A Deep Learning-Based Single-Source-Domain Generalization Fault Diagnosis Method

Abstract

This invention discloses a single-source domain generalized fault diagnosis method based on deep learning, belonging to the field of mechanical fault diagnosis. The method includes: extracting features through an encoder to obtain class-related and class-independent features; enhancing the class-related and class-independent features through a feature-level co-mixing enhancement module; processing the enhanced class-related and enhanced class-independent features through a decoder; performing local clustering on samples in the training set; training the diagnostic model; and inputting test set data into the trained diagnostic model to achieve fault diagnosis. This invention extracts class-related and class-independent features through a gradient orthogonality and redundancy suppression constraint encoder, and after feature enhancement and similarity filtering, selects appropriate enhancement data to expand the dataset. Finally, it enhances the distinction between categories through local clustering and trains the model to improve diagnostic accuracy.

Description

Technical Field

[0001] This invention relates to mechanical fault diagnosis technology, and in particular to a single-source domain generalization fault diagnosis method based on deep learning. Background Technology

[0002] With the surge in industrial monitoring data and the rapid advancement of artificial intelligence technology, research on data-driven fault diagnosis technology has experienced explosive growth. Deep learning has demonstrated remarkable performance in the field of fault diagnosis, automatically extracting deep-level feature information and rapidly responding to the health status of machines. Thanks to deep learning's ability to perform end-to-end automatic feature extraction, intelligent fault diagnosis solutions have evolved from traditional signal processing and model-based methods to modern deep learning-based approaches.

[0003] In reality, the data used to train models typically 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 different physical characteristics of machines, cross-domain fault diagnosis is necessary. Multi-source data contain similar fault characteristics because they come from machines with the same operating principles and fault mechanisms. By leveraging knowledge from related datasets, high-precision and robust diagnostic models can be achieved. The most important issue in building diagnostic models is acquiring domain-invariant knowledge while eliminating domain-specific information. Therefore, Domain Generalization-based Fault Diagnosis (DGFD) has been extensively 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 vibration signal domain generalization fault diagnosis techniques, a series of problems and challenges remain. Most domain generalization-based diagnostic models are built upon multi-source domain information, and their excellent performance comes at the cost of massive amounts of training data. In real-world industrial scenarios, considering the high costs and safety risks associated with data collection, directly adopting these multi-domain methods may be impractical. Summary of the Invention

[0005] Purpose of the invention: To address the above problems, the purpose of this invention is to provide a single-source domain generalized fault diagnosis method based on deep learning. This method improves diagnostic accuracy by introducing orthogonal operations and redundancy reduction in the encoder, extracting class-related and class-independent features, selecting appropriate augmentation data to expand the dataset, and finally enhancing the distinction between categories through local clustering.

[0006] Technical solution: The single-source domain generalization fault diagnosis method based on deep learning of the present invention includes the following steps:

[0007] Step 1: Collect single-source domain ball bearing vibration signals from the bearing housing closest to the ball bearing, segment the vibration signal sample segments, perform preprocessing, and divide the preprocessed vibration signal sample segments into a training set containing single-source domain data and a test set containing multi-source domain data.

[0008] Step 2: Extract features from the data in the training set using an encoder, and after applying constraints, obtain class-related features and class-independent features;

[0009] Step 3: Enhance class-related features and class-independent features through the feature-level co-mixing enhancement module to generate enhanced class-related features and enhanced class-independent features;

[0010] Step 4: The enhanced class-related features and enhanced class-independent features are processed by the decoder to generate the enhanced signal;

[0011] 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.

[0012] Step 6: Combine the original ball bearing vibration signal with the screened enhanced signal to construct a joint data space, and perform local clustering of the samples in the joint data space;

[0013] Step 7: Input the clustered joint data into the classifier to obtain the fault type, calculate the total loss value, and update the parameters of the diagnostic model in reverse to obtain the trained diagnostic model; input the test set data into the trained diagnostic model to realize fault diagnosis; the diagnostic model includes an encoder and a classifier.

