Heterogeneous working condition loom bearing fault recognition method and model based on adversarial domain discrimination

Abstract

The application provides a heterogeneous working condition loom bearing fault recognition method and model based on an adversarial domain discriminator, machine state data of a target domain is original machine data without labeling, a model is trained through the unlabeled data, the machine fault recognition performance of the target domain is greatly improved, and time and cost of manual labeling data are reduced, time and labor are saved, the model has a better application prospect in industrial production state monitoring, and has a better use effect on industrial machine health data; through adversarial training between two classifiers and adversarial training between a separator and a working condition discriminator, heterogeneous working condition data sets with domain bias are adjusted, so that the model also has good fault recognition accuracy in a target domain data set with large bias with a source domain.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of loom bearing fault identification, and particularly relates to a heterogeneous working condition loom bearing fault identification method and model based on an adversarial domain discriminator. BACKGROUND

[0002] With the continuous expansion of industrialization scale, the machine equipment used in many industrial manufacturing fields is becoming more and more complex, and the textile industry is no exception. However, due to the harsh working environment and the long-time operation of the loom, key components such as bearings are bound to fail, affecting economic benefits and the safety of workers. Therefore, it is essential to monitor the production status of the corresponding equipment during the operation of the loom equipment, and the most important one is loom bearing fault identification. At present, the loom bearing fault identification method for production status monitoring adopted by the industry is mostly based on the experience of workers. Such a way is not only low in efficiency and high in cost, but also cannot guarantee accuracy. Therefore, the industry gradually tries to use the loom equipment data to monitor the health status of the loom and discover faults in a timely manner. Such a way can improve efficiency, reduce cost and ensure diagnostic accuracy.

[0003] Most loom bearing fault recognition methods can achieve rapid and relatively accurate bearing fault recognition, which has higher efficiency than manual loom production state monitoring, but the generalization of these algorithms is poor, and they can only perform loom production state monitoring well on the benchmark data set. In actual industrial scenarios, data from different devices often have different distributions, which will cause the fault recognition ability of the model to drop sharply on new data sets. Although there are currently technical solutions that consider the problem of different data distributions, they align the data from the source domain and the target domain, and have achieved certain results. However, these methods do not consider that the fault classes of the loom data sampled from the target domain are less than those of the loom data from the source domain. Under the original working conditions, it is easy to collect machine data of many fault classes in a healthy state, but under the new working conditions, the machine fault classes under the heterogeneous working conditions are often less than those under the original environment. Therefore, an effective method is needed to solve the problem of fault recognition when the fault classes of the loom bearing under the heterogeneous working conditions are less than those under the original working conditions. For example, the loom vibration data collected from the health management system of the production system may contain both bearing and gearbox data. If these data are directly used for fault recognition of rolling bearings in looms under heterogeneous working conditions, on the one hand, the bearing vibration data under the heterogeneous working conditions have domain bias with the original bearing vibration data, and on the other hand, the redundant gearbox data collected may have a negative effect on the training of the model. These will constrain the fault recognition performance of the model. Therefore, an effective method is needed to improve the generalization ability of the model under the new heterogeneous working conditions with fewer data classes than the original working conditions, and to reduce the high cost of retraining the model.

[0004] Currently, there have been many studies on machine bearing fault recognition, such as the residual wide kernel deep convolutional autoencoder method and the fault recognition technology based on continuous wavelet transform and deep Q learning. These methods only study fault recognition itself and do not consider domain bias and differences in class sets, which cannot improve the generalization of the model.

[0005] For example, the patent document with Chinese invention patent application number CN201810491685.2 discloses a fault migration method and a method for analyzing the impact of fault migration on HVDC commutation failure. The invention uses the fault migration method to migrate the fault in the AC network to the converter bus of the high-voltage direct current transmission system, and analyzes the AC power grid to some extent, which has good performance.

[0006] For example, the patent document with the Chinese invention patent application number CN202010286527.0 discloses a fault discrimination method based on small sample self-learning, which does not need a large amount of labeled data for training, realizes the migration of fault recognition knowledge through feature extraction combined with transfer learning, improves the fault recognition accuracy in the target data, ensures the normal operation of the machine, and has great significance for the safety production of the factory workshop.

[0007] The above-mentioned prior art has the following two obvious defects, one is that these methods are mostly only for a specific type of data, and cannot generalize the model to multiple types of machines, and cannot complete the production state monitoring task under the condition of heterogeneous working conditions; on the other hand, these fault migration methods do not consider the case that the fault categories of the data collected under the heterogeneous working conditions are less than those under the original working conditions, which will lead to the problem of decline in model fault recognition performance when these methods are applied to actual industrial scenes.

[0008] Based on the technical problems of the prior art as described above, the present application provides a heterogeneous working condition loom bearing fault recognition method and model based on adversarial domain discrimination. SUMMARY

[0009] The present application provides a heterogeneous working condition loom bearing fault recognition method and model based on adversarial domain discrimination.

