Multi-noise robust cross-working-condition migration fault diagnosis method under time-varying working condition

Abstract

The invention belongs to the technical field of intelligent fault diagnosis of mechanical equipment, and discloses a multi-noise robust cross-working-condition migration fault diagnosis method under a time-varying working condition, and the method comprises the following steps: obtaining a signal sample, and carrying out the preprocessing of the signal sample; a correction network module and a domain adaptation network module are established, the correction network module is used for screening clean samples and re-marking and inhibiting feature noise, label noise and mixed noise of the feature noise and the label noise, and the domain adaptation network module is used for feature clustering and cross-working-condition migration diagnosis; establishing a multi-noise robust fault diagnosis training model; performing model training according to the loss function and an optimization algorithm; according to the method, the problem of low fault identification accuracy caused by various noise interferences under the time-varying working condition is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent fault diagnosis of mechanical equipment, and in particular to a multi-noise robust cross-operating condition migration fault diagnosis method under time-varying operating conditions. Background Art

[0002] With the development and progress of industry, mechanical equipment is widely used in various industries. 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 signal measured by the sensor contains more characteristic noise and fault-irrelevant components. At the same time, mechanical equipment works under high speed, impact and other working conditions, which is a typical non-stationary working condition. Unstable excitation and complex working conditions, mechanical vibration, electrical interference, pneumatic flow and other factors may generate 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 to study fault diagnosis, but the performance of fault diagnosis is largely determined by the quality of training data. In practical applications, the classification of fault data depends on manual labeling. The complicated labeling process inevitably leads to some errors, resulting in a large number of noise labels in the training samples. These noise labels greatly reduce the performance of fault diagnosis. In the face of scenarios with multiple noises such as characteristic noise, label noise or even a mixture of characteristic noise and label noise, the existing methods have the problem of low fault recognition accuracy in practical engineering applications. Summary of the invention

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

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

[0005] Offline training:

[0006] S1: Collect sample data and preprocess them 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 uses the source domain training data obtained in step S1 as input to obtain clean source domain data; the domain adaptation network module uses the clean source domain data and target domain training data as input to achieve 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: Perform training 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 to perform model training, and obtain a fault diagnosis test model;

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

[0011] Online troubleshooting:

[0012] The target domain data that needs to be diagnosed is input into the fault diagnosis model obtained in S5 to diagnose the fault type corresponding to the target domain data.

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

[0014] The preprocessing method is: segmenting the 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] Preferably, in step S2, the correction network module 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 autoencoder network, so as to extract fault feature information hidden 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 category of the sample;

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

[0019] The domain discriminator is composed of a fully connected neural network, which takes the features output by the feature extractor as input, and outputs samples from the source domain or the target domain.

[0020] Preferably, 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, regularization loss and sorting loss 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;

[0022] S2.2 Input the clean source domain data and unlabeled target domain training data obtained in step S2.1 into the domain adaptation network, extract the 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, and use the weighted classification loss, contrast loss, and weighted domain difference metric loss of the domain adaptation network classifier to optimize the feature extractor and domain discriminator to achieve domain knowledge transfer and distribution alignment.

[0023] Preferably, step S3 specifically includes 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 category distribution μ is calculated by the following formula s and the uncertainty weight s of the sample

[0025] w:

[0026]

[0027] Among them, C and V represent the corrected network classifier and weight generator;

[0028] The training of the feature extractor, the rectified network classifier, and the weight generator in the rectified network is done with a weighted classification loss:

[0029]

[0030] in, is the uncertainty weighted classification loss function, ||·||1 represents the L1 norm, represents the i-th sample Class labels that may be contaminated;

[0031] The uncertainty weight of the prediction w i Through 1 / w i 2 Effectively suppress the weight of noise samples and reduce classification loss, making the trained model more robust to noisy input data and reducing the impact of noise samples with large uncertainty weights on training loss;

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

[0033]

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

[0035] The uncertainty ranking loss is expressed as:

[0036]

[0037] Among them, δ represents the margin;

[0038] Assumptions and represents the target probability and model probability output. For sample i in the kth training, the target probability output of k+1 is calculated as:

[0039]

[0040] Among them, β represents momentum;

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

[0042]

[0043] Among them, <·,·> 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 re-labeled. The purpose of removing the noise label is achieved by selecting the re-labeled sample with the maximum prediction probability, providing relatively clean training data for training the adaptive network;

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

[0048]

[0049] In the formula, 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 average aggregated values ​​are arranged in ascending order by category, and the (n-th c ×p) samples as the threshold of the c-th type of fault to select samples, n c is the number of samples of class fault, 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 will be used to train the robust domain adaptation module.

