A bearing fault diagnosis method, device and equipment based on weighted confrontation

Abstract

The present invention discloses a bearing fault diagnosis method, device and equipment based on weighted adversarial, and relates to the technical field of rotating machinery fault detection. The present invention trains an auxiliary classifier and a label classifier in a weighted domain network model through labeled source domain data formed by historical vibration signals collected by rolling bearings, so that when the trained auxiliary classifier extracts the target domain entropy value of the target domain feature and uses posterior inference to fit the target domain entropy value to obtain the weight of the target domain feature, the feature distribution deviation of known classes in the source domain feature and the target domain feature is reduced, so that the trained label classifier can better distinguish between known classes and unknown classes in the target domain feature, and output the diagnosis result of the rolling bearing according to the known class features and the unknown class features, and the recognition accuracy of the open set cross-domain fault is greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of rotating machinery fault detection, and in particular to a bearing fault diagnosis method, device and equipment based on weighted confrontation. Background Art

[0002] Rolling bearings, as key components in rotating machinery, serve as a bridge between the stationary and rotating parts of equipment, but they are also the main source of failures. With the continuous development of industrial Internet and sensor technology, bearing fault diagnosis technology is becoming increasingly intelligent. Intelligent fault diagnosis collects a large amount of operating data through sensors, and then uses data mining technology to find hidden fault modes. Compared with traditional methods that rely on manual experience, it is more efficient and more accurate in diagnosis. Among them, the deep learning-based method automatically learns advanced features, providing a powerful tool for intelligent fault diagnosis.

[0003] The bearing fault diagnosis method based on deep learning aims to use the collected historical bearing data (training set) to train the model, and then apply it to determine the fault category of the current bearing (test set); this method is based on a premise assumption that the training set and test set of the bearing should obey the same distribution, but this assumption is difficult to meet in industrial scenarios; because during the production and operation of mechanical equipment, the working conditions such as the load and speed of the bearing change; load refers to the pressure acting on the bearing, such as when a crane lifts goods of different weights, the load borne by the bearing will change accordingly; speed refers to the speed at which the bearing rotates, and the corresponding speed will be adjusted according to the process requirements, such as the machine tool will switch to different speeds during rough processing and fine processing; this change in the working condition of the bearing causes the distribution of the collected training set and test set data to be different, resulting in a domain offset problem, that is, the model trained in the training set cannot be well generalized in the test set.

[0004] To solve the above problems, unsupervised domain adaptation technology is introduced in the field of fault diagnosis. Unsupervised domain adaptation is an effective transfer learning technology that can well transfer the fault knowledge learned from a labeled source domain data to another different but related unlabeled target domain data. Domain adaptation technology reduces the distribution difference between the source domain and the target domain by extracting domain-invariant features. However, most domain adaptation methods currently assume that the label space of the source domain and the target domain is consistent, which is called closed set cross-domain. However, in real industrial scenarios, considering the complex influence of the external environment of the bearing, such as high temperature, humidity, dust particles, acid and alkali and other environmental factors, the hazards to the bearing will overlap with each other, which may easily lead to unknown fault categories in the bearing, greatly increasing the difficulty of migration. The problem of inconsistency in the label space of two domains is called open set cross-domain; because the confusion of known class and unknown class samples in the target domain causes negative transfer in the alignment process of the two domains, that is, unknown class samples are misclassified as known classes, so the existing unsupervised domain adaptation methods cannot solve the open set cross-domain problem well; at this stage, most methods for studying the open set cross-domain problem use the output of a single sample of an auxiliary classifier or domain discriminator directly as a measurement indicator to judge the probability that the sample comes from a known class, ignoring the overall distribution information of the target domain data, so the obtained weights will be confused, and it is difficult to distinguish between known class samples and unknown class samples in the target domain, resulting in the distribution of known classes in the source domain and the target domain cannot be aligned, resulting in low recognition accuracy for open set cross-domain faults. Summary of the invention

[0005] The embodiments of the present invention provide a bearing fault diagnosis method, device and equipment based on weighted adversarial analysis, which can solve the problem that the current research on the open set cross-domain problem in the prior art is that most methods use the output of a single sample of an auxiliary classifier or a domain discriminator directly as a measurement indicator to judge the probability that the sample comes from a known class, ignoring the overall distribution information of the target domain data, so the obtained weights will be confused, and it is difficult to distinguish between known class samples and unknown class samples in the target domain, resulting in the distribution of known classes in the source domain and the target domain cannot be aligned, resulting in low recognition accuracy for open set cross-domain faults.