[0010] The present application adopts the following technical solutions:

[0011] A heterogeneous working condition loom bearing fault recognition method and model based on adversarial domain discrimination, comprising:

[0012] Step 1, the loom state input module defines the target domain data set O of the newly collected loom fault recognition t and the source domain data set O of the existing original loom fault recognition s ;

[0013] Step 2, a loom bearing feature extraction module is constructed based on a convolutional neural network, a loom bearing fault double diagnosis module is constructed based on the idea of adversarial training, the loom bearing feature extraction module receives data transmitted from the loom state input module, extracts sample features, finds target domain data samples outside the source domain support using the difference between the predictions of two classifiers about the target domain data samples, and aligns the data distribution of the source domain and the target domain;

[0014] Step 3, the loom bearing main property separation module extracts the main properties similar to the source domain and the target domain from each sample in the data set through the adversarial training between the main property separator and the working condition discriminator;

[0015] Step 4, the heterogeneous loom bearing feature weighting module weights the data features using the main properties, and respectively averages the weighted source domain features and target domain features by column, and then aligns the target domain and source domain heterogeneous weighted feature averages using the heterogeneous weighted feature alignment module;

[0016] Step 5, training the loom bearing fault identification model according to the source domain data samples and the target domain data samples;

[0017] Step 6, inputting the target domain data set O t into the trained loom bearing fault identification model to analyze the target domain loom state data, transmitting the two classifier prediction results into the fault state output module, taking the average value output, taking the class with the highest classification confidence as the loom bearing fault class, and completing the target domain loom bearing fault identification task.

[0018] Further, in step 1, the source domain O s and the target domain O t are represented as:

[0019]

[0020]

[0021] In the above formulas (1) and (2), O s represents the source domain loom state data set, O t represents the target domain loom state data set, is the i-th data of the source domain, n s is the number of source domain data samples, i represents the i-th data sample, represents the label of the source domain data sample, i.e., the bearing fault class to which it belongs, is the i-th data of the target domain, n t is the number of target domain data samples, i represents the i-th data sample, wherein the source domain contains bearing fault classes, represents the bearing fault class space of the source domain, and the bearing fault class set of the target domain is a subset of the bearing fault class set of the source domain.

[0022] Further, step 5 includes training of the loom bearing fault identification primary model:

[0023] Step 511, the loom state input module randomly selects a certain amount of source domain data samples and target domain data samples and transmits them into the loom bearing feature extraction module;

[0024] Step 512, after the source domain data samples and the target domain data samples are processed by the loom bearing feature extraction module, multi-dimensional features are obtained;

[0025] In step 513, the weaving machine bearing fault double diagnosis module obtains multi-dimensional features, and respectively obtains bearing fault class prediction probability values of the source domain data sample and the target domain data sample, and obtains a difference loss function based on the prediction probability values of the target domain data sample Maximizing the difference loss function Based on the prediction probability values of the source domain data sample, a classification cross-entropy loss is obtained With The first-level model training is completed.

[0026] Further, step 5 includes weaving machine bearing fault recognition secondary model training:

[0027] In step 521, the weaving machine bearing fault double diagnosis module obtains multi-dimensional features, obtains sample prediction probability values, and minimizes the difference loss function Based on the prediction values of the target domain data sample by the two classifiers, a class-level weight is obtained Based on the class-level weight A class-level weighted classification cross-entropy loss is obtained

[0028] In step 522, the weaving machine bearing main property separation module obtains separated weaving machine state main properties according to sample features; and a working condition discrimination loss is constructed according to a working condition discriminator

[0029] In step 523, the heterogeneous weaving machine bearing feature weighting module maps sample features to a main property regression loss according to sample features and weaving machine state main properties The gradient value of the input layer of the heterogeneous weaving machine bearing feature weighting module is obtained The gradient value As a feature-level weighting weight value, the dimension is the same as that of the corresponding feature; the weight value is multiplied by the corresponding feature at the element level to obtain the source domain and target domain data sample features of the feature-level weighting, and the source domain and target domain data sample features are respectively averaged by dimension and input into the heterogeneous weighted feature alignment module, and then an alignment loss function is constructed Align the weighted source domain and target domain features after averaging;

[0030] In step 524, the alignment difference loss function And The total training loss of the training stage is constructed The parameters of the weaving machine bearing feature extraction module and the weaving machine bearing fault double diagnosis module are updated by using the gradient descent algorithm for back propagation, the loss function is minimized, and the model realizes weaving machine bearing fault recognition in the target domain data;

[0031] Step 525, repeat step 524 for training until the maximum number of iterations is reached.

[0032] Further, in step 513, the difference loss function obtained based on the predicted probability value of the target domain data sample is expressed as:

[0033]

[0034] In the above formula, n t represents the number of target domain data samples, i represents the number of samples, represents the probability value vector of the i-th target domain data sample predicted by the classifier A belonging to each bearing fault category, represents the probability value vector of the i-th target domain data sample predicted by the classifier B belonging to each bearing fault category, and the difference between the two is taken to reflect the difference between the two classifiers in predicting the target domain loom state sample.

[0035] Further, in step 513, and is obtained by the following formula:

[0036]

[0037] In the above formula, n s represents the number of source domain data samples, represents the number of bearing fault categories of the source domain, i represents the number of samples, and j represents the number of bearing fault categories, represents the predicted probability value of the i-th source domain data sample belonging to bearing fault category j by the classifier A or B, represents 1 when , otherwise 0.