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

[0052] (1) The features of samples of the same type are taken as positive samples, and the features of samples of different types are taken as negative samples. Contrastive learning loss is constructed so that samples of the same type gather at the same cluster center and features of samples of different types move away from each other, thereby separating overlapping features. The constructed contrastive loss function is:

[0053]

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

[0055] Through contrast loss, sample features of the same type are clustered, and sample features of different types are pulled apart, overcoming the problem of overlapping and difficult separation of different fault features caused by time-varying working conditions;

[0056] (2) The domain difference metric used encourages the network to learn a discriminative model using clean data from the source domain and data from the target domain. The total loss is expressed as:

[0057]

[0058] in, 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] Preferably, the domain difference metric includes but is not limited to one of the maximum mean difference, multi-core maximum mean difference metric, correlation alignment metric and maximum marginal difference metric to reduce domain drift caused by changes in different working conditions and improve the diagnostic accuracy in cross-working condition processes.

[0060] Preferably, when the domain difference metric adopts the maximum marginal difference metric, the metric loss of the maximum difference metric is expressed as:

[0061]

[0062] Among them, disp (ρ) (f′,f) represents the marginal difference, which is defined as:

[0063]

[0064] Among them, ρ f′ is the assumed margin of the domain discriminator f′ of the domain adaptation network, h f (x) is the labeling function of the classifier f, Φ ρ is the ρ marginal loss;

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

[0066]

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

[0068]

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

[0070]

[0071] in, Represents sample x i The average aggregation of Represents sample x i The minimum value of the average aggregation over k iterations, Represents sample x i Average the aggregated maximum value over k iterations.

[0072] Preferably, the optimization algorithm includes but is not limited to one of a stochastic gradient descent algorithm, a root mean square transfer algorithm, and an adaptive moment estimation algorithm.

[0073] Preferably, step S5 is specifically as follows: the fault diagnosis test model is composed of a trained feature extractor ψ and a classifier f, and the fault samples in the target domain test data only verify the accuracy of the model output prediction labels. If the accuracy reaches the preset standard, the output prediction labels correspond to the categories of each fault sample to obtain the fault diagnosis model; otherwise, training continues according to the loss function and optimization algorithm.

[0074] In general, compared with the prior art, the multi-noise robust cross-operating condition migration fault diagnosis method under time-varying operating conditions provided by the present invention mainly has the following advantages:

[0075] Beneficial effects:

[0076] 1. The multi-noise robust cross-operating condition migration fault diagnosis method under time-varying operating conditions of the present invention has a high diagnostic accuracy under time-varying multi-noise conditions.

[0077] 2. The multi-noise robust cross-operating condition migration fault diagnosis method under time-varying operating conditions of the present invention does not require complex data preprocessing steps and can have good robustness to time-varying operating conditions, label noise, feature noise, and mixed noise of label noise and feature noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 It is a training flow chart of a multi-noise robust cross-operating condition migration fault diagnosis method under time-varying operating conditions provided by the present invention.

[0079] Figure 2 It is a structural block diagram of a multi-noise robust cross-operating condition migration fault diagnosis method under time-varying operating conditions provided by the present invention. DETAILED DESCRIPTION

[0080] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0081] See also Figure 1 and Figure 2 The present invention provides a multi-noise robust cross-condition migration fault diagnosis method under time-varying conditions. The fault diagnosis problem under multi-noise cross-condition conditions refers to the situation that the operating conditions (including changes in load, speed and other operating conditions) corresponding to the data sets in the test phase and the training phase of the model are inconsistent and the source domain contains characteristic noise, label noise and mixed noise samples of characteristic noise and label noise.

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

[0083] S1: Collect sample data and preprocess them to obtain source domain training data set and target domain training data set;

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

[0085] The preprocessing method is: segmenting the 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.