[0006] The embodiment of the present invention provides a bearing fault diagnosis method based on weighted confrontation, comprising the following steps:

[0007] Obtaining two-dimensional time-frequency domain images corresponding to vibration signals of rolling bearings under different working conditions; and obtaining historical vibration signals of rolling bearings;

[0008] The feature extractor is constructed using the domain adversarial network (DANN) and graph convolutional network, the auxiliary classifier including posterior inference is constructed using the Bayesian method, the domain discriminator and label classifier are constructed using the domain adversarial network (DANN) respectively, and the feature extractor, auxiliary classifier, domain discriminator and label classifier are cascaded to construct a weighted domain network model;

[0009] The historical vibration signals collected by the rolling bearing are set as labeled source domain data, and the vibration signals of the rolling bearing under different working conditions are set as target domain data; the auxiliary classifier is trained using the labeled source domain data to obtain the trained auxiliary classifier; the label classifier is trained using the labeled source domain data to obtain the trained label classifier; the trained auxiliary classifier, the trained label classifier, the feature extractor and the domain discriminator are cascaded to form a trained weighted domain network model;

[0010] The two-dimensional time-frequency domain image is input into the trained weighted domain network model; the feature extractor is used to extract the sample features of the target domain in the two-dimensional time-frequency domain image; the target domain entropy value of the target domain feature is obtained by the trained auxiliary classifier, and the target domain entropy value is fitted by the posterior inference contained in the trained auxiliary classifier to obtain the weight of the target domain feature; the domain discriminator is used to filter and remove the target domain features with small weight values; the trained label classifier is used to separate the known class features and the unknown class features in the filtered target domain features, and the fault diagnosis results of the rolling bearing are output according to the known class features and the unknown class features.

[0011] Preferably, the step of training the auxiliary classifier using labeled source domain data comprises:

[0012] Use labeled source domain data to train auxiliary classifiers To estimate the similarity between the target domain sample and the source domain sample, the loss equation is L c for:

[0013]

[0014] in: Representation sample The true label of represents the cross entropy loss; represents the labeled source domain dataset; S Indicates the source domain Source; Indicates the number of samples in the source domain; represents a feature extractor;

[0015] Auxiliary classifier By learning a K The decision boundary on the category, the entropy value of the unknown class samples in the target domain is higher than that of the known class samples;

[0016] A mixed model of two Beta distributions is established on the entropy value of the target domain samples, and the probability density function of the mixed distribution of the target domain entropy value is defined, and its equation is:

[0017]

[0018] in: Represents the target domain sample The entropy value of H represents the entropy function; and Respectively represent the mixing coefficients of the probability density functions of the known and unknown classes; the parametric equation obeyed by the probability density function of each part is:

[0019]

[0020] in: and Beta distribution parameters representing the known class part; and Parameters of the Beta distribution representing the unknown class portion.

[0021] Preferably, the updating of the distribution parameters includes:

[0022] The distribution parameters are calculated using the expectation-maximization (EM) algorithm. and Update and introduce hidden variables and Represents the posterior probability of each part of the entropy value;

[0023] Fixing parameters in the E step of the EM algorithm , use Bayes' rule to update the hidden variables, and the update equation is:

[0024]

[0025] in: ; ; represents the unlabeled target domain dataset; t Indicates the target domain Target; Represents the number of samples in the target domain; Indicates i target domain samples;

[0026] In the M step of the EM algorithm, the hidden variables in the E step are fixed, and the weighted moment method is used to update the parameters separately. , the update equation is:

[0027]

[0028] in: and Represent the weighted average and weighted variance estimates of entropy values ​​of known or unknown classes, respectively; and They represent the weighted average and weighted variance estimates of known class entropy values ​​respectively; and They represent the weighted average and weighted variance estimates of the unknown class entropy values, respectively. The calculation formula is:

[0029]

[0030] in: γ k represents the posterior probability of the entropy value, when k When it is 0, it indicates the posterior probability of the entropy value of the known class sample. k When it is 1, it indicates the posterior probability of the entropy value of the unknown class sample;

[0031] Iterate the E and M steps and calculate the posterior probability to obtain the probability that the target domain entropy value belongs to the known class as the weight of the corresponding sample. The calculation equation is:

[0032]

[0033] in: Represents the target domain sample The entropy value of H represents the entropy function; represents a feature extractor; represents the auxiliary classifier; The probability density function that represents the entropy value belonging to a known class; The probability density function that represents the entropy value belonging to the unknown class; Represents the posterior probability of the entropy value.