[0038] Further, in step 521, the difference loss becomes part of the total loss in the formal training phase to update the loom bearing feature extraction module, the classifier A and the classifier B, and the class-level weight The calculation formula is as follows:

[0039]

[0040]

[0041] In the above formula, n t represents the number of target domain data samples, i represents the number of samples, represents the probability value vector of the i-th target domain data sample predicted by the classifier A belonging to each bearing fault category, a probability value vector representing the probability of the i-th target domain data sample belonging to each bearing fault class predicted by the classifier B, representing the element with the maximum value in the weight vector is normalized by equation (6).

[0042] Further, in step 521, a new class-level weighted classification cross-entropy loss is constructed based on the class-level weight

[0043]

[0044] In equation (7), n s represents the number of source domain data samples, represents the number of bearing fault classes of the source domain data, i represents the sample number, and j represents the bearing fault class number, represents the predicted probability value of the i-th source domain data sample belonging to the bearing fault class j by the classifier A, represents the predicted probability value of the i-th source domain data sample belonging to the bearing fault class j by the classifier B, represents 1 when , otherwise 0, represents the element value in the j-th bearing fault class dimension of the weight vector.

[0045] Further, in step 522, a sample discrimination loss

[0046]

[0047] wherein n s and n t represent the number of source domain and target domain data samples, respectively, i represents the sample main property number, and k i represents the domain label true value of the i-th extracted sample loom state main property, when the main property is extracted from the data sample features of the source domain, k i is 1, otherwise, k i is 0, represents the probability value of the i-th sample main property belonging to the source domain predicted by the working condition discriminator, ranging from 0 to 1;

[0048] The loss function is constructed as:

[0049]

[0050] wherein n srepresents the number of source domain data samples, i represents the number of features, j represents the number of bearing fault categories, represents the number of bearing fault categories of source domain data, represents the predicted probability value of the bearing fault classifier about the i th source domain data sample feature principal property belonging to the bearing fault category j, represents 1 when , otherwise 0.

[0051] Further, in step 523, the regression loss function that maps the sample feature to its corresponding principal property is constructed

[0052]

[0053] wherein n s and n t respectively represent the number of source domain and target domain data samples, h i represents the principal property obtained by the i th sample feature in the heterogeneous loom bearing feature weighting module, represents the principal property of the i th sample obtained by the principal property separator, the second term of formula (10) is l1 norm regularization, γ represents the penalty coefficient of l1 regularization term, the value is set to 0.5, c e represents the e th parameter of the first layer of the heterogeneous loom bearing feature weighting module, the first layer has E parameters in total;

[0054] The obtained gradient value is taken as an absolute value and subjected to l2 normalization operation, and is subjected to element level multiplication operation with the sample feature:

[0055]

[0056] wherein N represents the l2 normalization operation, represents the element level multiplication operation, X represents the feature of the sample, and Z represents the obtained weighted feature;

[0057] The weighted features of the source domain and the target domain are respectively averaged by column, and are input into the heterogeneous weighted feature alignment module to construct the weighted feature alignment loss function between the source domain and the target domain

[0058]

[0059] wherein mean() represents the average value by column, Z t represents the weighted feature of the target domain, Z s represents the weighted feature of the source domain, the l2 norm square is calculated between the averaged source domain and target domain features, and the alignment loss function to update parameters of the loom bearing feature extraction module.

[0060] Further, in step 524, the total loss function As shown below:

[0061]

[0062] In the above formula, η and θ represent hyperparameters that need to be manually set by hand, represents a class-level weighted classification loss function, represents a difference loss function, represents a weighted feature alignment loss function.

[0063] The application further provides a loom bearing fault identification model, comprising a loom state input module, a loom bearing feature extraction module, a loom bearing fault double diagnosis module, a loom bearing main property separation module, a heterogeneous loom bearing feature weighting module, a heterogeneous weighted feature alignment module, and a fault state output module, wherein:

[0064] The loom state input module is used for processing the input loom data samples;

[0065] The loom bearing feature extraction module receives the processed source domain data samples and target domain data samples, extracts multi-dimensional features from the samples, and transmits the obtained multi-dimensional features to the loom bearing fault double diagnosis module, the loom bearing main property separation module, and the heterogeneous loom bearing feature weighting module for operation;

[0066] The loom bearing fault double diagnosis module is used for completing class-level weighting operation on the data samples, and simultaneously performing preliminary alignment on the data distribution of the source domain and the target domain through the adversarial training between the classifier A and the classifier B;

[0067] The loom bearing main property separation module is used for separating the main properties similar between the source domain and the target domain from the transmitted multi-dimensional features based on the idea of adversarial domain discrimination, and helping the heterogeneous loom bearing feature weighting module to perform feature-level weighting operation;

[0068] The heterogeneous loom bearing feature weighting module is used for performing feature-level weighting operation on the multi-dimensional features of the source domain and the target domain, and taking the average value of the weighted multi-dimensional features of the source domain and the target domain;

[0069] The heterogeneous weighted feature alignment module is used for aligning the distribution of the average weighted multi-dimensional features of the source domain and the target domain;

[0070] The loom bearing fault double diagnosis module, the heterogeneous weighted feature alignment module cooperate to train the loom bearing feature extraction module, input the result obtained by the loom bearing fault double diagnosis module into the fault state output module, obtain the final loom bearing fault recognition result, and complete the loom bearing fault recognition task.