[0086] S2: Establish a correction network module and a domain adaptation network module. The correction network module uses the source domain training data obtained in step S1 as input to obtain clean source domain data; the domain adaptation network module uses the clean source domain data and target domain training data as input to achieve domain knowledge migration and distribution alignment;

[0087] In step S2, the correction network module 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 source domain noise samples as input, and is mainly used to suppress various feature noises, label noises, and noises mixed with feature noises and label noises in the source domain samples, so as to obtain clean source domain training samples; the adaptive network module takes clean source domain noise and unlabeled target domain samples as input, and is used to transfer knowledge in the source domain to the target domain, realize cross-condition distribution alignment, and thus improve cross-domain diagnostic performance in noise scenarios.

[0089] 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 autoencoder network, so as to extract fault feature information hidden in the signal;

[0090] 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;

[0091] The weight generator is composed of a fully connected neural network, which takes the features output by the feature extractor as input and outputs the weight of each sample;

[0092] The domain discriminator is composed of a fully connected neural network, which takes the features output by the feature extractor as input, and outputs samples from the source domain or the target domain.

[0093] Step S2 specifically includes the following steps:

[0094] 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, regularization loss and sorting loss 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;

[0095] S2.2 Input the clean source domain data and unlabeled target domain training data obtained in step S2.1 into the domain adaptation network, extract the 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, and use the weighted classification loss, contrast loss, and weighted domain difference metric loss of the domain adaptation network classifier to optimize the feature extractor and domain discriminator to achieve domain knowledge transfer and distribution alignment.

[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: Perform training 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 to perform model training, and obtain a fault diagnosis test model;

[0098] 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 ), predict the category distribution μ s and the uncertainty weight w of the sample s Obtained by the following calculation:

[0099]

[0100] Among them, C and V represent the corrected network classifier and weight generator,

[0101] The training of the feature extractor, the rectified network classifier, and the weight generator in the rectified network is done with a weighted classification loss:

[0102]

[0103] in, is the uncertainty weighted classification loss function, represents the i-th sample Class labels that may be contaminated.

[0104] In the early stages of training, samples with noise exhibit larger classification losses and larger uncertainty weights. During the process of minimizing the loss, the neural network will first fit samples without noise before fitting samples with label noise or feature noise. Therefore, the predicted uncertainty weight w i Through 1 / w i 2 It effectively suppresses the weight of noise samples and reduces classification loss, making the trained model more robust to noisy input data and reducing the impact of uncertain noise samples on training loss.

[0105] To further eliminate the impact of label noise on the model, ranking regularization is used to associate noisy label samples with large uncertainty weights. Specifically, for each training batch of samples, their labels are randomly shuffled to obtain noisy samples. The weight generator generates new uncertainty weights by training with noise samples that disrupt the labels. Compared with the uncertainty weight w generated by the original data, due to the new label contains more wrong labels (noise) than Y, so the uncertainty weight Bigger. In the case of , a larger uncertainty weight is used to suppress label noise, thereby further reducing the impact of label noise samples on the model.

[0106] Assume that the current parameters of the weighted classifier V are θ, for each randomly shuffled batch θ is updated via a gradient descent step to The process is expressed as:

[0107]

[0108] in, represents the uncertainty weighted classification loss function, α represents the learning rate, represents the gradient, X represents the sample, represents a noisy label;

[0109] The uncertainty ranking loss is expressed as:

[0110]

[0111] Here, δ represents the margin and is set to 0.2.

[0112] In order to prevent over-learning and memorizing wrong labels, the target probability is estimated based on the obtained model, and an early regularization term is used to obtain the true target probability.

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

[0114]

[0115] Here, 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 expressed as:

[0117]

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

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

[0120] The total training loss is written as:

[0121]

[0122] After the correction network training is completed, the noise source domain is relabeled, and the sample with the largest prediction probability is selected to relabel the sample to achieve the purpose of removing the noise label, providing 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, the source domain samples for training are selected based on the aggregation of classification loss and uncertainty weights, and the average aggregation of samples is expressed as:

[0124]

[0125] In the formula, 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 average aggregated values ​​are arranged in ascending order by category, and the (n-th c ×p) samples as the threshold of the c-th type of fault to select samples, n c is the number of samples of class fault, and p is the selection ratio.