[0034] Preferably, the step of training a label classifier using labeled source domain data comprises:

[0035] Training a label classifier using labeled source domain data , to correctly classify known classes, the loss equation is:

[0036]

[0037] in: Indicates that each sample is assigned to K The probability of a known class; Indicates the number of samples in the source domain; represents the cross entropy loss; Representation sample The true label of represents the labeled source domain dataset;S Indicates the source domain Source; represents a feature extractor;

[0038] At the same time, the target domain samples Weight The larger the value, the greater the probability that the sample belongs to the known class; the smaller the weight of the target domain sample, that is, The larger the value, the greater the probability that the sample belongs to the unknown class; The target domain samples with large values ​​are assigned to the unknown class for training, and their loss equation is:

[0039]

[0040] in: Indicates the label of the unknown class; represents the probability of assigning the target domain sample to the unknown class; represents the unlabeled target domain dataset; t Indicates the target domain Target; Represents the number of samples in the target domain; Indicates i target domain samples.

[0041] Preferably, the training process of the weighted domain network model includes:

[0042] Pre-training feature extractor on source domain and auxiliary classifier , and obtain the initial weight of the target domain sample by fitting the mixed distribution of entropy values. The optimization equation is:

[0043]

[0044] in: The parameter value that represents the minimum value of the function in its domain; , They represent feature extractors and auxiliary classifier The weight parameter of , Respectively represent the minimum auxiliary classification loss The weight parameter value when ;

[0045] After obtaining the initial weights, start training the feature extractor alternately , auxiliary classifier , Domain Discriminator and label classifier ;

[0046] The alternating training is divided into two stages. In the first stage, the domain discriminator is learned by , and its optimization equation is:

[0047]

[0048] in: θ d Domain Discriminator The weight parameter of Indicates the minimum domain classification loss L d The weight parameter value when ;

[0049] In the second stage, the parameters are updated by and , and its optimization equation is:

[0050]

[0051] in: and They represent feature extractors and label classifier The learning parameters of and Respectively The minimum joint loss and Parameter value; represents the source domain fault category loss; represents the target domain fault category loss; represents the domain classification loss.

[0052] The embodiment of the present invention further provides a bearing fault diagnosis device based on weighted confrontation, comprising:

[0053] A detection module is used to obtain two-dimensional time-frequency domain images corresponding to vibration signals of rolling bearings under different working conditions; and to obtain historical vibration signals of rolling bearings;

[0054] Model module, using domain adversarial network DANN and graph convolutional network to build feature extractor, using Bayesian method to build auxiliary classifier including posterior inference, using domain adversarial network DANN to build domain discriminator and label classifier respectively, and cascading feature extractor, auxiliary classifier, domain discriminator and label classifier to build weighted domain network model;

[0055] The historical vibration signals collected by the rolling bearing are set as labeled source domain data, and the vibration signals of the rolling bearing under different working conditions are set as target domain data; the auxiliary classifier is trained using the labeled source domain data to obtain the trained auxiliary classifier; the label classifier is trained using the labeled source domain data to obtain the trained label classifier; the trained auxiliary classifier, the trained label classifier, the feature extractor and the domain discriminator are cascaded to form a trained weighted domain network model;

[0056] The diagnosis module is used to input the two-dimensional time-frequency domain image into the trained weighted domain network model; use the feature extractor to extract the sample features of the target domain in the two-dimensional time-frequency domain image; use the trained auxiliary classifier to obtain the target domain entropy value of the target domain feature, and fit the target domain entropy value through the posterior inference contained in the trained auxiliary classifier to obtain the weight of the target domain feature; use the domain discriminator to filter and remove the target domain features with small weight values; use the trained label classifier to separate the known class features and the unknown class features in the filtered target domain features, and output the fault diagnosis results of the rolling bearing according to the known class features and the unknown class features.

[0057] An embodiment of the present invention further provides an electronic device, including a memory and a processor;

[0058] The memory is used to store computer programs;

[0059] The processor is used to implement the steps of the above-mentioned bearing fault diagnosis method based on weighted confrontation when executing the computer program stored in the memory.

[0060] The embodiment of the present invention provides a bearing fault diagnosis method, device and equipment based on weighted confrontation. Compared with the prior art, the beneficial effects thereof are as follows:

[0061] The present invention uses labeled source domain data formed by historical vibration signals collected by rolling bearings to train auxiliary classifiers and label classifiers in a weighted domain network model, so that when the trained auxiliary classifier extracts the target domain entropy value of the target domain feature and uses posterior inference to fit the target domain entropy value to obtain the weight of the target domain feature, the feature distribution deviation of the known classes in the source domain feature and the target domain feature is reduced, so that the trained label classifier can better distinguish the known classes and unknown classes in the target domain feature, and output the diagnosis result of the rolling bearing according to the known class features and the unknown class features; the present invention clearly displays the overall distribution information of the target domain feature data, so that the acquired weights are clear, so as to clearly distinguish the known class samples and the unknown class samples in the target domain feature data, so that the feature distributions of the known classes in the source domain and the target domain are aligned, thereby greatly improving the recognition accuracy of open set cross-domain faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 A schematic diagram of the overall process of a bearing fault diagnosis method based on weighted confrontation provided by an embodiment of the present invention;

[0063] Figure 2 A schematic diagram of the overall framework of a network of a bearing fault diagnosis method based on weighted confrontation provided by an embodiment of the present invention;

[0064] Figure 3 Pre-training of a bearing fault diagnosis method based on weighted confrontation provided by an embodiment of the present invention Schematic diagram of the initial entropy value of the target domain outputted later. DETAILED DESCRIPTION

[0065] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention, so the present invention is not limited by the specific embodiments disclosed below.

