A multi-noise robust cross-condition transfer fault diagnosis method under time-varying conditions

Abstract

The present application belongs to the technical field of intelligent fault diagnosis of mechanical equipment, and discloses a multi-noise robust cross-condition migration fault diagnosis method under time-varying conditions, comprising the following steps: acquiring signal samples and preprocessing; establishing a correction network module and a domain adaptation network module, the correction network is used for screening clean samples and re-labeling noise of suppressing feature noise, label noise and mixed feature noise and label noise, and the domain adaptation network module is used for feature clustering and cross-condition migration diagnosis; a multi-noise robust fault diagnosis training model is established; model training is carried out according to a loss function and an optimization algorithm; online fault diagnosis, the present application solves the problem of low fault recognition accuracy caused by multiple noise interference under time-varying conditions.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent fault diagnosis of mechanical equipment, and particularly relates to a multi-noise robust cross-condition migration fault diagnosis method under time-varying conditions. BACKGROUND

[0002] With the development and progress of industry, mechanical equipment is widely used in various industries, and the demand for intelligent operation and maintenance of major equipment is increasing. Intelligent operation and maintenance is of great significance to the operation safety and function execution of mechanical equipment. However, the working environment of mechanical equipment is harsh, and the background noise is strong. The vibration signals measured by sensors contain a lot of characteristic noise and fault-independent components. At the same time, mechanical equipment works in high-speed, impact and other conditions, which is a typical non-stationary condition. Unstable excitation, complex working conditions, mechanical vibration, electrical interference, and pneumatic flow may all produce a large amount of characteristic noise. In recent years, with the development of data-driven technologies such as machine learning and deep learning, many researchers have begun to use data-driven methods for fault diagnosis research. However, the performance of fault diagnosis is largely determined by the quality of the training data. In practical applications, the fault data categories are marked by artificial, and the complex marking process inevitably contains some errors, resulting in a large amount of noise labels in the training samples, which greatly reduces the performance of fault diagnosis. In the presence of feature noise, label noise, or a mixture of feature noise and label noise, existing methods have the problem of low fault recognition accuracy in actual engineering applications. SUMMARY

[0003] In view of the above defects or improvement needs of the prior art, the present application provides a multi-noise robust cross-condition migration fault diagnosis method under time-varying conditions, which solves the problem of low fault recognition accuracy caused by the presence of noise when training network models in source domains containing multiple noises.

[0004] To achieve the above-mentioned purpose, according to one aspect of the present application, a multi-noise robust cross-condition migration fault diagnosis method under time-varying conditions is provided, comprising the following steps:

[0005] Offline training:

[0006] S1: Collect sample data and perform preprocessing to obtain source domain training data set and target domain training data set;

[0007] S2: Establish a correction network module and a domain adaptation network module. The correction network module takes the source domain training data obtained in step S1 as input to obtain clean source domain data. The domain adaptation network module takes the clean source domain data and the target domain training data as input to realize domain knowledge migration and distribution alignment;

[0008] S3: combining the correction network module and the domain adaptation network module established in step S2 to establish a multi-noise robust fault diagnosis training model;

[0009] S4: training according to a loss function and an optimization algorithm, inputting source domain training data and target domain training data into the fault diagnosis training model constructed in step S3 to perform model training, and obtaining a fault diagnosis test model;

[0010] S5 model testing: inputting samples of target domain test data into the fault diagnosis test model constructed in S4 in sequence, and outputting the fault categories of the samples;

[0011] Online fault diagnosis:

[0012] Inputting target domain data requiring fault diagnosis into the fault diagnosis model obtained in S5 to diagnose the fault type corresponding to the target domain data.

[0013] As a preferred, the source domain data in step S1 includes vibration, acceleration, current or acoustic emission signal data;

[0014] The preprocessing method is: segmenting sample data collected under different working conditions, unifying the sample data length, and normalizing the sample data to obtain a source domain training data set and a target domain training data set; wherein the source domain training data set is labeled and used for model training; the target domain training data is unlabeled and used for domain adaptation training of the model.

