Multi-source-domain unsupervised-domain adaptive bearing fault diagnosis method

Abstract

The invention is applied to the technical field of bearing fault diagnosis, and particularly discloses a multi-source-domain unsupervised-domain adaptive bearing fault diagnosis method, which comprises the following steps: S1, collecting vibration signals under different working conditions, and carrying out a preprocessing step to respectively obtain a plurality of source domain data sets and target domain data sets; and S2, establishing a network of the bearing fault diagnosis model. According to the multi-source-domain unsupervised domain adaptive bearing fault diagnosis method, a larger weight is given to a source domain sample with high similarity with a target domain through a sample weighting mechanism, and meanwhile, the confusion degree of the target domain sample is quantized and the weight is adjusted by means of prediction entropy; the method can effectively reduce the interference of dissimilar samples and easily-confused samples on the bearing diagnosis process, calculates the difference between a source domain and a target domain through the Worsteesson distance in combination with a multi-source domain information fusion module, and distributes the weight, thereby achieving the weighted fusion of multi-source domain classification results.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault diagnosis, and in particular to a multi-source domain unsupervised domain adaptive bearing fault diagnosis method. Background Art

[0002] Mechanical systems play an important role in industrial production. As an important component of mechanical systems, the safe and stable operation of rotating machinery is the guarantee of high-quality industrial production of mechanical systems. Rotating machinery is widely used in aerospace, marine survey, material mining and other fields, including generators, engines, internal combustion engines, steam turbines and other large rotor systems, as well as small equipment such as electric motors and pumps. Rotating machinery is prone to damage in harsh working environments and long-term high-load operation due to rolling bearings and gears. In order to ensure equipment safety, avoid production losses and prevent casualties, health monitoring and fault diagnosis of rotating machinery are required in the field of industrial intelligent operation and maintenance.

[0003] With the rapid development of industrial data analysis and artificial intelligence technology, data-driven methods have become the mainstream technology for solving rotating machinery fault diagnosis. Among them, deep learning-based methods have gradually become the focus and hotspot of research due to their powerful adaptive feature extraction and the lack of reliance on expert knowledge and manual feature design. However, in most existing deep learning-based rotating machinery fault diagnosis methods, a common limiting assumption is that the training data and test data are independent samples from the same distribution. In real scenarios, the operating environment and operating conditions of rotating machinery are usually complex and changeable, which leads to different distributions of training data and test data, thereby reducing the model's generalization ability of applying the pattern knowledge learned from labeled training data to unlabeled test data, ultimately affecting the accuracy of the model in fault diagnosis of rotating machinery. Summary of the Invention

[0004] The purpose of the present invention is to provide a multi-source domain unsupervised domain adaptive bearing fault diagnosis method to solve the problem raised in the above background technology that the operating environment and operating conditions of rotating machinery are complex and changeable, which affects the accuracy of the fault diagnosis model in rotating machinery fault diagnosis.

[0005] To achieve the above objectives, the present invention provides the following technical solutions: a multi-source domain unsupervised domain adaptive bearing fault diagnosis method, S1. collecting vibration signals under different working conditions, and obtaining multiple source domain data sets and target domain data sets through preprocessing steps; S2. Establish a network for the bearing fault diagnosis model, including six submodules: a shared feature extractor, a domain-specific feature extractor, a domain-specific fault classifier, a domain-specific discriminator, a domain-specific auxiliary discriminator, and multi-source domain information fusion. S3. Establish the objective function of the bearing fault diagnosis model and specify the corresponding model training strategy; S4. Inputting the training data set into the bearing fault diagnosis model and training the network structure of the bearing fault diagnosis model; S5. Input the target domain dataset into the trained bearing fault diagnosis model to implement bearing fault diagnosis; Preferably, the source domain dataset and the target domain dataset in S1 are specifically expressed as: Assume a total source domains, a single source domain dataset is represented as: ; The target domain dataset is represented as: ;

[0006] in, represents the source domain, Indicates the Source domain source domain samples, represents the set of all source domain samples, Indicates the The labels of source domain samples, where represents the number of fault categories, represents the set of all source domain sample labels, Indicates the number of samples in the source domain; represents the target domain, Indicates the target domain samples, represents the set of all target domain samples, Represents the number of samples in the target domain; The source domain is greater than or equal to two, that is, In all domains, including the source and target domains, there are three or more fault types, and the fault types are the same in all domains. The source domain is a historical operating condition where datasets of different fault types have been collected and the fault types in the datasets are known; the target domain is the operating condition where fault diagnosis is required and consists of a new operating condition sample that is different from all historical operating conditions, and the fault types under the new operating condition are the same as those in the historical conditions; The source domain and target domain follow and The probability distribution of ; represents the probability distribution obeyed by the source domain, Represents the probability distribution that the target domain obeys.

[0007] By adopting the above technical solution, the definitions and relationships of data in multiple source domains and target domains can be clarified, laying the foundation for the subsequent use of historical operating condition data to realize unsupervised fault diagnosis of new operating conditions.

