vectors, and yes and Length of the module; The enhanced signals z whose cosine similarity exceeds the threshold are selected and added to the training set.

[0011] Furthermore, step 6 includes: The original data and the filtered enhanced data are merged and the K-negative triplet loss is used to Enhance the distinction between categories, select the k negative samples farthest from the anchor point, and maximize the anchor point and negative samples The distance between the anchor points of the same type of samples is minimized and positive samples The distance between them is used to achieve clustering; Among them, K negative triple loss The formula is as follows: , Where, is the standard relaxation coefficient; the distance between samples is defined as follows: , Where, and It is a sample and Class-related features.

[0012] Furthermore, the formula for calculating the total loss value in step 7 is: , Where, is the weight parameter, is the feature decoupling loss, Triple loss for K; is the cross entropy loss, and the formula is: , Where, is the total number of samples, and are the true label and predicted probability of the i-th class of sample, respectively.

[0013] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: 1. The present invention uses an encoder with gradient orthogonality and redundancy suppression constraints to ensure that the different features extracted by the encoder are independent of each other at the gradient level, avoiding information interference between features and allowing the encoder to focus on core features that are valuable for category judgment; 2. The present invention combines the feature-level synergistic enhancement module to perform data expansion in the feature space. By comparing the fit between the enhanced data and the original data in terms of feature distribution and category consistency, the best enhanced data is selected and added to the training set, thus solving the problem of insufficient training data. 3. The present invention uses local clustering to make similar features more closely aggregated in the feature space, and the boundaries between different types of features are clearer, further enhancing the model's ability to identify different types of faults. Experimental results show that the fault diagnosis method described in the present invention can effectively improve the accuracy of signal diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a flowchart of the present invention; Figure 2 Comparison chart of the average accuracy of unit domain generalization of different methods. DETAILED DESCRIPTION

[0015] In order to make the purpose, technical solutions and advantages of this application more clear, this application is further described in detail below with reference to the accompanying drawings and embodiments.

[0016] Combine Figure 1 As shown, the single-source domain generalized fault diagnosis method based on deep learning described in this embodiment includes the following steps: Step 1: Collect the single-source domain ball bearing vibration signal from the bearing seat closest to the ball bearing, divide it into vibration signal sample segments, and then preprocess it. The preprocessed vibration signal sample segments are divided into a training set containing single-source domain data and a test set containing multi-source domain data.

[0017] In one example, raw data is obtained from multiple source domains. Specifically, in each source domain, a single-source domain ball bearing vibration signal is collected from the bearing seat closest to the ball bearing using an acceleration sensor and a rolling bearing test module. This signal is used as the original vibration signal, and the original vibration signal dataset is divided into a training set and a test set by slicing. The training set data is divided into 800 samples per category, and the test set data is divided into 200 samples per category, with each sample containing 3200 data points. All samples are normalized, and the single-source domain training data is randomly sampled three times. Accurate domain labels are assigned to them so that they correspond to domain labels 0, 1, and 2, respectively, to generate a data set with clear domain identification. These carefully processed data sets with different domain labels are used as the input source of the encoder to achieve effective training of the encoder, thereby laying a solid data foundation for subsequent model optimization and task execution.

[0018] Step 2: Extract features from the data in the training set through the encoder and constrain them to obtain class-related features and class-independent features.

[0019] Furthermore, step 2 includes: The vibration signal sample input to the encoder is recorded as , class-related features are obtained through the encoder Class-independent features , and then constrain class-related features through feature decoupling loss Class-independent features ; The feature decoupling loss includes orthogonal loss and redundancy suppression loss , the formulas are: , , Where, and They are class-related loss and class-independent loss relative to class-related features. Class-independent features The gradient, represents the L2 norm, and They are class-related and class-independent losses respectively; is a square matrix with element 1, is the identity matrix, used to mask the diagonal elements, is a class-related feature matrix, each row represents a class-related feature vector of a sample, is a class-independent matrix; represents the Hadamard product, represents the Frobenius norm, is the characteristic dimension; The feature decoupling loss expression is: .