[0015] As a preferred, the correction network module in step S2 includes a feature extractor, a correction network classifier, and a weight generator; the domain adaptation network module includes a feature extractor, a domain adaptation classifier, and a domain discriminator;

[0016] The feature extractor includes but is not limited to constructing feature extraction through one of a fully connected network, a deep convolutional network, a deep belief network, a deep residual network, and a deep auto-encoding network, for extracting hidden fault feature information in the signal;

[0017] The correction network classifier and the domain adaptation classifier are both composed of a fully connected neural network and a Softmax classifier, taking the features output by the feature extractor as input, and outputting the fault categories of the samples;

[0018] The weight generator is composed of a fully connected neural network, taking the features output by the feature extractor as input, and outputting the weight of each sample;

[0019] The domain discriminator is composed of a fully connected neural network, taking the features output by the feature extractor as input, and outputting whether the sample is from the source domain or the target domain.

[0020] As a preferred, step S2 specifically includes the following steps:

[0021] S2.1 input the noisy source domain training data into the correction network module, extract deep features through the correction network feature extractor, and input the deep features into the correction network classifier and weight generator, use the classification loss optimization of the correction network classifier to optimize the correction network feature extractor, classifier and weight extractor, suppress the noise samples in the source domain, and finally obtain clean source domain data and the weight of the sample;

[0022] S2.2 input the clean source domain data obtained in step S2.1 and the unlabeled target domain training data into the domain adaptation network, extract deep features of the source domain and target domain data through the feature extractor, and input the deep features into the domain adaptation classifier and domain discriminator, use the classification loss weighted by the domain adaptation network classifier, the contrast loss, and the weighted domain difference measurement loss to optimize the feature extractor and the domain discriminator, realize the migration and distribution alignment of the domain knowledge.

[0023] As a preferred, step S3 specifically comprises the following steps:

[0024] S3.1 in the correction network module, for the source domain noise input sample x s , the feature extractor φ outputs the extracted feature φ(x s ), the predicted class distribution μ s and the uncertainty weight s

[0025] w of the sample are obtained by calculation through the following formula:

[0026]

[0027] Wherein, C and V represent the correction network classifier and the weight generator;

[0028] The training of the feature extractor, the correction network classifier and the weight generator in the correction network is through the weighted classification loss:

[0029]

[0030] Wherein, is the uncertainty weighted classification loss function, ||·||1 represents the L1 norm, represents the possible contaminated class label of the i-th sample ;

[0031] The predicted uncertainty weight w i is through 1 / w i 2 The weight of the noise sample is effectively suppressed and the classification loss is reduced, so that the model after training has stronger robustness to the input data with noise, and the influence of the noise sample with large uncertainty weight on the training loss is reduced;

[0032] S3.2 Randomly shuffle the labels of training samples to obtain noisy samples The weight generator is trained by the noisy samples with shuffled labels to generate new uncertainty weights Assume the current parameters of the weight classifier V are θ, for each randomly shuffled batch θ is updated by a gradient descent to The process is represented as:

[0033]

[0034] where, represents the uncertainty weighted classification loss function;

[0035] The uncertainty ranking loss is represented as:

[0036]

[0037] where, δ represents the margin;

[0038] Assume and represent the target probability and model probability output, for sample i in the kth training, the target probability output in the k+1th is calculated as:

[0039]

[0040] where, β represents the momentum;

[0041] Using the estimated target probability output, for a training batch of B samples, the early regularization term is represented as:

[0042]

[0043] where, <·,·> represents the inner product operation;

[0044] The total training loss is:

[0045]

[0046] After the correction network training is completed, the noise source domain is relabeled, and the sample is relabeled by selecting the maximum predicted probability to achieve the purpose of removing the noise label, and relatively clean training data is provided for training the adaptive network;

[0047] S3.3 In the robust domain adaptation module, the clean training data obtained in step S2.1 and the target domain training data are used as input to train the model, and the source domain samples used for training are selected according to the aggregation of the classification loss and the uncertainty weight, and the average aggregation of the samples is represented as:

[0048]

[0049] wherein M represents the number of training times, represents the classification loss, w(x i represents the prediction uncertainty weight of the i-th sample; the values of the average aggregation in ascending order by class, the (n c ×p)th sample is selected as the threshold of the c-th fault, n c is the number of fault samples, and p is the selection ratio;

[0050] Through the above sample selection strategy, a multi-noise robust fault diagnosis training model is established, and cleaner samples are used to train the robust domain adaptation module.