[0008] Preferably, the vibration signal preprocessing method under different working conditions in S1 is the same, and the specific process of the vibration signal preprocessing is: The vibration signal is divided into samples without any overlap, expressed as , the length of each sample is The total number of samples is ; The segmented data set is normalized and divided into a training set and a test set in a ratio of 8:2. The training set and the test set are composed of a single-channel three-dimensional data set in the form of ; Indicates the number of samples in a training set or test set for a certain working condition.

[0009] By adopting the above technical solution, the vibration signal preprocessing methods of different working conditions are unified to ensure data consistency and improve diagnostic reliability.

[0010] Preferably, the network operation of establishing the bearing fault diagnosis model in S2 is as follows: S2.1: The shared feature extractor is represented as: G f (.): ; The domain-specific feature extractor is represented as: G e (.): ; The domain-specific fault classifier is expressed as: G y (.): ; The domain-specific discriminator is represented as: G d (.): ; The domain-specific auxiliary discriminator is expressed as: G d0 (.): ; in, X Represents samples from all source and target domains; Z s express X Through G f (.) Then the common feature space of all source and target domains is constructed; Z d express Through G e (.) The latter is aimed at the common feature space of a single source domain and a target domain; S Represents the probability space of fault types after model prediction; Represents the domain category probability space after model prediction; Represents the domain category probability space after model prediction; S2.2. The shared feature extractor consists of three convolutional modules, each of which consists of a one-dimensional convolutional layer, a batch normalization layer, and an activation function layer. The shared feature extractor maps samples from all domains to a common feature space. Extract downstream task feature vectors; S2.3. The domain-specific feature extractor consists of two convolution modules and an adaptive average pooling layer. The convolution module consists of a one-dimensional convolution layer, a batch normalization layer, and an activation function layer. The domain-specific feature extractor maps samples from a single source domain and a target domain in the shared feature extractor to a common feature space. Extract the downstream task feature vector from the source domain. Domain-specific feature extractors; S2.4. The domain-specific fault classifier consists of two fully connected layers, with the output node being the number of fault types. The domain-specific fault classifier maps the feature representations of the single source domain and target domain from the upstream domain-specific feature extractor to the fault type probability space. S The dimension of the output feature vector is consistent with the number of fault types, indicating the probability of belonging to a certain type of fault. The total number of source domains is domain-specific fault classifiers; S2.5. The domain-specific discriminator consists of three fully connected layers, with the output nodes being the number of domain types, which is 2. The domain-specific discriminator maps the feature representations of a single source domain and a target domain from the upstream domain-specific feature extractor to the domain type probability space. S d The dimension of the output feature vector is consistent with the number of domains, indicating the probability of belonging to a certain domain. The total number of source domains is A domain-specific discriminator; S2.6. The domain-specific auxiliary discriminator consists of three fully connected layers, with the output node being a binary classification. The domain-specific auxiliary discriminator maps the feature representations of a single source domain and a target domain from the upstream domain-specific feature extractor to a binary classification probability space. The dimension of the output feature vector is consistent with the output node, which indicates the probability of belonging to a certain domain. The probability can be converted into the weight of each sample. The total number of source domains is A domain-specific auxiliary discriminator; S2.7. The multi-source domain information fusion module obtains the weights of different domains by calculating the Wolstein distance between each pair of source domains and target domains, and performs weighted fusion on the classification results of multiple source domains to obtain the final classification result of the target domain.

[0011] By adopting the above technical solution, through multi-layer feature extraction and discriminator design, combined with multi-source domain information fusion, the accuracy and adaptability of unsupervised fault diagnosis from multi-source domains to target domains can be improved.

[0012] Preferably, the sample weighting module in S2.6 is specifically as follows: Given the extracted features Input auxiliary discriminator , and get the output of the auxiliary discriminator: ; in, is the domain label, is a binary classifier that labels source samples as 1 and target domain samples as 0. The closer to 0, the more similar the sample is to the target domain, and the closer the output is to 1, the more similar the sample is to the source domain. Therefore, the weight is Inversely proportional, it can be written as: ; Samples with high similarity to the target domain have greater weights, which can limit the negative impact of dissimilar samples. The weight function of each source sample is normalized. The normalized weight can more accurately reflect the relative importance of each sample in model training. The final weight of each sample in the source domain is expressed as follows: ; For the alignment of samples that are difficult to confuse in the target domain, prediction entropy is used to quantify the prediction uncertainty and assign different weights to target domain samples to promote sample alignment, as follows: Given the target domain sample discriminator G d The output is , then the uncertainty of the target domain sample can be expressed as: ; in, is the Dirichlet distribution parameter associated with category c, φ is the digamma function, which measures the degree of confusion of the sample by predicting entropy. The higher the entropy value, the easier it is to confuse the sample, and the lower the entropy value, the harder it is to confuse the sample. Normalize the weights of target domain samples: ; After adding weights to the source domain samples and target domain samples of the domain classifier, the objective function of the weighted adversarial network is It can be expressed as: ; in, represents the binary cross entropy function, and denote the number of samples in the source domain and the target domain respectively, and It represents the domain label.