[0066] See also Figure 1 The embodiment of the present invention provides a bearing fault diagnosis method based on weighted confrontation, comprising the following steps:

[0067] Step 1: After normalizing the collected raw data and performing overlapping sampling, the data is converted into a two-dimensional image as the input of the model.

[0068] Step 2: Build a weighted domain adaptation model, including feature extraction, weighting and classification modules.

[0069] Step 3: The feature extraction module extracts feature samples.

[0070] Step 4: The source domain features are used to calculate the loss through the classification module and the weighting module.

[0071] Step 5: The target domain features are used to calculate the similarity with the source domain through a weighted module, and small weights are assigned to non-shared samples.

[0072] Step 6: The loss obtained in each iteration is back-propagated to the model to update the parameters; when the iteration reaches the maximum number, the network model is determined for fault diagnosis.

[0073] In order to solve the negative transfer problem of unknown faults in the target domain, this paper proposes a weighted adversarial open-set cross-domain fault diagnosis method (Weighted Adversarial Open-Set Fault Diagnosis, WAOFD). The overall framework is as follows: Figure 2As shown; the framework mainly includes feature extractor , auxiliary classifier , domain discriminator and classifier Four modules; Feature extractor Use graph convolutional networks to learn the spatial structure features of the two domains; auxiliary classifier Used to obtain the weights for distinguishing known and unknown classes in the target domain; domain discriminator according to The weights obtained by the module are trained adversarially to align the known class distributions of the source domain and the target domain; It is used to accurately classify target domain samples or reject them as unknown classes. Specifically, the source domain is used to train an auxiliary classifier to obtain the entropy output of the target domain. After fitting the entropy distribution, the weight of the target domain sample is obtained through posterior inference. If the weight of the target domain sample is small, the domain discriminator will filter out the sample during the adversarial process. Finally, the trained classifier can accurately distinguish known classes and identify unknown classes.

[0074] Domain Adversarial Networks:

[0075] The Domain Adversarial Neural Network (DANN) is inspired by the Generative Adversarial Network (GAN). It learns the domain-invariant features of the source domain and the target domain through adversarial training, narrows the distribution difference between the two domains and realizes cross-domain diagnosis. The overall network mainly includes feature extractors , label classifier and domain discriminator Three parts: given source domain samples and target domain samples , Used to learn deep feature representations of samples from both domains, and then input source domain features into the classifier Training to enable it to accurately identify the fault type; input the two domain features into the domain discriminator Training is done to enable it to correctly distinguish whether samples come from the source domain or the target domain. The final network loss function is defined as follows:

[0076]

[0077] in: Respectively , and The learning parameters of and denote the source domain fault category loss and domain classification loss respectively; and represent the true category label and domain label respectively.

[0078] During training, by minimizing the source domain fault class loss , to achieve the correct classification of source domain fault categories; at the same time, the feature extractor and domain discriminator The two players conduct adversarial training through minimax games. By minimizing the domain classification loss Learning discriminative features makes it easier to distinguish the source domain of samples, is achieved by maximizing the domain classification loss Learning more general features to deceive To confuse it; finally, the update formula of the network parameters is:

[0079]

[0080] Weighted Adversarial Learning:

[0081] The unlabeled target domain contains both known class samples and unknown class samples. The purpose of weighted adversarial learning is to narrow the distribution difference of known classes in the source domain and the target domain, so that the classifier subsequently trained in the source domain can correctly distinguish the known classes in the target domain; domain adversarial network is an effective domain adaptation method, which trains the feature extractor to learn domain invariant features by maximizing the loss of the domain discriminator to achieve alignment of the two domains; however, adversarial learning does not perform well in open set cross-domain problems, mainly because in this process the feature distribution of the source domain and the entire target domain is aligned, and the unknown class samples will inevitably match with the source domain, and this mismatch will cause negative transfer; therefore, the present invention introduces weighted adversarial learning, using weights Representation sample The probability of belonging to a known class, the known class in the target domain will be given a larger weight, and the unknown class will be given a smaller weight. In the process of adversarial learning, the influence of the unknown class can be ignored, and the feature distribution of the known class in the target domain and the source domain can be aligned; the loss definition formula of weighted adversarial learning is as follows:

[0082]

[0083] in: =1 indicates source domain; =0 indicates the target domain; represents the binary cross entropy loss.