[0051] As preferred, the loss function in step S4 is:

[0052] (1) Taking the features of the same class samples as positive samples and the features of different class samples as negative samples, a contrastive learning loss is constructed to make the same class samples gather to the same clustering center and the features of different class samples move away from each other, so as to separate the overlapping features, and the constructed contrastive loss function is:

[0053]

[0054] wherein N represents the number of fault samples, S(·) represents the similarity measure, cosine similarity, Euclidean distance, etc., z i represents the i-th sample feature, represents the positive sample feature, represents the negative sample feature, and τ represents the proportion parameter;

[0055] Through the contrastive loss, the features of the same class samples are gathered, and the features of different class samples are moved away, so as to overcome the problem of difficult separation of different fault features caused by time-varying working conditions;

[0056] (2) The domain difference measure is used to make the network learn a discriminant model using clean data of the source domain and data of the target domain, and the total loss is represented as:

[0057]

[0058] wherein, represents the difference measure, ψ(x i ) represents the domain adaptation network feature extractor, f(·,·) represents the domain adaptation network classifier, and f' represents the domain adaptation network domain discriminator.

[0059] As preferred, the domain difference measure is constructed by one of the following, including but not limited to, maximum mean discrepancy, multi-kernel maximum mean discrepancy measure, correlation alignment measure and maximum margin discrepancy measure, to reduce the domain shift caused by different working conditions and improve the diagnostic accuracy in the cross-working condition process.

[0060] As preferred, when the domain difference measure adopts the maximum margin discrepancy measure, the measure loss of the maximum discrepancy measure is expressed as:

[0061]

[0062] wherein, disp (ρ) (f',f) represents the marginal discrepancy, defined as:

[0063]

[0064] wherein, ρ f′ is the margin assumed by the domain adaptation network domain discriminator f', h f (x) is the label function of the classifier f, Φ ρ is the ρ margin loss;

[0065] Minimizing MDD is an adversarial process, expressed as:

[0066]

[0067] Combining the robust domain adaptation module weighted classification loss and the contrast loss, the whole optimization process of the robust domain adaptation is expressed as:

[0068]

[0069] wherein, represents the expectation, σ represents the Softmax function, γ=exp(ρ) is used to protect the margin ρ, λ i represents the weight, defined according to the average aggregation as:

[0070]

[0071] wherein, represents the average aggregation of the sample x i , represents the maximum value of the average aggregation of the sample x i in k iterations, represents the minimum value of the average aggregation of the sample x i in k iterations.

[0072] As preferred, the optimization algorithm includes but is not limited to one of the following, including but not limited to, stochastic gradient descent algorithm, root mean square propagation algorithm, adaptive moment estimation algorithm.

[0073] As preferred, the step S5 is specifically: the fault diagnosis test model is composed of the trained feature extractor and the classifier, the fault samples in the target domain test data only verify the accuracy of the output prediction label of the model, if the accuracy reaches the preset standard, the output prediction label corresponds to the category of each fault sample, and the fault diagnosis model is obtained; otherwise, the training is continued according to the loss function and the optimization algorithm.

[0074] Overall, compared with the prior art, the cross-condition migration fault diagnosis method robust to multiple noises under time-varying conditions provided by the application mainly has the following advantages

[0075] Beneficial effects:

[0076] 1. The cross-condition migration fault diagnosis method robust to multiple noises under time-varying conditions has a high diagnostic accuracy under time-varying multiple noise conditions.

[0077] 2. The cross-condition migration fault diagnosis method robust to multiple noises under time-varying conditions does not need complex preprocessing steps for data, and has good robustness to time-varying working conditions, label noise, feature noise, and mixed noise of label noise and feature noise. BRIEF DESCRIPTION OF DRAWINGS

[0078] Figure 1 is a training flowchart of the cross-condition migration fault diagnosis method robust to multiple noises under time-varying conditions provided by the application.