[0013] By adopting the above technical solution and the sample weighting mechanism, the interference of dissimilar samples can be suppressed and the accuracy of cross-domain fault diagnosis can be enhanced.

[0014] Preferably, the specific process of obtaining the weight in S2.7 is: The specific formula of the Wolstein distance is as follows: ; Based on the calculation results of the above-mentioned Wolstein distance, a weight is assigned to each source domain. The source domain with less difference from the target domain should receive a higher weight, while the source domain with a large difference from the target domain should receive a lower weight. Specifically, the weight of each source domain can be calculated by the following formula: ; Subsequently, the domain-specific diagnostic information of each source domain is weighted to obtain the final diagnostic result of the target domain: .

[0015] By adopting the above technical solution, the difference between the source domain and the target domain can be quantified by calculating the Wolstein distance, which can improve the fault diagnosis accuracy in the target domain.

[0016] Preferably, the objective function of the bearing fault diagnosis model in S3 includes four loss functions, which are a source domain fault classification loss function, an auxiliary domain discriminator loss function, a weighted domain adversarial loss function, and a specific domain distribution difference loss function. During the training process of the bearing fault diagnosis model, the network parameters of each source domain fault diagnosis model are updated according to the loss function: S3.1. The source domain fault classification loss function is used to construct a mapping relationship between the high-dimensional fault features captured by the feature extractor and the fault categories. The source domain fault classification loss function uses labeled source domain data for supervised learning training. The source domain classification task first uses a shared feature extractor and a specific domain feature extractor to extract the segmented vibration signal into a specific domain feature vector. This is then input into the specific domain fault classifier, which outputs a fault class probability vector. The final single source domain fault classification objective function is expressed as follows: ; in, represents the cross entropy loss function; S3.2. The auxiliary domain discriminator loss function is used to generate better generation similarity weights. The auxiliary domain discriminator loss function will be jointly trained with the auxiliary classifier loss function to optimize the parameters of the auxiliary classifier. The specific expression of the auxiliary domain discriminator loss function is as follows: ; S3.3. The weighted domain adversarial loss function is used to align shared class features while reducing the interference of negative samples in the process, and assigns a weight to the source domain sample and the target domain that reflects its similarity with the source domain. The weight It is used to control the participation of target domain features in the feature alignment process and minimize the interference of unknown categories on the domain adaptation process. Its expression is as follows: ; S3.4. The domain-specific distribution difference minimization objective function is based on the features of the domain-specific feature extractor. Fc In terms of , it is used to reduce the difference in conditional distribution between the source domain and the target domain, and align the distribution of the two directly into the category space, which is conducive to the identification of fault categories. Its objective function is: ; S3.5. The training objectives of the fault diagnosis model include classifier loss , auxiliary domain discriminator loss , weighted domain adversarial loss and feature distribution difference loss , the optimization objective formula of the model is as follows: ; in, , and Corresponding to , and The regularization coefficient of , , are the network parameters of the shared feature extractor, domain-specific feature extractor, domain discriminator, and domain-assisted discriminator, respectively; The parameter update process formula is as follows: ; in, represents the learning rate, represents the partial derivative operator.

[0017] By adopting the above technical solution and co-optimizing multiple loss functions, the accuracy and stability of the model's cross-domain fault diagnosis can be enhanced.

[0018] Preferably, the specific process in S4 is as follows: Obtain a source domain training dataset and a target domain dataset according to the method in S1, i.e., the training dataset; The learning rate, optimizer, training batch size and number of training rounds of the network are selected, and the obtained training data set is input into the set fault diagnosis model to obtain the optimal fault diagnosis model.

[0019] By adopting the above technical solution and inputting the data set into the model training, the best fault diagnosis model can be obtained, thus ensuring the effectiveness of the model training.

[0020] Compared with the prior art, the beneficial effects of the present invention are: the multi-source domain unsupervised domain adaptive bearing fault diagnosis method: 1. This invention uses a sample weighting mechanism to assign greater weight to source domain samples with high similarity to the target domain. It also uses predicted entropy to quantify the degree of confusion of target domain samples and adjust the weights, effectively reducing the interference of dissimilar and easily confused samples on the bearing diagnosis process. Combined with a multi-source domain information fusion module, the difference between the source and target domains is calculated using the Wolstein distance and weights are assigned, achieving weighted fusion of multi-source domain classification results. This improves the accuracy of bearing fault diagnosis under different operating conditions. Taking sample differences into account in this process can effectively reduce the impact of negative samples on the model, thereby improving the diagnostic performance and stability of the fault diagnosis model. 2. The fault diagnosis model of the present invention is equipped with a shared feature extractor and a specific domain feature extractor, which not only extracts the fault features common to all domains, but also takes into account the specific features of a single source domain and a target domain. The feature distributions of the source domain and the target domain are aligned through adversarial training and other methods. The multi-source domain setting enables the model to learn knowledge from multiple historical working conditions, thereby enhancing the model's generalization ability and adaptability to bearing faults under different working conditions in complex and changeable industrial environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a schematic diagram of the process structure of the present invention; Figure 2 Schematic diagram of the model structure of the present invention; Figure 3 Schematic diagram of the confusion matrix of the present invention; Figure 4 This is a schematic diagram of the t-SNE dimensionality reduction distribution of the present invention. DETAILED DESCRIPTION

[0022] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0023] See also Figures 1-4 , the present invention provides a technical solution: a multi-source domain unsupervised domain adaptive bearing fault diagnosis method.