[0071] Compared with the prior art, the loom bearing fault recognition method based on the anti-domain discrimination has the following advantages:

[0072] 1. The loom bearing fault recognition method based on the anti-domain discrimination, the fault categories of the data of the target domain are less than the fault categories of the data of the source domain, in actual industrial scenes, the fault categories of the machine data of the target domain are often less than the fault categories of the data of the source domain, the performance of the loom bearing fault recognition model is not affected by the unique fault category samples of the source domain, the loom bearing fault recognition model maintains high-precision loom bearing fault recognition capability in the target domain data, and the class-level weighting mechanism and the feature-level weighting mechanism help the loom bearing fault recognition model to avoid the influence of the unique fault category samples of the source domain well.

[0073] 2. The loom bearing fault recognition method based on the anti-domain discrimination, the machine state data of the target domain are unlabeled original machine data, the model is trained through the unlabeled data, the machine fault recognition performance of the target domain is improved, the time and cost of manual labeling data are reduced, time and labor are saved, and the loom bearing fault recognition model has a better application prospect in industrial production state monitoring and a better use effect on industrial machine health data.

[0074] 3. The loom bearing fault recognition method based on the anti-domain discrimination, the anti-training between the two classifiers and the anti-training between the separator and the working condition discriminator are used to adjust the heterogeneous working condition data set with domain bias, so that the model also has good fault recognition accuracy in the target domain data set with large source domain bias, the generalization of the model is greatly improved, the model is more suitable for actual industrial scene needs, and the fault recognition performance is better. BRIEF DESCRIPTION OF DRAWINGS

[0075] Figure 1 The loom bearing fault recognition model is shown in the schematic diagram. DETAILED DESCRIPTION

[0076] In order to more clearly understand the above purpose, features and advantages of the present application, the present application will be further described in detail below with reference to the drawings and specific embodiments, and it should be explained that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0077] EMBODIMENT

[0078] As shown in Figure 1 The loom bearing fault identification model comprises a loom state input module, a loom bearing feature extraction module, a loom bearing fault double diagnosis module, a loom bearing main property separation module, a heterogeneous loom bearing feature weighting module, a heterogeneous weighted feature alignment module and a fault state output module, wherein:

[0079] The loom state input module is configured to process the input loom data sample.

[0080] The loom bearing feature extraction module receives the processed source domain data sample and target domain data sample, extracts multi-dimensional features from the sample, and transmits the obtained multi-dimensional features to the loom bearing fault double diagnosis module, the loom bearing main property separation module and the heterogeneous loom bearing feature weighting module for operation.

[0081] The loom bearing fault double diagnosis module is configured to complete the class-level weighting operation on the data sample, and simultaneously align the data distribution of the source domain and the target domain through the adversarial training between the classifier A and the classifier B. The construction of the loom bearing fault double diagnosis module and the loom bearing main property separation module is based on the idea of adversarial domain discrimination.

[0082] The loom bearing main property separation module is configured to separate the main properties similar between the source domain and the target domain from the transmitted multi-dimensional features based on the idea of adversarial domain discrimination, so as to help the subsequent heterogeneous loom bearing feature weighting module to perform feature-level weighting operation.

[0083] The heterogeneous loom bearing feature weighting module is configured to perform feature-level weighting operation on the multi-dimensional features of the source domain and the target domain, and take the average value of the weighted multi-dimensional features of the source domain and the target domain.

[0084] The heterogeneous weighted feature alignment module is configured to align the distribution of the average weighted multi-dimensional features of the source domain and the target domain.

[0085] The loom bearing fault double diagnosis module and the heterogeneous weighted feature alignment module cooperate with each other to train the loom bearing feature extraction module. The result obtained by the loom bearing fault double diagnosis module is input into the fault state output module to obtain the final loom bearing fault identification result, thereby completing the task of loom bearing fault identification.

[0086] In the model training data set, the heterogeneous source domain and target domain data sets are included. The data set of the source domain is represented as:

[0087]

[0088] Wherein, n s represents the number of source domain data samples, and i represents the sample number. the label of the source domain data sample, the i-th data of the source domain; the target domain data set is not annotated with bearing fault classes, denoted as:

[0089]

[0090] wherein n t is the number of target domain data samples, i represents the i-th data sample, represents the target domain data sample;

[0091] The source domain, i.e. the working condition in which the loom data sample and the corresponding bearing fault label can be obtained, the target domain, i.e. the heterogeneous working condition, in which the bearing fault label of the data sample cannot be obtained, only the loom data sample without label can be obtained, the source domain and the target domain are heterogeneous working conditions, so the data distribution of the sample is different; the source domain data sample and the target domain data sample correspond to the loom data sample collected in the source domain and the target domain respectively, which are the samples belonging to the corresponding source domain data set and target domain data set, and are used for model training; the source domain feature refers to the corresponding multi-dimensional feature generated by the loom bearing feature extraction module receiving the source domain data sample, and the target domain feature is the same.