[0079] Figure 2 is a structural block diagram of the cross-condition migration fault diagnosis method robust to multiple noises under time-varying conditions provided by the application. DETAILED DESCRIPTION

[0080] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the application and do not limit the application. In addition, the technical features involved in each embodiment of the application described below can be combined with each other as long as they do not conflict with each other.

[0081] Please refer to Figure 1 and Figure 2 , the application provides a cross-condition migration fault diagnosis method robust to multiple noises under time-varying conditions. The fault diagnosis problem under multiple noise cross-condition conditions refers to that the model is inconsistent in the test phase and the training phase data set corresponding to the working condition (including load, speed and other working condition changes), and the source domain contains feature noise, label noise and mixed noise samples of feature noise and label noise.

[0082] The fault diagnosis method mainly comprises the following steps:

[0083] S1: Collecting sample data and preprocessing to obtain source domain training data set and target domain training data set;

[0084] The source domain data in step S1 includes vibration, acceleration, current or acoustic emission signal data;

[0085] The preprocessing method is to segment the sample data collected under different working conditions, unify the sample data length, and normalize the sample data to obtain the source domain training data set and the target domain training data set; wherein the source domain training data set is labeled and used for model training; the target domain training data is unlabeled and used for domain adaptation training of the model.

[0086] S2: Establishing a correction network module and a domain adaptation network module, the correction network module taking the source domain training data obtained in step S1 as input to obtain clean source domain data; the domain adaptation network module taking the clean source domain data and the target domain training data as input to realize domain knowledge migration and distribution alignment;

[0087] The correction network module in step S2 includes a feature extractor, a correction network classifier and a weight generator; the domain adaptation network module includes a feature extractor, a domain adaptation classifier and a domain discriminator;

[0088] The correction network module takes the source domain noise sample as input, mainly used to suppress various feature noise, label noise and mixed noise of feature noise and label noise in the source domain sample, so as to obtain clean source domain training sample; the adaptive network module takes the clean source domain noise and the unlabeled target domain sample as input, used to migrate the knowledge in the source domain to the target domain, realize cross-condition distribution alignment, and thus improve the cross-domain diagnosis performance in the noise scene.

[0089] The feature extractor includes but is not limited to constructing feature extraction by one of full connection network, deep convolution network, deep belief network, deep residual network and deep auto-encoding network, used to extract hidden fault feature information in the signal;

[0090] The correction network classifier and the domain adaptation classifier are both composed of full connection neural network and Softmax classifier, taking the feature output by the feature extractor as input, and outputting the fault category of the sample;

[0091] The weight generator is composed of full connection neural network, taking the feature output by the feature extractor as input, and outputting the weight of each sample;

[0092] The domain discriminator is composed of full connection neural network, taking the feature output by the feature extractor as input, and outputting whether the sample comes from the source domain or the target domain.

[0093] Step S2 specifically comprises the following steps:

[0094] S2.1 inputting the noisy source domain training data into the correction network module, extracting deep features through the correction network feature extractor, and inputting the deep features into the correction network classifier and the weight generator, using the classification loss optimization of the correction network classifier, the regularization loss and the ranking loss to optimize the correction network feature extractor, the classifier and the weight extractor, suppressing the noisy samples in the source domain, and finally obtaining clean source domain data and the weight of the sample;

[0095] S2.2 inputting the clean source domain data obtained in step S2.1 and the unlabeled target domain training data into the domain adaptation network, extracting deep features of the source domain and target domain data through the feature extractor, and inputting the deep features into the domain adaptation classifier and the domain discriminator, using the weighted classification loss of the domain adaptation network classifier, the contrast loss and the weighted domain difference measurement loss to optimize the feature extractor and the domain discriminator, realizing the migration and distribution alignment of domain knowledge.

[0096] S3: combining the correction network module and the domain adaptation network module established in step S2 to establish a multi-noise robust fault diagnosis training model;

[0097] S4: training according to the loss function and the optimization algorithm, inputting the source domain training data and the target domain training data into the fault diagnosis training model constructed in step S3 for model training, and obtaining a fault diagnosis test model;

[0098] In the correction network module, for the source domain noisy input sample x s , the feature extractor φ outputs the extracted feature φ(x s ), the predicted class distribution μ s and the uncertainty weight w s of the sample are obtained by calculation as follows:

[0099]

[0100] Wherein, C and V represent the correction network classifier and the weight generator,

[0101] The training of the feature extractor, the correction network classifier and the weight generator in the correction network is through the weighted classification loss:

[0102]

[0103] Wherein, is the uncertainty weighted classification loss function, represents the possible contaminated class label of the i-th sample .