[0024] like Figure 1 As shown, taking a laboratory bearing dataset as an example, the steps of the method are as follows: S1. Collect vibration signals under different working conditions, and obtain source and target domain datasets through preprocessing and fast Fourier transform. The source domain dataset in S1 is represented as: ; The target domain dataset is represented as: ; in, represents the source domain, Indicates the Source domain source domain samples, represents the set of all source domain samples, Indicates the The labels of source domain samples, where represents the number of fault categories, represents the set of all source domain sample labels, Indicates the number of samples in the source domain; represents the target domain, Indicates the target domain samples, represents the set of all target domain samples, Represents the number of samples in the target domain; The dataset is divided into four source domains according to the four working conditions. The specific working condition details are shown in Table 1. The number of source domains is 3, which also means that the number of labeled historical working conditions is 3. Under each historical working condition, according to different damage modes, fault locations, damage levels and damage combinations, the specific fault information is shown in Table 2. Damage modes include pitting (P) and indentation (I), fault locations include the outer ring (OR) and inner ring (IR) of the bearing, damage combinations include single damage (S), repeated damage (R), and multiple damage (M), and the fault types are the same under all historical operating conditions;

[0025] Table 1

[0026] Table 2

[0027] The vibration signals collected under different working conditions in S1 are for a bearing fault diagnosis test platform. When it is running under certain working conditions, an acceleration sensor acquisition system is used to collect vibration acceleration signals at the motor that can reflect the vibration state of the bearing turntable. The bearing fault diagnosis test platform consists of a motor, a coupling, a front bearing, two flywheels, a distal bearing, and a load. The vibration signal preprocessing methods under different working conditions in S1 are the same, which is to divide the vibration signal into samples without any overlap. The specific process of preprocessing the historical working condition and new working condition data is as follows: The vibration signal of a certain fault category with a duration of 10 seconds and a sampling frequency of 50KHz, totaling 500500 data points, is divided into 488 samples with a length of 1024 data points. All samples are standardized according to the range of 0-1. The samples are normalized to the interval of 0-1, and then the training set and test set are divided into training set and test set respectively according to the ratio of 0.8 and 0.2. The training set and test set constitute a single-channel three-dimensional data set in the form of ; N represents the number of samples in a training set or test set for a certain working condition.