[0084] Posterior inference weights:

[0085] In this part, the present invention introduces how to obtain weights to evaluate the probability that each target domain sample belongs to a known class; specifically, firstly, the auxiliary classifier is trained using labeled source domain data. To estimate the similarity between the target domain sample and the source domain sample, the loss calculation formula is as follows:

[0086]

[0087] in: Representation sample The true label of represents the cross entropy loss; represents the labeled source domain dataset; S Indicates the source domain Source; Represents the number of samples in the source domain.

[0088] Auxiliary classifier By learning a K The decision boundary on the categories is given by It can be seen that the target domain unknown class samples will produce higher entropy values ​​than the known class samples; because the known class samples only have distribution differences, while the unknown class has both domain differences and semantic differences with the source domain, so When facing unknown class samples, there will be more uncertainty and the entropy value will be higher. According to the above information, the entropy value can be implicitly used as a weight to measure the target domain samples. The smaller the entropy value, the greater the probability that the sample belongs to the known class. However, there are certain limitations in directly using the entropy value of the target domain sample as an indicator to measure the unknown class and the known class: ① Through The obtained target domain entropy value is insensitive in highly uncertain and extremely sharp cases. For example, there is no difference in the entropy value when the predicted probability is (1, 0, 0, ...) and (0.5, 0.5, ...), and the entropy values ​​of known classes and unknown classes will be confused; ② Relying solely on the entropy value of a single sample is not conducive to understanding the data characteristics of the target domain from a global perspective; therefore, the present invention participates in adversarial training by fitting the distribution of all target domain entropy values ​​and taking the posterior probability that each entropy value belongs to a known class as the weight of the corresponding target domain sample.

[0089] The value range of entropy after normalization is [0, 1], so the present invention fits the entropy value distribution of the target domain by initializing Beta distribution. The formula of the probability density function of the distribution is as follows:

[0090]

[0091] Among them: distribution parameters , represents the gamma function.

[0092] According to the above information, the entropy values ​​of samples of known classes in the target domain are generally small, while the entropy values ​​of samples of unknown classes are generally large. The entropy values ​​of the target domain should present a bimodal distribution; Figure 3 For pre-training The initial entropy value result of the target domain outputted later also shows that the overall entropy value of the target domain samples presents a bimodal distribution; therefore, the present invention establishes a mixed model of two Beta distributions on the entropy value of the target domain samples, and defines the probability density function of the mixed distribution of the target domain entropy value as shown in the following formula:

[0093]

[0094] in: Represents the target domain sample The entropy value of H represents the entropy function; and Represent the mixing coefficients of the probability density functions of the known and unknown classes respectively.

[0095] The probability density function of each part obeys the following parameters:

[0096]

[0097] in: and Beta distribution parameters representing the known class part; and Parameters of the Beta distribution representing the unknown class portion.

[0098] Expectation Maximization (EM) is an iterative algorithm for estimating parameters of a probability model. The basic idea is to maximize the likelihood function by alternating two steps. The E step calculates the posterior probability of the latent variable based on the current parameter estimate; the M step uses the posterior probability obtained in the E step to maximize the expected log-likelihood function of the observed data, thereby obtaining a new round of parameter estimates. The following is a detailed introduction to using the EM algorithm to update distribution parameters. and , in order to better fit the mixed distribution of the target entropy value, we first introduce the hidden variable and To represent the posterior probability of each part of the entropy value; fix other parameters in step E , use the Bayesian rule to update the hidden variables, the calculation formula is as follows:

[0099]

[0100] in: ; ; represents the unlabeled target domain dataset; t Indicates the target domain Target; Represents the number of samples in the target domain; Indicates itarget domain samples.

[0101] In the M step, the hidden variables in the E step are fixed, and the weighted moment method is used to update the parameters separately. , the calculation formula is as follows:

[0102]

[0103] in: and They represent the weighted average and weighted variance estimates of known class entropy values ​​respectively; and They represent the weighted average and weighted variance estimates of the unknown class entropy values, respectively. The calculation formula is as follows:

[0104]

[0105] in: γ k represents the posterior probability of the entropy value, when k When it is 0, it indicates the posterior probability of the entropy value of the known class sample. k When it is 1, it indicates the posterior probability of the entropy value of the unknown class sample.