[0104] During the early training process, the samples with noise show larger classification loss and larger uncertainty weight, and the neural network will fit the samples without noise before fitting the samples with label noise or feature noise in the process of minimizing the loss training. Therefore, the uncertainty weight w i By 1 / w i 2 Effectively suppress the weight of noisy samples and reduce the classification loss, so that the trained model has stronger robustness to noisy input data, and reduces the influence of noisy samples with uncertainty on training loss.

[0105] To further eliminate the influence of label noise on the model, ranking regularization is used to associate noisy label samples and large uncertainty weights. Specifically, for each training batch of samples, randomly shuffle their labels to obtain noisy samples The weight generator is trained by the noisy samples with shuffled labels to generate new uncertainty weights Compared with the uncertainty weight w generated by the original data, because of the new label Y contains more mislabeled (noise) than Y, so the uncertainty weight is larger. In the case, the larger uncertainty weight is used to suppress the label noise, thereby further reducing the influence of the label noise sample on the model.

[0106] Assuming that the current parameters of the weight classifier V are θ, for each randomly shuffled batch θ is updated to The process is represented as:

[0107]

[0108] where, represents the uncertainty weighted classification loss function, α represents the learning rate, represents the gradient, X represents the sample, represents the noise label;

[0109] The uncertainty ranking loss is represented as:

[0110]

[0111] where δ represents the margin, which is set to 0.2.

[0112] In order to prevent overfitting and remember the wrong labels, based on the obtained model to estimate the target probability, an early regularization term is used to obtain the true target probability.

[0113] Assuming and The target probability and the probability output of the model are represented. For the sample i in the kth training, the target probability output of the k+1th training is calculated as:

[0114]

[0115] where β represents momentum, and is set to 0.9.

[0116] Using the estimated target probability output, for a training batch of B samples, the early regularization term is represented as:

[0117]

[0118] where <·,·> represents the inner product operation.

[0119] Because the deep learning model tends to fit clean samples in the early training process, the estimated target effectively obtains the clean label probability, and the purpose of removing the noisy label is achieved by re-labeling the sample by selecting the maximum prediction probability. Early regularization forces the output probability generated by the network to be more similar to the target probability, preventing learning and memorizing the wrong label.

[0120] The total training loss is written as:

[0121]

[0122] After the correction network training is completed, the noise source domain is re-labeled, and the purpose of removing the noisy label is achieved by re-labeling the sample by selecting the maximum prediction probability, which provides relatively clean training data for training the adaptive network.

[0123] In the robust domain adaptation module, clean source domain data and target domain training samples are used as input to train the model. Specifically, in order to further improve the robustness of the model to noise, source domain samples used for training are selected according to the aggregation of classification loss and uncertainty weight, and the average aggregation of samples is represented as:

[0124]

[0125] In the formula, M represents the number of training, represents the classification loss, and w(x i represents the prediction uncertainty weight of the i th sample. The values of the average aggregation are arranged in ascending order by class, and the (n c ×p) th sample is selected as the threshold of the c th class fault, n c is the number of samples of the class fault, and p is the selection proportion.

[0126] Through the above sample selection strategy, a multi-noise robust fault diagnosis training model is established, and cleaner samples are used to train the robust domain adaptation module, further improving the noise robustness of the model.

[0127] In order to overcome the problem of overlapping and difficult separation of different fault features caused by time-varying working conditions, the same sample features are taken as positive samples, and different sample features are taken as negative samples to construct a contrast learning loss, so that the same sample features are gathered to the same clustering center, and the different sample features are far away from each other, so as to realize the separation of overlapping features, and the constructed contrast loss function is:

[0128]

[0129] Wherein, N represents the number of fault samples, S(·) represents the similarity measure, cosine similarity, Euclidean distance, etc. are adopted, z i represents the i-th sample feature, represents the positive sample feature, represents the negative sample feature, and τ represents the proportion parameter. Through the contrast loss, the same sample features are gathered, and the different sample features are far away from each other, overcoming the problem of overlapping and difficult separation of different fault features caused by time-varying working conditions.