[0014] Further, step 2 includes:

[0015] The vibration signal sample input to the encoder is denoted as... Class-related features are obtained through the encoder. Class-independent features Then, the loss constraint is used to decouple the features and constrain the class-related features. Class-independent features ;

[0016] The feature decoupling loss includes orthogonality loss. and redundancy suppression loss The formulas are as follows:

[0017] ,

[0018] ,

[0019] Where, and These are class-related loss and class-independent loss relative to class-related features. Class-independent features gradient, Describing the L2 norm, and These are class-dependent and class-independent losses, respectively. It is a square matrix with an element of 1. It is an identity matrix, used to mask the diagonal elements. It is a class-related feature matrix, where each row represents the class-related feature vector of a sample. It is a class-independent matrix; This represents the Hadamard product. Denotes the Frobenius norm. It is the feature dimension;

[0020] The expression for feature decoupling loss is: .

[0021] Furthermore, step 3 includes:

[0022] Class-related features Class-independent features The input data is fed into the feature-level co-mixing enhancement module for co-mixing data enhancement, generating class-related enhanced features. Class-independent augmentation features The formulas are as follows:

[0023] ,

[0024] ,

[0025] Where, Indicates the enhancement coefficient. Indicates by random index The sorted class-related features, Indicates by random index The class-independent features after arrangement.

[0026] Furthermore, step 4 includes:

[0027] Enhance class-related features Class-independent augmentation features The input is fed to the decoder, where the two feature vectors are merged into a joint feature along the feature dimension. Then, the feature vectors are gradually reduced in dimensionality through a linear layer, a convolutional layer, and ReLU, eventually becoming an enhanced signal z mapped to the original signal dimension.

[0028] Furthermore, step 5 includes:

[0029] The enhanced signal z and the original mechanical vibration signal are subjected to fast Fourier transform to obtain the analytical signal. The negative frequency part of the obtained analytical signal is set to zero, and the positive frequency part is retained.

[0030] Then, perform an inverse fast Fourier transform on the spectrum of the analytic signal to obtain the complex form of the analytic signal. Based on the properties of the Hilbert transform, analytic signals Represented as:

[0031] ,

[0032] In the formula, It is the original ball bearing vibration signal. It is the imaginary unit. It is the Hilbert transform result of the original ball bearing vibration signal;

[0033] The envelope value of the original ball bearing vibration signal is calculated using the following formula:

[0034] ,

[0035] The cosine similarity is calculated based on the envelope value, using the following formula:

[0036] ,

[0037] In the formula, and These represent the envelope values ​​of the original ball bearing vibration signal and the enhanced signal, respectively. It is the dot product of two vectors. and yes and The modulus length;

[0038] Enhanced signals z with cosine similarity exceeding a threshold are selected and added to the training set.

[0039] Furthermore, step 6 includes:

[0040] The original data and the filtered augmented data are merged, and K-negative triple loss is applied. 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 them, while minimizing the anchor points of samples of the same class. and positive samples Clustering is achieved by considering the distance between them;

[0041] Among them, K suffers triple loss. The formula is as follows:

[0042] ,

[0043] Where, It is the standard relaxation coefficient; the distance between samples. The definition is as follows:

[0044] ,

[0045] Where, and It is a sample and Class-related features.

[0046] Furthermore, the formula for calculating the total loss value in step 7 is as follows:

[0047] ,

[0048] Where, For weight parameters, For feature decoupling loss, K suffers a triple loss; The cross-entropy loss is expressed by the following formula:

[0049] ,

[0050] Where, The total number of samples, and These are the true label and predicted probability of the i-th class of the sample, respectively.

[0051] Beneficial effects: Compared with the prior art, the significant advantages of this invention are:

[0052] 1. This 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, avoid information interference between features, and allow the encoder to focus on the core features that are valuable for category judgment.