[0092] Constructing the loom bearing feature extraction module and the loom bearing fault double diagnosis module:

[0093] The loom bearing fault double diagnosis module utilizes the idea of adversarial training, increases the difference loss of the prediction of the target domain data sample in the pre-training stage

[0094]

[0095] wherein n t represents the number of target domain data samples, represents the class prediction probability value of classifier A about the i-th sample, represents the class prediction probability value of classifier B about the i-th sample, the difference loss is increased in the pre-training stage, and the parameters of the two classifiers are updated accordingly, helping the model to find the samples of the source domain support edge, which is beneficial to the alignment of the target domain and the source domain data distribution in the next step, the difference loss is reduced in the formal training stage, and the loom bearing feature extraction module and the two classifiers are updated according to the difference loss, helping the model to align the source domain and the target domain data distribution, so that the model can also have good fault recognition performance in the target domain data.

[0096] In the pre-training stage, the following is also used:

[0097]

[0098] This allows the model to have a preliminary classification ability on the source domain data samples, helping the model to perform class-level weighting operations in the formal stage, where j represents the bearing fault category number. The number of bearing fault categories representing the source domain data. This represents the probability that classifiers A / B will predict a data sample from source domain i as bearing fault category j. Representative When the time condition is met, its value is 1; otherwise, it is 0. For the true label of bearing failure category of sample i, use To update the parameters of the loom bearing feature extraction module, classifier A, and classifier B during the pre-training phase;

[0099] In addition, to avoid the negative impact of source domain-specific bearing fault category samples on the model, a dual diagnosis module for loom bearing faults is used to calculate class-level weights during the formal training phase. The calculation process consists of two steps:

[0100]

[0101]

[0102] in, Representative seeking The value of the largest element in the vector, representing the class-level weight, is used to predict the bearing fault category of the target domain data samples using two classifiers. Samples are likely to be classified into a shared bearing fault category between the source and target domains. Therefore, samples with bearing fault categories unique to the source domain are assigned smaller weight values ​​to reduce their impact on model parameter updates. Based on this weight value... Construct a class-weighted classification loss function:

[0103]

[0104] Where, n s Represents the number of data samples in the source domain. The number of bearing fault categories represents the number of data points in the source domain, where i represents the data number and j represents the bearing fault category number. This represents the predicted probability value of classifier A for the i-th source domain data sample belonging to bearing fault category j. This represents the predicted probability value of classifier B for the i-th source domain data sample belonging to bearing fault category j. Representative When the time condition is met, the value is 1; otherwise, the value is 0. Represents weight The magnitude of the element value corresponding to the bearing fault category dimension in the vector, the loom bearing feature extraction module and the two classifiers, as shownFigure 1 The parameters of the loom bearing feature extraction module and the two classifiers are updated using the loss function, so that the model can perform fault identification on the samples of the target domain, while avoiding the negative impact of the samples of the unique bearing fault categories of the source domain on the model performance to a certain extent;

[0105] The loom bearing main property separation module, heterogeneous loom bearing feature weighting module, and heterogeneous weighted feature alignment module are constructed to weight the feature dimensions of the data, so as to better align the data of the target domain with the data of the source domain, and also to weight the samples in a more fine-grained manner, thereby avoiding the loss of useful information in the source domain unique category samples;

[0106] In the loom bearing main property separation module, the main property separator, the working condition discriminator, and the fault classifier are included. The input of the module is the sample features extracted by the loom bearing feature extraction module, and the output is the main property separated from the sample features by the main property separator. The main property separator and the working condition discriminator constitute an adversarial training relationship, and a loss function is constructed:

[0107]

[0108] Wherein, n s and n t represent the number of data samples of the source domain and the target domain respectively, i represents the sample number, represents the probability value of the main property of the i-th sample predicted by the working condition discriminator belonging to the source domain working condition, k i represents the domain label true value of the i-th main property, and the training target of the working condition discriminator is to minimize that is, to judge the source of the main property obtained by the separator as much as possible, and the training target of the main property separator is to maximize that is, to separate the main property similar to the source domain and the target domain from the features as much as possible to confuse the working condition discriminator; the adversarial training of the two makes the main property separator able to separate the main property similar to the source domain and the target domain as much as possible to help the subsequent feature weighting. In addition, in order to ensure the effectiveness of the main property, a fault classifier is added, and a classification cross-entropy loss function is constructed:

[0109]

[0110] Wherein, represents the probability value of the i-th source domain data sample belonging to the bearing fault category j predicted by the fault classifier, and after the restriction of is applied, the main property separated by the main property separator will be more meaningful, rather than simply generating some meaningless vectors such as unit vectors in order to confuse the working condition discriminator. The loss function of the loom bearing main property separation module does not affect the parameter update of other modules.