[0106] Then, the mixing coefficient and It can be calculated by the following formula:

[0107]

[0108] Iterate the above steps E and M for a certain number of times, and finally calculate the posterior probability to obtain the probability that the target domain entropy value belongs to the known class as the weight of the corresponding sample. The formula is as follows:

[0109]

[0110] in: Represents the target domain sample The entropy value of The probability density function that represents the entropy value belonging to a known class; The probability density function that represents the entropy value belonging to the unknown class; Represents the posterior probability of the entropy value.

[0111] Classifier train:

[0112] Through the previous weighted adversarial feature alignment, this part will train a classifier To accurately classify target domain samples or reject them as unknown classes; is a K +1 extended classifier, including Kknown classes and one unknown class; first, use labeled source domain data to train To enable it to achieve correct classification of known classes, the loss formula is as follows:

[0113]

[0114] in: Indicates that each sample is assigned to K The probability of a known class; Indicates the number of samples in the source domain; represents the cross entropy loss; Representation sample The true label of represents the labeled source domain dataset; S Indicates the source domain Source.

[0115] In addition, training is required It has the ability to identify unknown classes in the target domain; according to the posterior inference described above, the target domain samples Weight The larger the value, the greater the probability that the sample belongs to the known class. The larger the value, the smaller the sample The higher the probability of belonging to the unknown class, the higher the probability of belonging to the unknown class. The target domain samples with larger values ​​are assigned to the unknown class for training, and the loss definition formula is:

[0116]

[0117] in: Indicates the label of the unknown class; represents the probability of assigning the target domain sample to the unknown class; represents the unlabeled target domain dataset; t Indicates the target domain Target; Represents the number of samples in the target domain; Indicates i target domain samples.

[0118] Training process:

[0119] Update the network based on the loss mentioned above , , and , , , and Respectively represent the parameters of the corresponding network; first pre-trained on the source domain and , and then obtain the initial weight of the target domain sample by simply fitting the mixed distribution of entropy values. The optimization formula is:

[0120]

[0121] in: The parameter value that represents the minimum value of the function in its domain; , They represent feature extractors and auxiliary classifier The weight parameter of , Respectively represent the minimum The weight parameter value at .

[0122] After obtaining the initial weights, start alternating training , , and ; The alternation consists of two stages. The first stage learns the domain discriminator through the following optimization formula ;have:

[0123]

[0124] in: θ d Domain Discriminator The weight parameter of Indicates the minimum L d The weight parameter value at .

[0125] In the second stage, the parameters are updated through the following optimization formula and , and continuously update using the above formula ; Its optimization formula is:

[0126]

[0127] in: and They represent feature extractors and classifier The learning parameters of and Respectively The minimum joint loss and Parameter value; represents the source domain fault category loss; represents the target domain fault category loss; represents the domain classification loss.

[0128] The present invention proposes a new framework for solving cross-domain fault diagnosis, which mainly includes four modules: feature extractor, auxiliary classifier, domain discriminator and classifier; the feature extractor uses a graph convolutional network, which generalizes the convolutional neural network to graph data, and learns node embedding by aggregating neighbor node features through information transmission, not only learning the feature representation of connected samples, but also transmitting and updating information through the connection relationship between samples at a deeper level to capture the intrinsic relationship of the data, so that more complex feature representation can be learned; because the posterior inference in the auxiliary classifier module is based on the Bayesian formula, the relationship between the class and the sample can be better modeled in the form of conditional probability, thereby obtaining more comprehensive and accurate information, which is more helpful to accurately distinguish known classes from unknown classes, and the weights obtained through the posterior inference can better adapt to the uncertainty of data in the target domain, and improve the robustness and generalization of the model; because the known classes in the target domain will be given a larger weight and the unknown classes will be given a smaller weight in the domain discriminator module, the influence of the unknown classes can be ignored in the process of adversarial learning, and the feature distribution alignment of the known classes in the target domain and the source domain can be achieved; by expanding the dimension of the classifier to K+1, training enables it to prepare to classify target domain samples or reject them as unknown classes.

[0129] The embodiment of the present invention further provides a bearing fault diagnosis device based on weighted confrontation, comprising:

[0130] The detection module is used to obtain the two-dimensional time-frequency domain images corresponding to the vibration signals of the rolling bearing under different working conditions; and to obtain the historical vibration signals of the rolling bearing.

[0131] The model module uses the domain adversarial network DANN and graph convolutional network to build a feature extractor, uses the Bayesian method to build an auxiliary classifier containing posterior inference, uses the domain adversarial network DANN to build a domain discriminator and label classifier respectively, and cascades the feature extractor, auxiliary classifier, domain discriminator and label classifier to construct a weighted domain network model.