[0053] 2. This invention combines a feature-level comixing enhancement module to expand the data in the feature space. By comparing the fit between the enhanced data and the original data in terms of feature distribution, category consistency, etc., the highest quality enhanced data is selected and added to the training set, thus solving the problem of insufficient training data.

[0054] 3. This invention enables similar features to be more tightly clustered in the feature space through local clustering, and the boundaries between different types of features are clearer, which further enhances the model's ability to identify different types of faults. Experimental results show that the fault diagnosis method described in this invention can effectively improve the accuracy of signal diagnosis. Attached Figure Description

[0055] Figure 1 This is a flowchart of the present invention;

[0056] Figure 2 A comparison chart of the average accuracy of unit domain generalization for different methods. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments.

[0058] Combination Figure 1 As shown in this embodiment, the single-source domain generalized fault diagnosis method based on deep learning includes the following steps:

[0059] Step 1: Collect single-source domain ball bearing vibration signals from the bearing housing closest to the ball bearing, segment the vibration signal sample segments, and preprocess them. Divide the preprocessed vibration signal sample segments into a training set containing single-source domain data and a test set containing multi-source domain data.

[0060] In one example, raw data is acquired from multiple source domains. Specifically, within each source domain, vibration signals from the ball bearing housing closest to the ball bearing are collected using an accelerometer and a rolling bearing testing module. These signals are used as the raw vibration signals. The raw vibration signal dataset is then segmented into training and testing sets. The training set contains 800 samples per class, and the testing set contains 200 samples per class, with each sample containing 3200 data points. All samples are normalized, and the single-source domain training data is sampled three times randomly. Precise domain labels are assigned to these samples, corresponding to domain labels 0, 1, and 2, generating datasets with clearly defined domain identifiers. These finely processed datasets with different domain labels are used as the input source for the encoder to achieve effective encoder training, thus laying a solid data foundation for subsequent model optimization and task execution.

[0061] Step 2: Extract features from the data in the training set using an encoder, and then constrain them to obtain class-related features and class-independent features.

[0062] Further, step 2 includes:

[0063] The vibration signal sample input to the encoder is denoted as... Class-related features are obtained through the encoder. Class-independent features Then, the loss constraint is used to decouple the features and constrain the class-related features. Class-independent features ;

[0064] The feature decoupling loss includes orthogonality loss. and redundancy suppression loss The formulas are as follows:

[0065] ,

[0066] ,

[0067] In the formula, and These are class-related loss and class-independent loss relative to class-related features. Class-independent features gradient, Represents the L2 norm. and These are class-dependent and class-independent losses, respectively. It is a square matrix with an element of 1. It is an identity matrix, used to mask the diagonal elements. It is a class-related feature matrix, where each row represents the class-related feature vector of a sample. It is a class-independent matrix; This represents the Hadamard product. Describing the Frobenius norm, It is the feature dimension;

[0068] The expression for feature decoupling loss is: .

[0069] In this example, the encoder consists of two convolutional neural network (CNN) subnetworks and four fully connected layers. First, convolution operations are performed using convolutional kernels. Then, features are progressively extracted and condensed through batch normalization, ReLU activation, and max pooling. After passing through four neural network layers, the dimensionality of the feature space is reduced, resulting in class-related features. Class-independent features By performing orthogonal operations and redundancy suppression constraints on the two types of features extracted, the gradient directions of the two types are made orthogonal, and the correlation between class-related features, class-independent features, and feature dimensions is reduced. Orthogonal loss. By minimizing class correlation loss Class-independent loss This is achieved through the dot product of gradients. Class-related loss. Class-independent loss The formula is as follows:

[0070] ,

[0071] ,

[0072] in, It is the cross-entropy loss function. and These are the class branch classifier and the class-independent branch classifier, respectively. and These are the input signals. Category tags and domain tags.

[0073] Step 3: Enhance class-related features and class-independent features through the feature-level comixing enhancement module to generate enhanced class-related features and enhanced class-independent features.