[0028] S2. Establish a network for the bearing fault diagnosis model, including six submodules: a shared feature extractor, a domain-specific feature extractor, a domain-specific fault classifier, a domain-specific discriminator, a domain-specific auxiliary discriminator, and multi-source domain information fusion. S2.1. The shared domain feature extractor consists of a one-dimensional convolutional neural network. The network includes three convolutional blocks. Each convolutional block consists of a convolution layer, a batch normalization layer, a pooling layer, and an activation layer. The convolution kernel size is 64, 3, 3. The shared feature extractor maps samples from all domains to a common feature space. Z s Extract downstream task feature vectors; S2.2. The domain-specific feature extractor consists of a one-dimensional convolutional neural network. The network includes two convolutional blocks. Each convolutional block consists of a convolution layer, a batch normalization layer, a pooling layer, and an activation layer. The convolution kernel size is 3,3. The domain-specific feature extractor concatenates the feature representations of the domain-invariant feature extractor and the feature representations of the domain-enhanced feature extractor and maps them to a common feature space. D d Extract the downstream task feature vector from , therefore, there are 2 domain-specific feature extractors; S2.3. The domain-specific fault classifier consists of three fully connected layers, with the output nodes representing the number of fault types. From front to back, they are a fully connected layer with an activation layer and 128 output nodes, a fully connected layer with an activation layer and 32 output nodes, and a fully connected layer without an activation layer and 2 output nodes. The domain-specific fault classifier maps the feature representations of the single source domain and target domain from the upstream feature extractor to the fault type probability space. S In , the output feature vector has 7 elements, which are the number of fault types, indicating the probability of belonging to a certain type of fault. Therefore, there are 3 domain-specific fault classifiers. S2.4. The domain-specific discriminator consists of two fully-connected layers, with output nodes equal to the number of domain types. From front to back, they are: a fully-connected layer with an activation layer and 128 output nodes, and a fully-connected layer without an activation layer and 2 output nodes. The domain-specific discriminator maps the feature representations of a single source domain and a target domain from the upstream feature extractor to a domain discriminator designed to distinguish features from the source domain or the target domain. The domain discriminator is adversarially trained with the feature extractor to make the feature distributions of the source and target domains similar. S2.5. The domain-specific auxiliary discriminator consists of two fully connected layers, with the output nodes being the number of domain types. From front to back, there is a fully connected layer with an activation layer and 128 output nodes, and a fully connected layer without an activation layer and 2 output nodes. The domain-specific fault classifier maps the feature representations of the single source domain and target domain from the upstream feature extractor to the fault type probability space. S In , the output feature vector has 13 elements, which represent the probability of belonging to a certain type of fault. Therefore, there are 3 domain-specific fault classifiers. S2.6. The multi-source domain information fusion module calculates the Wolstein distance between each pair of source and target domains to obtain the weights of different domains. It then performs a weighted fusion of the classification results of multiple source domains to obtain the final classification result of the target domain. Specifically, the auxiliary discriminator is described in S2.5 as follows: Given the extracted features Input auxiliary discriminator , we can get the output of the auxiliary discriminator: ; in, is the domain label, is a binary classifier that labels source samples as 1 and target domain samples as 0. The closer to 0, the more similar the sample is to the target domain, and the closer the output is to 1, the more similar the sample is to the source domain. Therefore, the weight should be Inversely proportional, it can be written as: ; In this way, samples with high similarity to the target domain have greater weights, which can limit the negative impact of dissimilar samples. In order to obtain a relatively reasonable weight value, the weight function of each source sample is normalized. The normalized weight can more accurately reflect the relative importance of each sample in model training. Finally, the weight of each sample in the source domain is: ; In order to more effectively align these difficult-to-confuse samples, prediction entropy is used to quantify the prediction uncertainty and give different weights to the target domain samples to promote sample alignment. Given the target domain sample discriminator G d The output is , then the uncertainty of the target domain sample can be expressed as: ; in, is the parameter of the Dirichlet distribution associated with class c, It is a digamma function that measures the degree of confusion of a sample by predicting entropy. The higher the entropy value, the easier it is to confuse the sample, and the lower the entropy value, the harder it is to confuse the sample. Based on this, the weights of the target domain samples are adjusted to reduce the weights of easily confused samples and increase the weights of hard-to-confuse samples. In order to ensure the consistency and standardization of the weights, the weights of the target domain samples are normalized: ; After adding weights to the source domain samples and target domain samples of the domain classifier, the objective function of the weighted adversarial network is It can be expressed as: ; The specific description of weight acquisition is: The Wolstein distance is a metric used to measure the difference between two probability distributions. It is particularly suitable for comparing distributions in high-dimensional space. The specific formula is as follows: ; Based on the calculation results of the above-mentioned Wolstein distance, a weight is assigned to each source domain. The source domain with less difference from the target domain should receive a higher weight, while the source domain with a larger difference from the target domain should receive a lower weight. Specifically, the weight of each source domain can be calculated by the following formula: ; Subsequently, the domain-specific diagnostic information of each source domain is weighted to obtain the final diagnostic result of the target domain: .

[0029] S3. Establish the objective function of the bearing fault diagnosis model and specify the corresponding model training strategy; The loss function of the bearing fault diagnosis model in S3 includes four loss functions, namely, the source domain fault classification loss function, the auxiliary domain discriminator loss function, the weighted domain adversarial loss function, and the specific domain distribution difference loss function; The source domain fault classification loss function is used to construct the mapping relationship between the high-dimensional fault features captured by the feature extractor and the fault category. The source domain fault classification loss function uses labeled source domain data for supervised learning training. The source domain classification task first uses a shared feature extractor and a specific domain feature extractor to extract the segmented vibration signal into a specific domain feature vector; then it is input into the specific domain fault classifier and outputs a fault class probability vector. The final single source domain fault classification objective function is expressed as follows: ; in, represents the cross entropy loss function; The auxiliary domain discriminator loss function is used to generate better generation similarity weights. The auxiliary domain discriminator loss function will be jointly trained with the auxiliary classifier loss function to optimize the parameters of the auxiliary classifier. The specific expression of the auxiliary domain discriminator loss function is as follows: ; The weighted domain adversarial loss function aims to align shared class features while reducing the interference of negative samples in the process, and assigns a weight to the source domain sample and the target domain that reflects its similarity to the source domain. It is used to control the participation of target domain features in the feature alignment process and minimize the interference of unknown categories on the domain adaptation process. Its expression is as follows: ; The domain-specific distribution difference minimization objective function is used to reduce the difference in conditional distribution between the source domain and the target domain for the features that pass through the domain-specific feature extractor, and align their distributions directly into the category space, which is conducive to the identification of fault categories. Its objective function is: ; The training objectives of the fault diagnosis model include the classifier loss , auxiliary domain discriminator loss , weighted adversarial loss and feature distribution difference loss , the optimization objectives of the model are as follows: ; in, , and Corresponding to , and The regularization coefficient of , , They are the network parameters of the shared feature extractor, the specific domain feature extractor, the domain discriminator and the domain auxiliary discriminator. The auxiliary domain discriminator is trained to assign more refined weights to the source domain samples. Their losses are specifically used to optimize their own parameters and will not participate in the training of the feature extractor. During the training process, by minimizing , the features extracted by forward propagation can be used to enhance the clustering effect of the classifier on source domain samples; the domain discriminator distinguishes the features from the source domain or the target domain, so its goal is to minimize In order to enhance the similarity of feature distribution between source and target domains, the feature extractor must maximize and , the parameters of the feature extractor The overall training goal is to minimize the loss function. The minus sign is added to indicate the maximization objective. The following formula shows the parameter update process: ; in, represents the learning rate, represents the partial derivative operator; S4 inputs the training data set into the bearing fault diagnosis model and trains the network structure of the bearing fault diagnosis model; The network structure of the bearing fault diagnosis model in S4 is specifically composed of six sub-modules: shared domain feature extractor, specific domain feature extractor, specific domain fault classifier, specific domain discriminator, specific domain auxiliary discriminator, and multi-source domain information fusion. The training is also aimed at the optimization of the six sub-modules. The specific process of S4 is as follows: S4.1 Obtain the source domain training dataset and the target domain dataset, i.e., the training dataset, according to the method in S1; S4.2 Select the learning rate, optimizer, training batch size and number of training rounds of the network, input the obtained training data set into the set bearing fault diagnosis model, and obtain the optimal bearing fault diagnosis model; the learning rate of the network is 0.001, the optimizer is Adam optimizer, the training batch size is 64 samples per batch; the number of training rounds is 100 iterations; among them, the weight coefficients of the domain enhancement fault classification objective function, the specific domain distribution difference minimization objective function, and the classifier regularization objective function are , in order to make its impact on the network structure smaller, As the number of iterations changes, its value is: ; in, is 10, K represents the 𝑘th iteration of the training process, Will change from 0 to 1 with the number of iterations; S5 inputs the target domain data set into the trained bearing fault diagnosis model to realize bearing fault diagnosis.