[0111] After the principal property separator is trained, the principal properties it obtains will be input into the heterogeneous loom bearing feature weighting module. The heterogeneous loom bearing feature weighting module receives data features from the loom bearing feature extraction module and constructs a regression loss function:

[0112]

[0113] Among them, h i The principal properties obtained by the feature weighting module of heterogeneous loom bearings through data feature mapping are represented by this module. The principal property represents the principal property corresponding to the true i-th sample input to the principal property separator. By mapping features to the corresponding principal properties, the gradient value of the input layer can be obtained. This gradient value is used as the weight of the feature to characterize the contribution of each dimension of the feature to the alignment of the source and target domain data. The second part of the formula is to ensure the sparsity of the input layer gradient, where γ represents the manually set penalty coefficient, e represents the input layer parameter number, E represents the total number of input layer parameters, and c e This represents the e-th parameter of the input layer; this regression loss does not affect the parameter updates of other modules. After obtaining the gradient value, it is processed... The processing involves taking the absolute value and performing N(l2 normalization) operations, followed by element-wise multiplication with the sample feature X. The operation yields a weighted feature Z. The weighted features of the source and target domains are averaged along their respective feature dimensions and then input into the heterogeneous weighted feature alignment module to align the average weighted features of the source and target domains, thus constructing the loss function.

[0114]

[0115] Here, mean represents the average value along the feature dimension. Minimizing this loss updates the parameters of the loom bearing feature extraction module, helping the model align the data distribution between the source and target domains.

[0116] The overall loss of the model during the formal training phase is:

[0117]

[0118] Where η and θ are manually set hyperparameters. Represents the class-weighted classification loss. This represents the loss due to the difference in predictions between data samples in the target domain. To align the weighted feature loss, the gradient descent algorithm is used for back propagation to update the parameters of the loom bearing feature extraction module and the loom bearing fault double diagnosis module, minimize the loss function, realize the feature alignment training of the source domain data and the target domain data, and finally realize the loom bearing fault recognition task of the target domain.

[0119] The loom bearing fault recognition method based on the anti-domain discrimination of heterogeneous working conditions comprises:

[0120] Step 1, the loom state input module defines the newly collected loom fault recognition target domain data set O t and the existing original loom fault recognition source domain data set O s .

[0121] Step 2, the loom bearing feature extraction module is constructed based on the convolutional neural network, and the loom bearing fault double diagnosis module is constructed based on the idea of anti-training. The loom bearing feature extraction module receives the data transmitted from the loom state input module, extracts the sample features, finds the target domain data samples outside the source domain support by using the difference between the predictions of the two classifiers about the target domain data samples, and aligns the data distribution of the source domain and the target domain.

[0122] Step 3, the loom bearing main property separation module separates the main properties of the source domain and the target domain from each sample in the data set through the anti-training between the main property separator and the working condition discriminator.

[0123] Step 4, the heterogeneous loom bearing feature weighting module weights the data features by using the main properties, and respectively takes the average values of the weighted source domain features and target domain features by column, and then aligns the heterogeneous weighted features of the target domain and the source domain by using the heterogeneous weighted feature alignment module.

[0124] Step 5, the loom bearing fault recognition model is trained according to the source domain data samples and the target domain data samples.

[0125] Step 6, the target domain data set O t is input into the trained loom bearing fault recognition model, the loom state data of the target domain is analyzed, the prediction results of the two classifiers are input into the fault state output module, the average value is output, the class with the maximum classification confidence is taken as the fault class of the loom bearing, and the loom bearing fault recognition task of the target domain is completed.

[0126] In step 5 of the above embodiment, the loom bearing fault recognition first-level model training comprises:

[0127] Step 511, the loom state input module randomly selects a certain amount of source domain data samples and target domain data samples and transmits them into the loom bearing feature extraction module;

[0128] Step 512, after the source domain data samples and the target domain data samples are processed by the loom bearing feature extraction module, multi-dimensional features are obtained;

[0129] Step 513, the loom bearing fault double diagnosis module obtains the multi-dimensional features and respectively obtains the bearing fault class prediction probability values of the source domain data samples and the target domain data samples, and obtains the difference loss based on the prediction probability values of the target domain data samples Maximize the difference loss Based on the prediction probability values of the source domain data samples, the classification cross-entropy loss is obtained And Complete the first model training.

[0130] In step 5 of the above embodiment, the loom bearing fault recognition secondary model training is also included:

[0131] Step 521, the loom bearing fault double diagnosis module obtains the multi-dimensional features, obtains the sample prediction probability values, and minimizes the difference loss The difference loss here The difference loss used in step 513 It is the same loss function, except that in step 513, the goal of training is to maximize the loss, while in this step, the goal of training is to minimize the loss, based on the prediction values of the two classifiers for the target domain data samples, to obtain the class-level weight Based on the weight, the class-level weighted classification cross-entropy loss is obtained

[0132] Step 522, the loom bearing main property separation module obtains the separated loom state main properties according to the sample features; and constructs a working condition discrimination loss according to the working condition discriminator