[0126] Through the above sample selection strategy, a multi-noise robust fault diagnosis training model is established. Cleaner samples will be 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 features of samples of the same type are used as positive samples, and samples of different types are used as negative samples to construct contrastive learning loss, so that samples of the same type gather to the same cluster center, and features of samples of different types are far away from each other, thereby achieving separation of overlapping features. The constructed contrastive loss function is:

[0128]

[0129] Where N represents the number of fault samples, S(·) represents the similarity measure, using 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 ratio parameter. Through contrast loss, sample features of the same type are aggregated, and sample features of different types are pulled apart, 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 changes in different working conditions, domain difference metrics are used to reduce domain drift and domain differences. The domain difference metrics used include but are not limited to the maximum mean difference (MMD), multi-kernel maximum mean difference metric (MK-MMD), correlation alignment metric (CORAL) and maximum marginal difference metric (MDD) to reduce domain drift and reduce the domain drift problem caused by changes in different working conditions, and improve the diagnostic accuracy in the cross-working process. Specifically, through a clean source domain, the classification loss of the source domain is minimized, and the difference metric prompts the network to use the clean data of the source domain and the target domain data to learn a discriminative model, which generalizes well to the target domain. The total loss is expressed as:

[0131]

[0132] in, 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 using the maximum marginal difference metric (MDD). The metric loss of the maximum difference metric is expressed as:

[0133]

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

[0135]

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

[0137] By minimizing the classification loss of the source domain through a clean source domain, the MDD difference metric encourages the network to learn a discriminative model using clean data from the source domain and data from the target domain, which generalizes well to the target domain. Minimizing MDD is an adversarial process, expressed as:

[0138]

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

[0140]

[0141] in, represents expectation, σ represents Softmax function, γ=exp(ρ) is used to protect margin ρ, λ i represents the weight, which 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 transfer algorithm, and an adaptive moment estimation algorithm.

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

[0145] The details are as follows: The fault diagnosis test model consists of a trained feature extractor ψ and a classifier f. The labels of the fault samples in the target domain test data are only used to verify the accuracy of the model output prediction labels. If the accuracy reaches the preset standard, the output prediction labels correspond to the categories of each fault sample, and the fault diagnosis model is obtained; otherwise, training continues according to the loss function and optimization algorithm.

[0146] The fault diagnosis method uses the designed training process to achieve cross-operating condition fault feature distribution migration in multi-noise scenarios, which can effectively improve the accuracy of fault diagnosis and improve the level of intelligent fault diagnosis of equipment.

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

[0148] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A multi-noise robust cross-operating condition migration fault diagnosis method under time-varying operating conditions, characterized in that: The following steps are involved: Offline training: S1: Collect sample data and preprocess them to obtain source domain training data set and target domain training data set; S2: Establish a correction network module and a domain adaptation network module. The correction network module uses the source domain training data obtained in step S1 as input to obtain clean source domain data; the domain adaptation network module uses the clean source domain data and target domain training data as input to achieve domain knowledge migration and distribution alignment; 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; S4: Perform training 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 to perform model training, and obtain a fault diagnosis test model; S5 model testing: input the samples of the target domain test data into the fault diagnosis test model constructed in S4 in sequence, and output the fault category of the samples; Online troubleshooting: The target domain data that needs to be diagnosed is input 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-operating condition migration fault diagnosis method under time-varying operating conditions according to claim 1, characterized in that: In step S1, the source domain data includes vibration, acceleration, current or acoustic emission signal data; The preprocessing method is: segmenting the 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.

3. The multi-noise robust cross-operating condition migration fault diagnosis method under time-varying operating conditions according to claim 1, characterized in that: In step S2, the correction network module 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; 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 autoencoder network, so as to extract fault feature information hidden 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, which takes the features output by the feature extractor as input and outputs the weight of each sample; The domain discriminator is composed of a fully connected neural network, which takes the features output by the feature extractor as input, and outputs samples from the source domain or the target domain.

4. The multi-noise robust cross-operating condition migration fault diagnosis method under time-varying operating conditions as claimed in claim 2, characterized in that: Step S2 specifically includes 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 weight generator. Use the classification loss optimization, regularization loss and sorting loss 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 and unlabeled target domain training data obtained in step S2.1 into the domain adaptation network, extract the 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, and use the weighted classification loss, contrast loss, and weighted domain difference metric loss of the domain adaptation network classifier to optimize the feature extractor and domain discriminator to achieve domain knowledge transfer and distribution alignment.