[0030] The target domain dataset is an unlabeled dataset under new working conditions, with a total of 384 samples and a sample length of 1024 data points. The target domain dataset is input into the trained model in batches of 64 samples, with a total of 6 batches. The trained bearing fault diagnosis model is a bearing fault diagnosis model trained according to the S4 training strategy. It is divided into four source domains according to the four working conditions of the dataset, namely PT1 (D1, D2, D3→D4), PT2 (D2, D3, D4→D1), PT3 (D3, D4, D1→D2) and PT4 (D4, D1, D2→D3). In order to reduce the randomness of the rotating machinery fault diagnosis results under variable working conditions, 10 repeated experiments were carried out.

[0031] The comparison method is as follows: CNN: It is a deep learning model that does not involve domain adaptation. It is trained with source domain data and directly predicts target domain samples. DAN: A classic transfer learning method that uses MMD to measure the difference in marginal probability distribution between the source and target domains. This method plays an auxiliary role by reducing the difference in marginal probability distribution between different domains. DANN: Based on DAN, an auxiliary discriminator is added to narrow the gap between the source and target domains through adversarial training and extract domain-invariant features; MSSA: Multi-source sub-domain adaptation network, which uses the same extractor to extract deep features in the time domain or frequency domain and uses the local maximum mean error as the domain adaptation loss; ADACL: It divides the multi-source domain into multiple single-source domain transfer tasks, combines adversarial and distribution alignment strategies to achieve knowledge transfer from each source domain, and finally achieves the final multi-source domain recognition result by integrating predictors; MFSAN: Use domain-specific feature extractors to extract common knowledge between different source and target domains, and combine them with domain-specific classifiers to achieve target domain recognition. Finally, strategic averaging is used to obtain the final recognition result.

[0032] Table 3

[0033] It can be concluded from the data in Table 3 that the proposed method achieved the best accuracy in all four tasks, which proves the effectiveness of the proposed method.

[0034] Working principle: By preprocessing the collected vibration signals of different working conditions, multi-source domain and target domain data sets are obtained after preprocessing, and a model containing a multi-source fusion module is constructed. Subsequently, through the sample weighting mechanism, high weights are assigned to source domain samples similar to the target domain, and the confusion of target domain samples is quantified and weighted using the predicted entropy to reduce interference. The multi-source fusion module uses the Wolstein distance to calculate the difference between the source domain and the target domain and assign weights, and weightedly fuses the classification results. The model training process is combined with classification loss, auxiliary discrimination loss, weighted adversarial loss and distribution difference loss to optimize network parameters, so that the feature distributions of the source domain and the target domain are aligned, and finally target domain bearing fault diagnosis is achieved.

[0035] While the embodiments of the present invention have been shown and described, it will be apparent to those skilled in the art that various changes, modifications, substitutions, and alterations can be made to the embodiments without departing from the principles and spirit of the invention.