[0133] Step 523, the heterogeneous loom bearing feature weighting module maps the sample features to the main property regression loss according to the sample features and the loom state main properties Obtain the gradient value of the input layer of the heterogeneous loom bearing feature weighting module Gradient value As a feature-level weighting weight value, the dimension is the same as that of the corresponding feature; the weight value and the corresponding feature are multiplied element by element to obtain the source domain and target domain data sample features of the feature-level weighting, and the source domain and target domain data sample features are respectively averaged by dimension and transmitted into the heterogeneous weighted feature alignment module, and then the alignment loss function is constructed align the weighted source domain and target domain features after averaging;

[0134] Step 524, using the alignment loss function and constructing the total training loss of the training stage using the gradient descent algorithm to perform back propagation, update the parameters of the loom bearing feature extraction module and the loom bearing fault double diagnosis module, minimize the loss function, and realize the loom bearing fault recognition of the model in the target domain data;

[0135] Step 525, repeat step 524 for training until the maximum number of iterations is reached.

[0136] The present application is not limited to the above-mentioned embodiments, and the above-mentioned embodiments and descriptions are only to illustrate the principles of the present application, and various changes and improvements can be made without departing from the spirit and scope of the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims.

Claims

1. A method for identifying bearing faults of a heterogeneous operating condition loom based on adversarial domain discrimination, characterized in that, Comprising: Step 1, the loom state input module defines a newly collected target domain data set O of loom fault recognition t and a source domain data set O of original loom fault recognition that already exists s ; Step 2, based on the convolutional neural network, a loom bearing feature extraction module is constructed, and a loom bearing fault double diagnosis module is constructed based on the idea of adversarial training. The loom bearing feature extraction module receives data from the loom state input module, extracts sample features, and finds target domain data samples outside the source domain support by using the difference between the two classifiers about the target domain data sample prediction, and aligns the data distribution of the source domain and the target domain; Step 3, the loom bearing main property separation module separates the main properties of the loom bearing through the adversarial training between the main property separator and the working condition discriminator, and extracts the main properties similar to the source domain and the target domain from each sample in the data set; Step 4, the heterogeneous loom bearing feature weighting module weights the data features using the main properties, and takes the average value of the weighted source domain features and target domain features respectively, and then aligns the heterogeneous weighted features of the target domain and the source domain using the heterogeneous weighted feature alignment module; Step 5, according to the source domain data sample and the target domain data sample, the loom bearing fault recognition model is trained: Step 511, the loom state input module randomly selects a certain amount of source domain data samples and target domain data samples, and transmits them into the loom bearing feature extraction module; Step 512, after the source domain data sample and the target domain data sample data are processed by the loom bearing feature extraction module, multi-dimensional features are obtained; In step 513, the loom bearing fault double diagnosis module obtains multi-dimensional features, and respectively obtains bearing fault class prediction probability values of the source domain data samples and the target domain data samples, and obtains a difference loss based on the prediction probability values of the target domain data samples Maximizing the difference loss Based on the prediction probability values of the source domain data samples, a classification cross-entropy loss is obtained With Complete primary model training; The loom bearing fault recognition two-level model training includes: In step 521, the weaving machine bearing fault double diagnosis module obtains multi-dimensional features, obtains sample prediction probability values, and minimizes the difference loss Based on the prediction values of the target domain data samples by the two classifiers, class-level weights are obtained Based on the class-level weights Class-level weighted classification cross-entropy loss is obtained At step 522, the loom bearing main property separation module obtains separated loom state main properties according to the sample features; and constructs a working condition discrimination loss according to the working condition discriminator Step 523, the heterogeneous loom bearing feature weighting module maps the sample features to the regression loss of the main properties according to the sample features and the main properties of the loom state Get the gradient value of the input layer of the heterogeneous loom bearing feature weighting module Gradient value As the weight value of feature-level weighting, the dimension is the same as that of the corresponding feature; element-level multiplication operation is performed on the weight value and the corresponding feature to obtain the source domain and target domain data sample features of feature-level weighting; the source domain and target domain data sample features are respectively averaged by dimension and input into the heterogeneous weighted feature alignment module, and then the alignment loss function is constructed Align the weighted source domain and target domain features after averaging Step 524, using an alignment loss function and constructing the total training loss of the training stage using a gradient descent algorithm to perform back propagation, update the parameters of the loom bearing feature extraction module and the loom bearing fault double diagnosis module, minimize the loss function, and realize the loom bearing fault recognition of the model in the target domain data; Step 525, repeat step 524 for training until the maximum number of iterations is reached; Step 6, the target domain data set O t The loom state data of the target domain is analyzed by inputting into the trained loom bearing fault identification model, the two classifier prediction results are transmitted into the fault state output module, the mean value is output, the class with the maximum classification confidence is taken as the fault class of the loom bearing, and the loom bearing fault identification task of the target domain is completed.

2. The method for recognizing bearing fault of heterogeneous working condition loom shaft based on adversarial domain discrimination according to claim 1, characterized in that, In step 1, the source domain O s and the target domain O t are represented: In the above formulas (1) and (2), O s represents the source domain loom state data set, O t represents the target domain loom state data set, is the i th data of the source domain, n s is the number of source domain data samples, i represents the i th data sample, represents the label of the source domain data sample, i.e., the bearing fault category to which it belongs, is the i th data of the target domain, n t is the number of target domain data samples, i represents the i th data sample, wherein the source domain contains fault categories, represents the fault category space of the source domain, and the bearing fault category set of the target domain is a subset of the bearing fault category set of the source domain.