[0020] In this example, the encoder consists of two convolutional neural network (CNN) sub-networks and four fully connected layers. First, convolution is performed through the convolution kernel, and then batch normalization, ReLU activation function and maximum pooling are used to gradually extract and concentrate features. After the four layers of neural networks, the dimension of the feature space is reduced to obtain class-related features. Class-independent features By performing orthogonal operations and redundant suppression constraints on the two types of features extracted, the gradient directions of the two are made orthogonal to each other and the correlation between class-related features and class-irrelevant features and feature dimensions is reduced. Orthogonal loss By minimizing the class-dependent loss Class-independent loss The dot product of the gradient is used to achieve the class-related loss. Class-independent loss The formula is as follows: , , in, is the cross entropy loss function, and They are category branch classifier and class-independent branch classifier, and The input signals are The category labels and domain labels of .

[0021] Step 3: The class-related features and class-irrelevant features are enhanced by a feature-level co-mixing enhancement module to generate enhanced class-related features and enhanced class-irrelevant features.

[0022] Furthermore, step 3 includes: Class-related features Class-independent features Input into the feature-level synergistic enhancement module for synergistic data enhancement to generate class-related enhanced features Class-independent enhancement features , the formulas are: , , Where, represents the enhancement coefficient, sampled from the Beta distribution, , is the hyperparameter of feature-level cointegration data augmentation, Indicates that the random index The class-related features after arrangement, Indicates that the random index Class-independent features after permutation.

[0023] The same type and Through the feature-level co-mixing data enhancement module, new class-related and class-independent features are created, and the category labels of the new features are retained and given new domain labels.

[0024] Step 4: Process the enhancement-related features and the enhancement-irrelevant features through a decoder to generate an enhanced signal.

[0025] Furthermore, step 4 includes: Enhance the class-related features Class-independent enhancement features The input is sent to the decoder, and the two feature vectors are merged into a joint feature along the feature dimension. Then the dimension is gradually reduced through the linear layer, convolutional layer and ReLU, and finally becomes the enhanced signal z mapped to the original signal dimension.

[0026] In this example, the decoder is used to transform the dual feature vectors, i.e., the class-related enhanced features Class-independent enhancement features Decoding the original signal and restoring the original signal dimension from the abstract features. This process takes two 256-dimensional feature vectors, merges them along the feature dimension into a 512-dimensional joint feature, and then gradually reduces the dimension through linear layers, convolutional layers, and ReLU layers, ultimately becoming enhanced data mapped to the original signal dimension.

[0027] Step 5: Calculate the cosine similarity between the enhanced signal and the original ball bearing vibration signal, and select the enhanced signal based on the cosine similarity.

[0028] Furthermore, step 5 includes: Perform fast Fourier transform on the enhanced signal z and the original mechanical vibration signal to obtain an analytical signal, set the negative frequency part of the obtained analytical signal to zero, and retain the positive frequency part; Then perform inverse fast Fourier transform on the spectrum of the analytical signal to obtain the complex form of the analytical signal , according to the properties of Hilbert transform, the analytical signal Expressed as: , Where, is the original ball bearing vibration signal, is an imaginary unit, is the Hilbert transform result of the original ball bearing vibration signal; Calculate the envelope value of the original ball bearing vibration signal using the formula: , Calculate the cosine similarity based on the envelope value. The formula is: , Where, and represent the envelope value of the original ball bearing vibration signal and the envelope value of the enhanced signal, respectively. is the dot product of two vectors, and yes and Length of the module; The enhanced signals z whose cosine similarity exceeds the threshold are selected and added to the training set.

[0029] In one example, enhanced signals z with a cosine similarity greater than 85% are selected and added to the training set.

[0030] In step 6, the original ball bearing vibration signal and the filtered enhanced signal are combined to construct a joint data space, and local sample clustering is performed to achieve preliminary sample classification. The k-negative triple loss constraint is used to give the samples in the training set a rough distribution, such as samples of the same type are close together and samples of different types are far apart, so that the final classification can be more accurate.

[0031] Furthermore, step 6 includes: The original data and the filtered enhanced data are merged into a joint data space, and K negative triplet loss is used to Enhance the distinction between categories, select the k negative samples farthest from the anchor point, and maximize the anchor point and negative samples The distance between the anchor points of the same type of samples is minimized and positive samples The distance between them is used to achieve clustering; Among them, K negative triple loss The formula is as follows: , Where, is the standard relaxation coefficient; the distance between samples is defined as follows: , Where, and It is a sample and Class-related features.