[0074] Furthermore, step 3 includes:

[0075] Class-related features Class-independent features The input data is fed into the feature-level co-mixing enhancement module for co-mixing data enhancement, generating class-related enhanced features. Class-independent augmentation features The formulas are as follows:

[0076] ,

[0077] ,

[0078] In the formula, This represents the enhancement coefficient, sampled from the Beta distribution. , It is a hyperparameter for feature-level comixing data enhancement. Indicates by random index The sorted class-related features, Indicates by random index The class-independent features after arrangement.

[0079] The same type and The feature-level comixing data augmentation module creates new class-related and class-independent features, retains the new feature category labels, and assigns new domain labels.

[0080] Step 4: The enhanced class-related features and enhanced class-independent features are processed by the decoder to generate an enhanced signal.

[0081] Furthermore, step 4 includes:

[0082] Enhance class-related features Class-independent augmentation features The input is fed to the decoder, where the two feature vectors are merged into a joint feature along the feature dimension. Then, the feature vectors are gradually reduced in dimensionality through a linear layer, a convolutional layer, and ReLU, eventually becoming an enhanced signal z mapped to the original signal dimension.

[0083] In this example, the decoder is used to transform the dual feature vectors, i.e., class-related augmented features. Class-independent augmentation features Decoding back to the original signal involves recovering the original signal dimensions from abstract features. By receiving two 256-dimensional feature vectors, they are merged along the feature dimensions into a 512-dimensional joint feature. Dimensionality is then gradually reduced through linear layers, convolutional layers, and ReLU, ultimately resulting in enhanced data mapped back to the original signal dimensions.

[0084] 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.

[0085] Furthermore, step 5 includes:

[0086] The enhanced signal z and the original mechanical vibration signal are subjected to fast Fourier transform to obtain the analytical signal. The negative frequency part of the obtained analytical signal is set to zero, and the positive frequency part is retained.

[0087] Then, perform an inverse fast Fourier transform on the spectrum of the analytic signal to obtain the complex form of the analytic signal. Based on the properties of the Hilbert transform, analytic signals Represented as:

[0088] ,

[0089] In the formula, It is the original ball bearing vibration signal. It is the imaginary unit. It is the Hilbert transform result of the original ball bearing vibration signal;

[0090] The envelope value of the original ball bearing vibration signal is calculated using the following formula:

[0091] ,

[0092] The cosine similarity is calculated based on the envelope value, using the following formula:

[0093] ,

[0094] In the formula, and These represent the envelope values ​​of the original ball bearing vibration signal and the enhanced signal, respectively. It is the dot product of two vectors. and yes and The modulus length;

[0095] Enhanced signals z with cosine similarity exceeding a threshold are selected and added to the training set.

[0096] In one example, an enhanced signal z with a cosine similarity of more than 85% is selected and added to the training set.

[0097] Step 6: Combine the original ball bearing vibration signal with the filtered enhanced signal to construct a joint data space, perform local clustering of samples, and achieve preliminary sample classification. By using the k-negative triple loss constraint, the samples in the training set are first given a rough distribution, such as samples of the same class being close together and samples of different classes being far apart, so that the final classification can be more accurate.

[0098] Furthermore, step 6 includes:

[0099] The original data and the filtered augmented data are merged to construct a joint data space, which is then analyzed using K-negative triple loss. 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 them, while minimizing the anchor points of samples of the same class. and positive samples Clustering is achieved by considering the distance between them;

[0100] Among them, K suffers triple loss. The formula is as follows:

[0101] ,

[0102] In the formula, It is the standard relaxation coefficient; the distance between samples. The definition is as follows:

[0103] ,

[0104] In the formula, and It is a sample and Class-related features.

[0105] Step 7: Input the clustered joint data into the classifier to obtain the fault type, calculate the total loss value, and update the parameters of the diagnostic model in reverse to obtain the trained diagnostic model; input the test set data into the trained diagnostic model to realize fault diagnosis; the diagnostic model includes an encoder and a classifier.