3. The method of claim 1, wherein, In step 513, the difference loss obtained based on the prediction probability value of the target domain data sample is represented as: In the above formula, n t represents the number of target domain data samples, i represents the i-th data sample, represents the probability value vector of the i-th target domain data sample belonging to each bearing fault category predicted by the classifier A, represents the probability value vector of the i-th target domain data sample belonging to each bearing fault category predicted by the classifier B, and the difference between the two and the absolute value can reflect the difference in the prediction of the two classifiers for the target domain loom state sample.

4. The method for recognizing bearing fault of heterogeneous working condition loom shaft based on adversarial domain discrimination according to claim 1 or 3, characterized in that, In step 513, With By the following formula: In the above formula, n s represents the number of source domain data samples, represents the number of bearing fault classes of the source domain data, i represents the i-th data sample, and j represents the number of bearing fault classes, represents the predicted probability value of the classifier A or B that the i-th source domain data sample belongs to the bearing fault class j, represents the predicted probability value of the classifier A or B that the i-th source domain data sample belongs to the bearing fault class j, is 1, otherwise 0.

5. The method of claim 1, wherein, In step 521, the difference loss is minimized to update the loom bearing feature extraction module, classifier A and classifier B, class level weights The calculation of the public announcement is as follows: In the above formula, n t represents the number of target domain data samples, i represents the i-th data sample, represents the probability value vector of the i-th target domain data sample belonging to each bearing fault category predicted by the classifier A, represents the probability value vector of the i-th target domain data sample belonging to each bearing fault category predicted by the classifier B, represents the element with the maximum value in the vector, and the weight is normalized by formula (6).

6. The method for recognizing bearing fault of heterogeneous working condition loom shaft based on adversarial domain discrimination according to claim 1 or 5, characterized in that, In step 521, the class-level weight is applied Constructing a new class-level weighted classification cross-entropy loss In formula (7), n s represents the number of source domain data samples, represents the number of bearing fault categories of the source domain data, i represents the i-th data sample, and j represents the number of bearing fault categories, represents the predicted probability value of the i-th source domain data sample belonging to the bearing fault category j by the classifier A, represents the predicted probability value of the i-th source domain data sample belonging to the bearing fault category j by the classifier B, represents 1 when , otherwise 0, represents the weight value of the element corresponding to the j bearing fault category dimension in the weight vector.

7. The method for recognizing bearing fault of heterogeneous working condition loom shaft based on adversarial domain discrimination according to claim 1 or 5, characterized in that, In step 522, the working condition discrimination loss is constructed Where, n s and n t Let represent the number of data samples in the source domain and the target domain, respectively, where i represents the i-th data sample, and k represents the number of data samples in the target domain. i The domain label truth value represents the principal property of the i-th sample loom state extracted. When the principal property is extracted from the features of the data samples in the source domain, k i The value of k is 1; otherwise, k i The value is 0. The probability value that the principal property of the i-th sample, predicted by the working condition discriminator, belongs to the source domain is represented by the value of the working condition discriminator. The range is 0 to 1; Construction Loss function: wherein n s represents the number of source domain data samples, i denotes the i-th data sample, j represents the number of bearing fault classes, represents the number of bearing fault classes of the source domain data, represents the predicted probability value of the i-th source domain data sample belonging to the bearing fault class j with respect to the feature principal property of the fault classifier, represents 1 when 0 otherwise.

8. The method for recognizing bearing fault of heterogeneous working condition loom shaft based on adversarial domain discrimination according to claim 1 or 5, characterized in that, In step 523, a regression loss function is constructed that maps the sample features to their corresponding principal properties wherein, n s and n t respectively represent the number of source domain and target domain data samples, h i represents the principal property of the i-th sample mapped by the i-th sample feature in the heterogeneous loom bearing feature weighting module, represents the principal property of the i-th sample obtained by the principal property separator, the second term of formula (10) is the l1 norm regularization, γ represents the penalty coefficient of the l1 regularization term, the value is set to 0.5, c e represents the e-th parameter of the first layer of the heterogeneous loom bearing feature weighting module, and the first layer has E parameters in total; The gradient values obtained Taking the absolute value and performing an l2 normalization operation, and element-level multiplication with the sample features: wherein N represents a l2normalization operation, represents an element-wise multiplication operation, X represents a feature of a sample, and Z represents a resulting weighted feature; The weighted features of the source domain and the target domain are respectively averaged by column, and are transmitted into the heterogeneous weighted feature alignment module to construct a weighted feature alignment loss function between the source domain and the target domain where mean() represents the average value by column, Z t represents the weighted features of the target domain, Z s represents the weighted features of the source domain, and the square of the l2 norm between the source domain and target domain features after averaging is calculated, and the parameters of the loom bearing feature extraction module are updated according to the alignment loss function .

Images