[0032] In step 7, the clustered joint data is input into the classifier to obtain the fault type, and the total loss value is calculated. The parameters of the diagnostic model are updated in reverse to obtain a trained diagnostic model. The test set data is input into the trained diagnostic model to realize the diagnosis of the fault. The diagnostic model includes an encoder and a classifier.

[0033] Furthermore, the formula for calculating the total loss value in step 7 is: , Where, , , is the weight parameter, is the cross entropy loss, and the formula is: , Where, is the total number of samples, and are the true label and predicted probability of the i-th class of the sample, respectively. Cross-entropy loss is the loss function used for iterative classifier training. The combined data after cointegration enhancement and local clustering is input into the classifier to predict the fault type. The classifier consists of an independent single-layer fully connected network with an input dimension of 256 and an output dimension of 12. The diagnostic model is trained by determining the fault type. The test set data is then imported into the trained diagnostic model for diagnosis, and the classification results are measured using cross-entropy loss.

[0034] The joint data after cointegration enhancement and local clustering is input to train the model with the goal of accurately judging the fault type, and the test set data is input into the trained model for diagnosis.

[0035] The method described in this paper was applied to vibration signals and compared using a trained diagnostic network. Accuracy was used as a detection metric for comparative analysis on a similar bearing dataset from the University of Paderborn, Germany. The dataset consisted of bearings with both artificial and real damage, with varying fault sizes (Level 1 ≤ 2 mm or Level 2 > 2 mm). Accelerated life failures were primarily caused by fatigue pitting and artificial failures. Artificial failures were induced by EDM, drilling, and electric engraving techniques, while real failures were detected using an accelerated life test bench. The failures were classified based on damage location (inner or outer ring) and severity (Level 1 to 2). The dataset contained 12 bearing fault categories, including Level 1 EDM inner ring damage, Level 1 EDM outer ring damage, Level 1 drilling outer ring damage, Level 1 electric engraving inner ring damage, Level 1 electric engraving outer ring damage, Level 1 inner ring pitting damage, Level 1 outer ring pitting damage, Level 2 drilling outer ring damage, Level 2 electric engraving inner ring damage, Level 2 electric engraving outer ring damage, Level 2 inner ring pitting damage, and Level 2 outer ring pitting damage. The data acquisition frequency is 64 kHz, and each data acquisition lasts 4 seconds. The acquisition covers four different operating environments, each with different speeds, load torques, and radial forces. In this example, four unit domains were selected for testing: training domains 6 through 9. Training domain 6 operates at a speed of 1500 rpm, a load torque of 0.1 Nm, and a radial force of 1000 N; training domain 7 operates at a speed of 1500 rpm, a load torque of 0.7 Nm, and a radial force of 400 N; training domain 8 operates at a speed of 1500 rpm, a load torque of 0.7 Nm, and a radial force of 1000 N; and training domain 9 operates at a speed of 900 rpm, a load torque of 0.7 Nm, and a radial force of 1000 N. One of the unit domains was used as the training domain for diagnostic model training, and the other three unit domains were used as target domains for diagnostic model testing. The comparison methods used include combined invariance and specificity network (DGNIS), empirical risk minimization (ERM) and conditional contrast method (CCDG). The experimental analysis results of the average accuracy of unit domain generalization of different methods are as follows: Figure 2 and as shown in Table 1.

[0036] Table 1

[0037] The detailed analysis results of the unit domain generalization accuracy of different methods are shown in Table 2.

[0038] Table 2

[0039] Table 2 Detailed list of unit domain generalization accuracy of different methods By comparing the accuracy of different methods in various training domains, it can be seen that the single-source domain generalization fault diagnosis method based on deep learning described in the present invention has an average accuracy higher than that of the combined invariance and specificity network, the minimum empirical risk, and the conditional comparison method in different training domains such as 6, 7, 8, and 9, showing obvious performance advantages.