[0130] In order to reduce the domain drift problem caused by different working conditions, the domain difference measure is used to reduce the domain drift and domain difference. The domain difference measure adopted includes but is not limited to one of maximum mean difference (MMD), multi-kernel maximum mean difference measure (MK-MMD), correlation alignment measure (CORAL) and maximum marginal difference measure (MDD) to construct, which is used to reduce the domain drift and reduce the domain drift problem caused by different working conditions, and improve the diagnosis accuracy in the process of cross-working condition. Specifically, by using a clean source domain, the classification loss of the source domain is minimized, and the difference measure promotes the network to learn a discriminant model using the clean data of the source domain and the target domain data. The model is well generalized to the target domain. The total loss is represented as:

[0131]

[0132] Wherein, represents the difference measure, ψ(x i ) represents the domain adaptation network feature extractor, f(·,·) represents the domain adaptation network classifier, and f' represents the domain adaptation network domain discriminator. The following is an example of maximum marginal difference measure (MDD), and the measurement loss of maximum difference measure is represented as:

[0133]

[0134] Wherein, disp (ρ) (f′,f) represents the marginal difference, which is defined as:

[0135]

[0136] where ρ f′ is the margin assumed by the domain adaptation network domain discriminator f', h f (x) is the labeling function of the classifier f, Φ ρ is the ρ margin loss.

[0137] By minimizing the classification loss of the source domain, the MDD discrepancy measure encourages the network to learn a discriminative model that generalizes well to the target domain using both the clean data of the source domain and the target domain data. Minimizing MDD is an adversarial process, denoted as:

[0138]

[0139] Combining the robust domain adaptation module weighted classification loss and the contrastive loss, the entire optimization process of the robust domain adaptation is represented as:

[0140]

[0141] where, denotes expectation, σ denotes the Softmax function, γ=exp(ρ) is used to protect the margin ρ, λ i denotes weight, and is defined according to the average aggregation as:

[0142]

[0143] The optimization algorithm includes but is not limited to one of a stochastic gradient descent algorithm, a root mean square propagation algorithm, and an adaptive moment estimation algorithm.

[0144] S5 model testing: input the sample of the target domain test data into the fault diagnosis test model constructed by S4 in turn, and output the fault category of the sample;

[0145] Specifically, the fault diagnosis test model is composed of the trained feature extractor ψ and the classifier f, and the labels of the fault samples in the target domain test data are only used to verify the accuracy of the predicted labels output by the model. If the accuracy reaches the preset standard, the class of each fault sample corresponding to the predicted label is output, and the fault diagnosis model is obtained. Otherwise, the training is continued according to the loss function and the optimization algorithm.

[0146] The fault diagnosis method uses the designed training process to realize the cross-condition fault feature distribution migration under a multi-noise scene, and can effectively improve the accuracy of fault diagnosis and improve the level of intelligent fault diagnosis of equipment.

[0147] Online fault diagnosis: input the target domain data that needs to be diagnosed into the fault diagnosis model obtained by S5 to diagnose the fault type corresponding to the target domain data.

[0148] It is to be understood that the above description is intended to be illustrative and not restrictive. Many other embodiments will be apparent to those of skill in the art upon reading the above description. The scope of the application should therefore, be determined not with reference to the above description, but should instead be determined with reference to the appended claims, along with their full scope of equivalents. The disclosure of all articles and references referred to herein are incorporated by reference in their entirety.