Claims

1. A multi-source domain unsupervised domain adaptive bearing fault diagnosis method, characterized by: The following steps are involved: S1. Collect vibration signals under different working conditions and obtain multiple source domain datasets and target domain datasets through preprocessing steps; S2. Establish a network for the bearing fault diagnosis model, including six submodules: a shared feature extractor, a domain-specific feature extractor, a domain-specific fault classifier, a domain-specific discriminator, a domain-specific auxiliary discriminator, and multi-source domain information fusion. S3. Establish the objective function of the bearing fault diagnosis model and specify the corresponding model training strategy; S4. Inputting the training data set into the bearing fault diagnosis model and training the network structure of the bearing fault diagnosis model; S5. Input the target domain dataset into the trained bearing fault diagnosis model to implement bearing fault diagnosis.

2. The multi-source domain unsupervised domain adaptive bearing fault diagnosis method according to claim 1 is characterized by: The source domain dataset and target domain dataset in S1 are specifically expressed as: Assume a total source domains, a single source domain dataset is represented as: ； The target domain dataset is represented as: ； in, represents the source domain, Indicates the Source domain source domain samples, represents the set of all source domain samples, Indicates the Source domain The labels of source domain samples, where represents the number of fault categories, represents the set of all source domain sample labels, Indicates the source domain The number of samples; represents the target domain, Indicates the target domain samples, represents the set of all target domain samples, Represents the number of samples in the target domain; The source domain is greater than or equal to two, that is, In all domains, including the source and target domains, there are three or more fault types, and the fault types are the same in all domains. The source domain is a historical operating condition where datasets of different fault types have been collected and the fault types in the datasets are known; the target domain is the operating condition where fault diagnosis is required and consists of a new operating condition sample that is different from all historical operating conditions, and the fault types under the new operating condition are the same as those in the historical conditions; The source domain and target domain follow and The probability distribution of ; represents the probability distribution obeyed by the source domain, Represents the probability distribution that the target domain obeys.

3. The multi-source domain unsupervised domain adaptive bearing fault diagnosis method according to claim 1 is characterized by: The vibration signal preprocessing method under different working conditions in S1 is the same, and the specific process of the vibration signal preprocessing is: The vibration signal is divided into samples without any overlap, represented as , the length of each sample is The total number of samples is ; The segmented data set is normalized and divided into a training set and a test set in a ratio of 8:

2. The training set and the test set are composed of a single-channel three-dimensional data set in the form of ; Indicates the number of samples in a training set or test set for a certain working condition.

4. The multi-source domain unsupervised domain adaptive bearing fault diagnosis method according to claim 1 is characterized by: The network operation of establishing the bearing fault diagnosis model in S2 is as follows: S2.1: The shared feature extractor is represented as: G f (.): ; The domain-specific feature extractor is represented as: G e (.): ; The domain-specific fault classifier is expressed as: G y (.): ; The domain-specific discriminator is represented as: G d (.): ; The domain-specific auxiliary discriminator is expressed as: G d0 (.): ; in, X Represents samples from all source and target domains; Z s express X After G f (.) Then the common feature space of all source and target domains is constructed; Z d express Z s After G e (.) The latter is aimed at the common feature space of a single source domain and a target domain; S Represents the probability space of fault types after model prediction; S d Represents the domain category probability space after model prediction; S e Represents the domain category probability space after model prediction; S2.

2. The shared feature extractor consists of three convolutional modules, each of which consists of a one-dimensional convolutional layer, a batch normalization layer, and an activation function layer. The shared feature extractor maps samples from all domains to a common feature space. Extract downstream task feature vectors; S2.

3. The domain-specific feature extractor consists of two convolution modules and an adaptive average pooling layer. The convolution module consists of a one-dimensional convolution layer, a batch normalization layer, and an activation function layer. The domain-specific feature extractor maps samples from a single source domain and a target domain in the shared feature extractor to a common feature space. Extract the downstream task feature vector from the source domain. A domain-specific feature extractor; S2.

4. The domain-specific fault classifier consists of two fully connected layers, with the output node being the number of fault types. The domain-specific fault classifier maps the feature representations of the single source domain and target domain from the upstream domain-specific feature extractor to the fault type probability space. S The dimension of the output feature vector is consistent with the number of fault types, indicating the probability of belonging to a certain type of fault. The total number of source domains is domain-specific fault classifiers; S2.

5. The domain-specific discriminator consists of three fully connected layers, with the output nodes being the number of domain types, which is 2. The domain-specific discriminator maps the feature representations of a single source domain and a target domain from the upstream domain-specific feature extractor to the domain type probability space. S d The dimension of the output feature vector is consistent with the number of domains, indicating the probability of belonging to a certain domain. The total number of source domains is A domain-specific discriminator; S2.

6. The domain-specific auxiliary discriminator consists of three fully connected layers, with the output node being a binary classification. The domain-specific auxiliary discriminator maps the feature representations of a single source domain and a target domain from the upstream domain-specific feature extractor to a binary classification probability space. The dimension of the output feature vector is consistent with the output node, which indicates the probability of belonging to a certain domain. The probability can be converted into the weight of each sample. The total number of source domains is A domain-specific auxiliary discriminator; S2.

7. The multi-source domain information fusion module obtains the weights of different domains by calculating the Wolstein distance between each pair of source domains and target domains, and performs weighted fusion on the classification results of multiple source domains to obtain the final classification result of the target domain.