Claims

1. A single-source domain generalized fault diagnosis method based on deep learning, characterized by: The following steps are involved: Step 1: Collect a single-source domain ball bearing vibration signal from the bearing seat closest to the ball bearing, divide it into vibration signal sample segments, and then preprocess it. The preprocessed vibration signal sample segments are divided into a training set containing single-source domain data and a test set containing multi-source domain data; Step 2: Extract features from the data in the training set through the encoder and constrain them to obtain class-related features and class-independent features; Step 3: Enhance the class-related features and class-irrelevant features through a feature-level co-mixing enhancement module to generate enhanced class-related features and enhanced class-irrelevant features; Step 4: Process the enhancement-related features and the enhancement-irrelevant features through a decoder to generate an enhanced signal; Step 5: Calculate the cosine similarity between the enhanced signal and the original ball bearing vibration signal, and select the enhanced signal based on the cosine similarity; Step 6: Merge the original ball bearing vibration signal and the filtered enhanced signal to construct a joint data space, and perform local clustering on the samples in the joint data space; In step 7, the clustered joint data is input into the classifier to obtain the fault type, and the total loss value is calculated. The parameters of the diagnostic model are updated in reverse to obtain a trained diagnostic model. The test set data is input into the trained diagnostic model to realize the diagnosis of the fault. The diagnostic model includes an encoder and a classifier.

2. The single-source domain generalization fault diagnosis method based on deep learning according to claim 1 is characterized in that: Step 2 includes: The vibration signal sample input to the encoder is recorded as , class-related features are obtained through the encoder Class-independent features , and then constrain class-related features through feature decoupling loss Class-independent features ; The feature decoupling loss includes orthogonal loss and redundancy suppression loss , the formulas are: ， ， Where, and They are class-related loss and class-independent loss relative to class-related features. Class-independent features The gradient, represents the L2 norm, and They are class-related and class-independent losses respectively; is a square matrix with element 1, is the identity matrix, used to mask the diagonal elements, is a class-related feature matrix, each row represents a class-related feature vector of a sample, is a class-independent matrix; represents the Hadamard product, represents the Frobenius norm, is the characteristic dimension; The feature decoupling loss expression is: .

3. The single-source domain generalization fault diagnosis method based on deep learning according to claim 2 is characterized in that: Step 3 includes: Class-related features Class-independent features Input into the feature-level synergistic enhancement module for synergistic data enhancement to generate class-related enhanced features Class-independent enhancement features , the formulas are: ， ， Where, represents the enhancement coefficient, Indicates that the random index The class-related features after arrangement, Indicates that the random index Class-independent features after permutation.

4. The single-source domain generalization fault diagnosis method based on deep learning according to claim 3 is characterized in that: Step 4 includes: Enhance the class-related features Class-independent enhancement features The input is sent to the decoder, and the two feature vectors are merged into a joint feature along the feature dimension. Then the dimension is gradually reduced through the linear layer, convolutional layer and ReLU, and finally becomes the enhanced signal z mapped to the original signal dimension.

5. The single-source domain generalization fault diagnosis method based on deep learning according to claim 4 is characterized in that: Step 5 includes: Perform fast Fourier transform on the enhanced signal z and the original mechanical vibration signal to obtain an analytical signal, set the negative frequency part of the obtained analytical signal to zero, and retain the positive frequency part; Then perform inverse fast Fourier transform on the spectrum of the analytical signal to obtain the complex form of the analytical signal , according to the properties of Hilbert transform, the analytical signal Expressed as: ， Where, is the original ball bearing vibration signal, is an imaginary unit, is the Hilbert transform result of the original ball bearing vibration signal; Calculate the envelope value of the original ball bearing vibration signal using the formula: ， Calculate the cosine similarity based on the envelope value. The formula is: ， Where, and represent the envelope value of the original ball bearing vibration signal and the envelope value of the enhanced signal, respectively. is the dot product of two vectors, and yes and Length of the module; The enhanced signals z whose cosine similarity exceeds the threshold are selected and added to the training set.

6. The single-source domain generalization fault diagnosis method based on deep learning according to claim 5 is characterized in that: Step 6 includes: The original data and the filtered enhanced data are merged and the K-negative triplet loss is used to Enhance the distinction between categories, select the k negative samples farthest from the anchor point, and maximize the anchor point and negative samples The distance between the anchor points of the same type of samples is minimized and positive samples The distance between them is used to achieve clustering; Among them, K negative triple loss The formula is as follows: ， Where, is the standard relaxation coefficient; the distance between samples is defined as follows: ， Where, and It is a sample and Class-related features.

7. The single-source domain generalization fault diagnosis method based on deep learning according to any one of claims 1 to 6, characterized in that: The formula for calculating the total loss value in step 7 is: ， Where, is the weight parameter, is the feature decoupling loss, Triple loss for K; is the cross entropy loss, and the formula is: ， Where, is the total number of samples, and are the true label and predicted probability of the i-th class of sample, respectively.

Images