Claims

1. A multi-noise robust cross-condition transfer fault diagnosis method under time-varying conditions, characterized in that, The method comprises the following steps: Offline training: S1: Collect sample data and pre-process to obtain source domain training data set and target domain training data set; the source domain training data set comprises vibration, acceleration, current or acoustic emission signal data; wherein the source domain training data set is labeled and used for model training; the target domain training data set is unlabeled and used for domain adaptation training of the model; S2: Establish a correction network module and a domain adaptation network module; the correction network module takes the source domain training data obtained in step S1 as input to obtain clean source domain data; the domain adaptation network module takes the clean source domain data and the target domain training data as input to realize domain knowledge transfer and distribution alignment; Step S2 specifically comprises the following steps: S2.1: input the noisy source domain training data into the correction network module, extract deep features through the correction network feature extractor, and input the deep features into the correction network classifier and the weight generator, use the classification loss optimization, regularization loss and ranking loss optimization of the correction network classifier to optimize the correction network feature extractor, classifier and weight extractor, suppress the noise samples in the source domain, and finally obtain clean source domain data and sample weights; S2.2: input the clean source domain data obtained in step S2.1 and the unlabeled target domain training data into the domain adaptation network, extract deep features of the source domain and target domain data through the feature extractor, and input the deep features into the domain adaptation classifier and the domain discriminator, use the weighted classification loss, contrast loss and weighted domain difference measurement loss of the domain adaptation network classifier to optimize the feature extractor and domain discriminator, realize domain knowledge transfer and distribution alignment; S3: combine the correction network module and the domain adaptation network module established in step S2 to establish a multi-noise robust fault diagnosis training model; S4: train according to the loss function and the optimization algorithm, input the source domain training data and the target domain training data into the fault diagnosis training model constructed in step S3 for model training, and obtain a fault diagnosis test model; S5 model test: input the samples of the target domain test data into the fault diagnosis test model constructed in S4 in sequence, and output the fault categories of the samples; Online fault diagnosis: input the target domain data that needs to be diagnosed into the fault diagnosis model obtained in S5 to diagnose the fault type corresponding to the target domain data.

2. The multi-noise robust cross-condition migration fault diagnosis method under time-varying working conditions according to claim 1, characterized in that, The pre-processing method in step S1 is to segment the sample data collected under different working conditions, unify the sample data length, and normalize the sample data to obtain the source domain training data set and the target domain training data set.

3. The multi-noise robust cross-condition migration fault diagnosis method under time-varying working conditions according to claim 1, characterized in that, The correction network module in step S2 comprises a feature extractor, a correction network classifier and a weight generator; the domain adaptation network module comprises a feature extractor, a domain adaptation classifier and a domain discriminator; The feature extractor comprises but is not limited to one of a fully connected network, a deep convolutional network, a deep belief network, a deep residual network and a deep auto-encoding network for constructing feature extraction, and is used to extract hidden fault feature information in the signal; The correction network classifier and the domain adaptation classifier are both composed of a fully connected neural network and a Softmax classifier, taking the features output by the feature extractor as input, and outputting the fault category of the sample; The weight generator is composed of a fully connected neural network, taking the features output by the feature extractor as input, and outputting the weight of each sample; The domain discriminator is composed of a fully connected neural network, taking the features output by the feature extractor as input, and outputting whether the sample is from the source domain or the target domain.

4. The multi-noise robust cross-condition migration fault diagnosis method under time-varying working conditions according to claim 1, characterized in that, Step S3 specifically includes the following steps: S3.1 In the correction network module, for the source domain noisy input sample , a feature extractor outputs extracted features , a prediction class distribution is obtained by calculating and an uncertainty weight of the sample : wherein, and denote the correction network classifier and weight generator, respectively. The training of the feature extractor, the correction network classifier and the weight generator in the correction network is through a weighted classification loss: wherein, is an uncertainty-weighted classification loss function, denotes an L1 norm, denotes the i th sample may be contaminated class labels; Predicted uncertainty weight By The weight of the noise sample is effectively inhibited and the classification loss is reduced, so that the model after training has stronger robustness to the input data with noise, and the influence of the noise sample with large uncertainty weight on the training loss is reduced. S3.2 Randomly shuffle the labels of the training samples to obtain noisy samples The weight generator is trained by the noisy samples with shuffled labels to generate new uncertainty weights ; Assume the current parameters of the weight classifier are For each randomly shuffled batch , is updated by a gradient descent to This process is denoted as: wherein, denotes an uncertainty-weighted classification loss function, denotes a learning rate, denotes a gradient, denotes a sample, denotes a noisy label; The uncertainty ranking loss is represented as: wherein represents a margin; Assume and represent the target probability and model probability outputs, for the k sample in the i , k +1th iteration of target probability output computation. wherein denotes momentum; With the estimated target probability output, for a training batch of B samples, the early regularization term is expressed as: wherein denotes an inner product operation; The total training loss is: After the correction network is trained, the noise source domain is relabeled, and the sample is relabeled by selecting the maximum prediction probability to achieve the purpose of removing the noise label, and relatively clean training data is provided for training the adaptive network; S3.3 In the robust domain adaptation module, the clean training data obtained in step S2.1 and the target domain training data are used as input to train the model, and the source domain samples for training are selected according to the aggregation of the classification loss and the uncertainty weight, and the average aggregation of the samples is represented as: In the formula, M denotes the number of training times, denotes the classification loss, denotes the prediction uncertainty weight of the i th sample; the values of the average aggregation in ascending order by class, the th sample as the threshold of the c th class fault to select the sample, is the number of samples of the class fault, is the selection ratio; Through the above sample selection strategy, a multi-noise robust fault diagnosis training model is established, and cleaner samples are used to train the robust domain adaptation module.