[0132] The historical vibration signals collected by the rolling bearing are set as labeled source domain data, and the vibration signals of the rolling bearing under different working conditions are set as target domain data; the labeled source domain data are used to train the auxiliary classifier to obtain the trained auxiliary classifier; the labeled source domain data are used to train the label classifier to obtain the trained label classifier; the trained auxiliary classifier, the trained label classifier, the feature extractor and the domain discriminator are cascaded to form a trained weighted domain network model.

[0133] The diagnosis module is used to input the two-dimensional time-frequency domain image into the trained weighted domain network model; use the feature extractor to extract the sample features of the target domain in the two-dimensional time-frequency domain image; use the trained auxiliary classifier to obtain the target domain entropy value of the target domain feature, and fit the target domain entropy value through the posterior inference contained in the trained auxiliary classifier to obtain the weight of the target domain feature; use the domain discriminator to filter and remove the target domain features with small weight values; use the trained label classifier to separate the known class features and the unknown class features in the filtered target domain features, and output the fault diagnosis results of the rolling bearing according to the known class features and the unknown class features.

[0134] An embodiment of the present invention further provides an electronic device, including a memory and a processor.

[0135] The memory is used to store computer programs.

[0136] When the processor is used to execute the computer program stored in the memory, the steps of the above-mentioned bearing fault diagnosis method based on weighted confrontation are implemented.

[0137] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A bearing fault diagnosis method based on weighted confrontation, characterized in that: The following steps are involved: Obtaining two-dimensional time-frequency domain images corresponding to vibration signals of rolling bearings under different working conditions; and obtaining historical vibration signals of rolling bearings; The feature extractor is constructed using the domain adversarial network (DANN) and graph convolutional network, the auxiliary classifier including posterior inference is constructed using the Bayesian method, the domain discriminator and label classifier are constructed using the domain adversarial network (DANN) respectively, and the feature extractor, auxiliary classifier, domain discriminator and label classifier are cascaded to construct a weighted domain network model; The historical vibration signals collected by the rolling bearing are set as labeled source domain data, and the vibration signals of the rolling bearing under different working conditions are set as target domain data; the auxiliary classifier is trained using the labeled source domain data to obtain the trained auxiliary classifier; the label classifier is trained using the labeled source domain data to obtain the trained label classifier; the trained auxiliary classifier, the trained label classifier, the feature extractor and the domain discriminator are cascaded to form a trained weighted domain network model; Input the two-dimensional time-frequency domain image into the trained weighted domain network model; use the feature extractor to extract the sample features of the target domain in the two-dimensional time-frequency domain image; use the trained auxiliary classifier to obtain the target domain entropy value of the target domain feature, and fit the target domain entropy value through the posterior inference contained in the trained auxiliary classifier to obtain the weight of the target domain feature; Use the domain discriminator to filter and remove target domain features with small weight values; The trained label classifier is used to separate the known class features and the unknown class features in the screened target domain features, and the fault diagnosis results of the rolling bearing are output based on the known class features and the unknown class features.

2. A bearing fault diagnosis method based on weighted confrontation according to claim 1, characterized in that: The step of using labeled source domain data to train an auxiliary classifier includes: Use labeled source domain data to train auxiliary classifiers To estimate the similarity between the target domain sample and the source domain sample, the loss equation is L c for: ； in: Representation sample The true label of represents the cross entropy loss; represents the labeled source domain dataset; S Indicates the source domain Source; Indicates the number of samples in the source domain; represents a feature extractor; Auxiliary classifier By learning a K The decision boundary on the category, the entropy value of the unknown class samples in the target domain is higher than that of the known class samples; A mixed model of two Beta distributions is established on the entropy value of the target domain samples, and the probability density function of the mixed distribution of the target domain entropy value is defined, and its equation is: ； in: Represents the target domain sample The entropy value of H represents the entropy function; and Respectively represent the mixing coefficients of the probability density functions of the known and unknown classes; the parametric equation obeyed by the probability density function of each part is: ； in: and Beta distribution parameters representing the known class part; and Parameters of the Beta distribution representing the unknown class portion.

3. A bearing fault diagnosis method based on weighted confrontation according to claim 2, characterized in that: The updating of the distribution parameters comprises: The distribution parameters are calculated using the expectation-maximization (EM) algorithm. and Update and introduce hidden variables and Represents the posterior probability of each part of the entropy value; Fixing parameters in the E step of the EM algorithm , use Bayes' rule to update the hidden variables, and the update equation is: ； in: ; ; represents the unlabeled target domain dataset; t Indicates the target domain Target; Represents the number of samples in the target domain; Indicates i target domain samples; In the M step of the EM algorithm, the hidden variables in the E step are fixed, and the weighted moment method is used to update the parameters separately. , the update equation is: ； in: and Represent the weighted average and weighted variance estimates of entropy values ​​of known or unknown classes, respectively; and They represent the weighted average and weighted variance estimates of known class entropy values ​​respectively; and They represent the weighted average and weighted variance estimates of the unknown class entropy values, respectively. The calculation formula is: ； in: γ k represents the posterior probability of the entropy value, when k When it is 0, it indicates the posterior probability of the entropy value of the known class sample. k When it is 1, it indicates the posterior probability of the entropy value of the unknown class sample; Iterate the E and M steps and calculate the posterior probability to obtain the probability that the target domain entropy value belongs to the known class as the weight of the corresponding sample. The calculation equation is: ； in: Represents the target domain sample The entropy value of H represents the entropy function; represents a feature extractor; represents the auxiliary classifier; The probability density function that represents the entropy value belonging to a known class; The probability density function that represents the entropy value belonging to the unknown class; Represents the posterior probability of the entropy value.