5. The multi-source domain unsupervised domain adaptive bearing fault diagnosis method according to claim 4 is characterized by: The sample weighting module in S2.6 is specifically as follows: Given the extracted features Input auxiliary discriminator , and get the output of the auxiliary discriminator: ； in, is the domain label, is a binary classifier that labels source samples as 1 and target domain samples as 0. The closer to 0, the more similar the sample is to the target domain, and the closer the output is to 1, the more similar the sample is to the source domain. Therefore, the weight is Inversely proportional, it can be written as: ； Samples with high similarity to the target domain have greater weights, which can limit the negative impact of dissimilar samples. The weight function of each source sample is normalized. The normalized weight can more accurately reflect the relative importance of each sample in model training. The final weight of each sample in the source domain is expressed as follows: ； For the alignment of samples that are difficult to confuse in the target domain, prediction entropy is used to quantify the prediction uncertainty and assign different weights to target domain samples to promote sample alignment, as follows: Given the target domain sample discriminator G d The output is , then the uncertainty of the target domain sample can be expressed as: ； in, is the Dirichlet distribution parameter associated with category c, φ is the digamma function, which measures the degree of confusion of the sample by predicting entropy. The higher the entropy value, the easier it is to confuse the sample, and the lower the entropy value, the harder it is to confuse the sample. Normalize the weights of target domain samples: ； After adding weights to the source domain samples and target domain samples of the domain classifier, the objective function of the weighted adversarial network is It can be expressed as: ； in, represents the binary cross entropy function, and denote the number of samples in the source domain and the target domain respectively, and It represents the domain label.

6. The multi-source domain unsupervised domain adaptive bearing fault diagnosis method according to claim 4 is characterized by: The specific process of obtaining the weight in S2.7 is as follows: The specific formula of the Wolstein distance is as follows: ； Based on the calculation results of the above-mentioned Wolstein distance, a weight is assigned to each source domain. The source domain with less difference from the target domain should receive a higher weight, while the source domain with a large difference from the target domain should receive a lower weight. Specifically, the weight of each source domain can be calculated by the following formula: ； Subsequently, the domain-specific diagnostic information of each source domain is weighted to obtain the final diagnostic result of the target domain: 。 7. The multi-source domain unsupervised domain adaptive bearing fault diagnosis method according to claim 6 is characterized by: The objective function of the bearing fault diagnosis model in S3 includes four loss functions, which are a source domain fault classification loss function, an auxiliary domain discriminator loss function, a weighted domain adversarial loss function, and a specific domain distribution difference loss function. During the training process of the bearing fault diagnosis model, the network parameters of each source domain fault diagnosis model are updated according to the loss function: S3.

1. The source domain fault classification loss function is used to construct a mapping relationship between the high-dimensional fault features captured by the feature extractor and the fault categories. The source domain fault classification loss function uses labeled source domain data for supervised learning training. The source domain classification task first uses a shared feature extractor and a specific domain feature extractor to extract the segmented vibration signal into a specific domain feature vector. This is then input into the specific domain fault classifier, which outputs a fault class probability vector. The final single source domain fault classification objective function is expressed as follows: ； in, represents the cross entropy loss function; S3.

2. The auxiliary domain discriminator loss function is used to generate better generation similarity weights. The auxiliary domain discriminator loss function will be jointly trained with the auxiliary classifier loss function to optimize the parameters of the auxiliary classifier. The specific expression of the auxiliary domain discriminator loss function is as follows: ； S3.

3. The weighted domain adversarial loss function is used to align shared class features while reducing the interference of negative samples in the process, and assigns a weight to the source domain sample and the target domain that reflects its similarity with the source domain. The weight It is used to control the participation of target domain features in the feature alignment process and minimize the interference of unknown categories on the domain adaptation process. Its expression is as follows: ； S3.

4. The domain-specific distribution difference minimization objective function is based on the features of the domain-specific feature extractor. Fc In terms of , it is used to reduce the difference in conditional distribution between the source domain and the target domain, and align the distribution of the two directly into the category space, which is conducive to the identification of fault categories. Its objective function is: ； S3.

5. The training objectives of the fault diagnosis model include classifier loss , auxiliary domain discriminator loss , weighted domain adversarial loss and feature distribution difference loss , the optimization objective formula of the model is as follows: ； in, , and Corresponding to , and The regularization coefficient of , , are the network parameters of the shared feature extractor, domain-specific feature extractor, domain discriminator, and domain-assisted discriminator, respectively; The parameter update process formula is as follows: ； in, represents the learning rate, represents the partial derivative operator.

8. The multi-source domain unsupervised domain adaptive bearing fault diagnosis method according to claim 1 is characterized by: The specific process in S4 is as follows: Obtain a source domain training dataset and a target domain dataset according to the method in S1, i.e., the training dataset; The learning rate, optimizer, training batch size and number of training rounds of the network are selected, and the obtained training data set is input into the set fault diagnosis model to obtain the optimal fault diagnosis model.

Images