5. The multi-noise robust cross-condition migration fault diagnosis method under time-varying working conditions according to claim 1, characterized in that, The loss function in step S4 is: (1) Taking the features of the same class samples as positive samples and the features of different class samples as negative samples, a contrastive learning loss is constructed to make the same class samples gather to the same clustering center and the features of different class samples move away from each other, so as to separate the overlapping features, and the constructed contrastive loss function is: wherein, represents the number of samples indicating a fault, represents a similarity measure, either cosine similarity or Euclidean distance, represents the feature of the i th sample, represents the feature of the positive sample, represents the feature of the negative sample, represents a proportionality parameter; Through the contrastive loss, the features of the same class samples are gathered, and the features of the different class samples are moved away, overcoming the problem of difficult separation of different fault features caused by time-varying working conditions; (2) The domain difference measurement promotes the network to learn a discriminant model using clean data of the source domain and data of the target domain, and the total loss is represented as: wherein, denotes a difference measure, denotes a domain adaptation network feature extractor, denotes a domain adaptation network classifier, denotes a domain adaptation network domain discriminator.

6. The multi-noise robust cross-condition migration fault diagnosis method in time-varying working conditions according to claim 5, characterized in that, The domain difference measurement includes but is not limited to one of maximum mean difference, multi-kernel maximum mean difference measurement, correlation alignment measurement and maximum margin difference measurement, which is constructed to reduce domain drift caused by different working condition changes and improve the diagnosis accuracy in the cross-working condition process.

7. The multi-noise robust cross-condition migration fault diagnosis method in time-varying working conditions according to claim 6, characterized in that, When the domain difference measurement adopts the maximum margin difference measurement, the measurement loss of the maximum difference measurement is represented as: wherein represents the marginal difference, defined as: wherein, is a domain adaptation network domain discriminator assumed margin, is a classifier a labeling function, is marginal loss; Minimizing MDD is an antagonistic process, represented as: Combining the weighted classification loss and the contrastive loss of the robust domain adaptation module, the whole optimization process of the robust domain adaptation is represented as: wherein, denotes the expectation, denotes the Softmax function, for protecting the margin , denotes the weight, defined as the average aggregation: wherein, denotes the average aggregation of the samples , denotes the average aggregation of the samples in k the minimum value of the average aggregation of the samples in the maximum value of the average aggregation of the samples k .

8. The multi-noise robust cross-condition migration fault diagnosis method in time-varying working conditions according to Claim 1, characterized in that, The optimization algorithm includes but is not limited to one of the stochastic gradient descent algorithm, the root mean square propagation algorithm and the adaptive moment estimation algorithm.

9. The multi-noise robust cross-condition migration fault diagnosis method in time-varying working conditions according to Claim 1, characterized in that, The step S5 specifically comprises: the fault diagnosis test model is composed of the trained feature extractor and the classifier , the fault samples in the target domain test data only verify the accuracy of the predicted labels output by the model, if the accuracy reaches a preset standard, the predicted labels output by the model correspond to the categories of the fault samples, and a fault diagnosis model is obtained; otherwise, the training is continued according to a loss function and an optimization algorithm.

Images