4. A bearing fault diagnosis method based on weighted confrontation according to claim 1, characterized in that: The method of using labeled source domain data to train a label classifier includes: Training a label classifier using labeled source domain data , to correctly classify known classes, the loss equation is: ； in: Indicates that each sample is assigned to K The probability of a known class; Indicates the number of samples in the source domain; represents the cross entropy loss; Representation sample The true label of represents the labeled source domain dataset; S Indicates the source domain Source; represents a feature extractor; At the same time, the target domain samples Weight The larger the value, the greater the probability that the sample belongs to the known class; the smaller the weight of the target domain sample, that is, The larger the value, the greater the probability that the sample belongs to the unknown class; The target domain samples with large values ​​are assigned to the unknown class for training, and their loss equation is: ； in: Indicates the label of the unknown class; represents the probability of assigning the target domain sample to the unknown class; represents the unlabeled target domain dataset; t Indicates the target domain Target; Represents the number of samples in the target domain; Indicates i target domain samples.

5. The bearing fault diagnosis method based on weighted confrontation according to claim 1 is characterized in that: The training process of the weighted domain network model includes: Pre-training feature extractor on source domain and auxiliary classifier , and obtain the initial weight of the target domain sample by fitting the mixed distribution of entropy values. The optimization equation is: ； in: The parameter value that represents the minimum value of the function in its domain; , They represent feature extractors and auxiliary classifier The weight parameter of , Respectively represent the minimum auxiliary classification loss The weight parameter value when ; After obtaining the initial weights, start training the feature extractor alternately , auxiliary classifier , Domain Discriminator and label classifier ; The alternating training is divided into two stages. In the first stage, the domain discriminator is learned by , and its optimization equation is: ； in: θ d Domain Discriminator The weight parameter of Indicates the minimum domain classification loss L d The weight parameter value when ; In the second stage, the parameters are updated by and , and its optimization equation is: ； in: and They represent feature extractors and label classifier The learning parameters of and Respectively The minimum joint loss and Parameter value; represents the source domain fault category loss; represents the target domain fault category loss; represents the domain classification loss.

6. A bearing fault diagnosis device based on weighted confrontation, characterized in that: A detection module is used to obtain two-dimensional time-frequency domain images corresponding to vibration signals of rolling bearings under different working conditions; and to obtain historical vibration signals of rolling bearings; Model module, using domain adversarial network DANN and graph convolutional network to build feature extractor, using Bayesian method to build auxiliary classifier including posterior inference, using domain adversarial network DANN to build domain discriminator and label classifier respectively, and cascading feature extractor, auxiliary classifier, domain discriminator and label classifier to build weighted domain network model; The historical vibration signals collected by the rolling bearing are set as labeled source domain data, and the vibration signals of the rolling bearing under different working conditions are set as target domain data; the auxiliary classifier is trained using the labeled source domain data to obtain the trained auxiliary classifier; the label classifier is trained using the labeled source domain data to obtain the trained label classifier; the trained auxiliary classifier, the trained label classifier, the feature extractor and the domain discriminator are cascaded to form a trained weighted domain network model; The diagnostic module is used to input the two-dimensional time-frequency domain image into the trained weighted domain network model; use the feature extractor to extract the sample features of the target domain in the two-dimensional time-frequency domain image; use the trained auxiliary classifier to obtain the target domain entropy value of the target domain feature, and fit the target domain entropy value through the posterior inference contained in the trained auxiliary classifier to obtain the weight of the target domain feature; Use the domain discriminator to filter and remove target domain features with small weight values; The trained label classifier is used to separate the known class features and the unknown class features in the screened target domain features, and the fault diagnosis results of the rolling bearing are output based on the known class features and the unknown class features.

7. An electronic device, characterized in that: including memory and processor; The memory is used to store computer programs; The processor is used to implement the steps of a bearing fault diagnosis method based on weighted confrontation as described in any one of claims 1 to 5 when executing the computer program stored in the memory.

Images