[0106] Furthermore, the formula for calculating the total loss value in step 7 is as follows:

[0107] ,

[0108] In the formula, , , For weight parameters, The cross-entropy loss is expressed by the following formula:

[0109] ,

[0110] Where, The total number of samples, and These are the true label and predicted probability of the i-th class of the sample, respectively. Cross-entropy loss is the loss function used during iterative training of the classifier. The joint data after comixing enhancement and local clustering is input into the classifier to predict the fault type. The classifier consists of independent single-layer fully connected networks with an input dimension of 256 and an output dimension of 12. The diagnostic model is trained based on the fault type judgment. Then, test set data is imported into the trained diagnostic model to perform diagnosis, and cross-entropy loss is used to measure the classification results.

[0111] The model is trained with the goal of accurately identifying fault types by inputting joint data that has been enhanced by comixing and local clustering, and the test set data is input into the trained model for diagnosis.

[0112] The method described in this invention was applied to vibration signals and compared using a trained diagnostic network. Accuracy was used as the detection metric for comparative analysis on the same bearing dataset from the University of Paderborn, Germany. The dataset consists of bearings with both human-caused and actual damage, with different bearing fault sizes (Level 1 ≤ 2 mm or Level 2 > 2 mm). Accelerated life failure was mainly caused by fatigue pitting and human-caused failure. Human-caused failure was induced by electrical discharge machining (EDM), drilling, and electro-engraving techniques, while actual failure was obtained through an accelerated life testing bench. Faults were classified according to damage location (inner or outer ring) and severity (Level 1 to 2). The dataset contains 12 bearing fault categories, including Level 1 EDM inner ring damage, Level 1 EDM outer ring damage, Level 1 drilled outer ring damage, Level 1 electro-engraving inner ring damage, Level 1 electro-engraving outer ring damage, Level 1 inner ring pitting damage, Level 1 outer ring pitting damage, Level 2 drilled outer ring damage, Level 2 electro-engraving inner ring damage, Level 2 electro-engraving outer ring damage, Level 2 inner ring pitting damage, and Level 2 outer ring pitting damage. The data acquisition frequency was 64 kHz, with each data acquisition session lasting 4 seconds. The acquisition range covered four different operating environments, each with different rotational speed, load torque, and radial force. In this example, four unit domains were selected for testing: training domains 6 through 9. Training domain 6 operated at a rotational speed of 1500 rpm, a load torque of 0.1 Nm, and a radial force of 1000 N; training domain 7 operated at a rotational speed of 1500 rpm, a load torque of 0.7 Nm, and a radial force of 400 N; training domain 8 operated at a rotational speed of 1500 rpm, a load torque of 0.7 Nm, and a radial force of 1000 N; and training domain 9 operated at a rotational speed of 900 rpm, a load torque of 0.7 Nm, and a radial force of 1000 N. One unit domain was used as the training domain for diagnostic model training, while the other three unit domains were used as target domains for diagnostic model testing. The comparison methods used include Combined Invariant and Specific Networks (DGNIS), Empirical Minimum Risk (ERM), and Conditional Comparison (CCDG). Experimental analysis results of the average accuracy of unit domain generalization for different methods are shown below. Figure 2 As shown in Table 1.

[0113] Table 1

[0114]

[0115] The detailed analysis results of the unit domain generalization accuracy of different methods are shown in Table 2.

[0116] Table 2

[0117]

[0118] Table 2 shows the detailed accuracy of unit domain generalization for different methods. By comparing the accuracy of different methods in each training domain, it can be seen that the single-source domain generalization fault diagnosis method based on deep learning described in this invention has a higher average accuracy in different training domains such as 6, 7, 8, and 9 than the methods combining invariant and specific networks, empirical risk minimization, and conditional comparison, demonstrating a significant performance advantage.

Claims

1. A single-source domain generalized fault diagnosis method based on deep learning, characterized in that, Includes the following steps: Step 1: Collect single-source domain ball bearing vibration signals from the bearing housing closest to the ball bearing, segment the vibration signal sample segments, perform preprocessing, and divide the preprocessed vibration signal sample segments 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 using an encoder, and after applying constraints, obtain class-related features and class-independent features; Step 3: Enhance class-related features and class-independent features through the feature-level co-mixing enhancement module to generate enhanced class-related features and enhanced class-independent features; Step 4: The enhanced class-related features and enhanced class-independent features are processed by the decoder to generate the 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: Combine the original ball bearing vibration signal with the screened enhanced signal to construct a joint data space, and perform local clustering of the samples in the joint data space; Step 7: Input the clustered joint data into the classifier to obtain the fault type, calculate the total loss value, and update the parameters of the diagnostic model in reverse to obtain the trained diagnostic model; input the test set data into the trained diagnostic model to realize fault diagnosis; 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, characterized in that, Step 2 includes: The vibration signal sample input to the encoder is denoted as... Class-related features are obtained through the encoder. Class-independent features Then, the loss constraint is used to decouple the features and constrain the class-related features. Class-independent features ; The feature decoupling loss includes orthogonality loss. and redundancy suppression loss The formulas are as follows: ， ， In the formula, and These are class-related loss and class-independent loss relative to class-related features. Class-independent features gradient, Describing the L2 norm, and These are class-dependent and class-independent losses, respectively. It is a square matrix with an element of 1. It is an identity matrix, used to mask the diagonal elements. It is a class-related feature matrix, where each row represents the class-related feature vector of a sample. It is a class-independent matrix; This represents the Hadamard product. Describing the Frobenius norm, It is the feature dimension; The expression for feature decoupling loss is: .

3. The single-source domain generalization fault diagnosis method based on deep learning according to claim 2, characterized in that, Step 3 includes: Class-related features Class-independent features The input data is fed into the feature-level co-mixing enhancement module for co-mixing data enhancement, generating class-related enhanced features. Class-independent augmentation features The formulas are as follows: ， ， In the formula, Indicates the enhancement coefficient. Indicates by random index The sorted class-related features, Indicates by random index The class-independent features after arrangement.

4. The single-source domain generalization fault diagnosis method based on deep learning according to claim 3, characterized in that, Step 4 includes: Enhance class-related features Class-independent augmentation features The input is fed to the decoder, where the two feature vectors are merged into a joint feature along the feature dimension. Then, the feature vectors are gradually reduced in dimensionality through a linear layer, a convolutional layer, and ReLU, eventually becoming an 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, characterized in that, Step 5 includes: The enhanced signal z and the original mechanical vibration signal are subjected to fast Fourier transform to obtain the analytical signal. The negative frequency part of the obtained analytical signal is set to zero, and the positive frequency part is retained. Then, perform an inverse fast Fourier transform on the spectrum of the analytic signal to obtain the complex form of the analytic signal. Based on the properties of the Hilbert transform, analytic signals Represented as: ， In the formula, It is the original ball bearing vibration signal. It is the imaginary unit. It is the Hilbert transform result of the original ball bearing vibration signal; The envelope value of the original ball bearing vibration signal is calculated using the following formula: ， The cosine similarity is calculated based on the envelope value, using the following formula: ， In the formula, and These represent the envelope values ​​of the original ball bearing vibration signal and the enhanced signal, respectively. It is the dot product of two vectors. and yes and The modulus length; Enhanced signals z with cosine similarity exceeding a 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, characterized in that, Step 6 includes: The original data and the filtered augmented data are merged, and K-negative triple loss is applied. 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 them, while minimizing the anchor points of samples of the same class. and positive samples Clustering is achieved by considering the distance between them; Among them, K suffers triple loss. The formula is as follows: ， In the formula, It is the standard relaxation coefficient; the distance between samples. The definition is as follows: ， In the formula, 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: ， In the formula, For weight parameters, For feature decoupling loss, K suffers a triple loss; The cross-entropy loss is expressed by the following formula: ， In the formula, The total number of samples, and These are the true label and predicted probability of the i-th class of the sample, respectively.

Images