5. The multi-noise robust cross-operating condition migration fault diagnosis method under time-varying operating 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 noise input sample x s , the feature extractor φ outputs the extracted feature φ(x s ), the predicted category distribution μ is calculated by the following formula s and the uncertainty weight w of the sample s : Among them, C and V represent the corrected network classifier and weight generator; The training of the feature extractor, the rectified network classifier, and the weight generator in the rectified network is done with a weighted classification loss: in, is the uncertainty weighted classification loss function, ||·||1 represents the L1 norm, represents the i-th sample Class labels that may be contaminated; The uncertainty weight of the prediction w i Through 1 / w i 2 Effectively suppress the weight of noise samples and reduce classification loss, making the trained model more robust to noisy input data and reducing the impact of noise samples with large uncertainty weights on training loss; S3.2 Randomly shuffle the labels of training samples to obtain noise samples The weight generator generates new uncertainty weights by training with noise samples that disrupt the labels. Assume that the current parameters of the weighted classifier V are θ, for each randomly shuffled batch θ is updated via a gradient descent as The process is expressed as: in, represents the uncertainty weighted classification loss function, α represents the learning rate, represents the gradient, X represents the sample, represents a noisy label; The uncertainty ranking loss is expressed as: Among them, δ represents the margin; Assumptions and represents the target probability and model probability output. For sample i in the kth training, the target probability output of k+1 is calculated as: Among them, β represents momentum; Using the estimated target probability output, for a training batch of B samples, the early regularization term is expressed as: Among them, <·,·> represents the inner product operation; The total training loss is: After the correction network training is completed, the noise source domain is re-labeled. The purpose of removing the noise label is achieved by selecting the re-labeled sample with the maximum prediction probability, providing relatively clean training data for training the adaptive network; S3.3 In the robust domain adaptation module, the clean training data and target domain training data obtained in step S2.1 are used as input to train the model. The source domain samples for training are selected according to the aggregation of classification loss and uncertainty weights. The average aggregation of samples is expressed as: In the formula, 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 average aggregated values ​​are arranged in ascending order by category, and the (n-th c ×p) samples as the threshold of the c-th type of fault to select samples, n c is the number of samples of class fault, p is the selection ratio; Through the above sample selection strategy, a multi-noise robust fault diagnosis training model is established, and cleaner samples will be used to train the robust domain adaptation module.

6. The multi-noise robust cross-operating condition migration fault diagnosis method under time-varying operating conditions according to claim 1, characterized in that: The loss function in step S4 is: (1) The features of samples of the same type are taken as positive samples, and the features of samples of different types are taken as negative samples. Contrastive learning loss is constructed so that samples of the same type gather at the same cluster center and features of samples of different types move away from each other, thereby separating overlapping features. The constructed contrastive loss function is: Where N represents the number of fault samples, S(·) represents the similarity measure, using cosine similarity, Euclidean distance, etc., z i represents the i-th sample feature, represents the positive sample feature, represents the negative sample feature, τ represents the scale parameter; Through contrast loss, sample features of the same type are clustered, and sample features of different types are pulled apart, overcoming the problem of overlapping and difficult separation of different fault features caused by time-varying working conditions; (2) The domain difference metric used encourages the network to learn a discriminative model using clean data from the source domain and data from the target domain. The total loss is expressed as: in, 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.

7. The multi-noise robust cross-operating condition migration fault diagnosis method under time-varying operating conditions according to claim 6, characterized in that: The domain difference metric includes but is not limited to one of the maximum mean difference, multi-core maximum mean difference metric, correlation alignment metric and maximum marginal difference metric to reduce domain drift caused by changes in different working conditions and improve the diagnostic accuracy in cross-working condition processes.

8. The multi-noise robust cross-operating condition migration fault diagnosis method under time-varying operating conditions as claimed in claim 7, characterized in that: When the domain difference metric adopts the maximum marginal difference metric, the metric loss of the maximum difference metric is expressed as: Among them, disp (ρ) (f′,f) represents the marginal difference, which is defined as: Among them, ρ f′ is the assumed margin of the domain discriminator f′ of the domain adaptation network, h f (x) is the labeling function of the classifier f, Φ ρ is the ρ marginal loss; Minimizing MDD is an adversarial process, expressed as: Combining the weighted classification loss and contrast loss of the robust domain adaptation module, the entire optimization process of robust domain adaptation is expressed as: in, represents expectation, σ represents Softmax function, γ=exp(ρ) is used to protect margin ρ, λ i represents the weight, which is defined according to the average aggregation as: in, Represents sample x i The average aggregation of Represents sample x i The minimum value of the average aggregation over k iterations, Represents sample x i Average the aggregated maximum value over k iterations.

9. The multi-noise robust cross-operating condition migration fault diagnosis method under time-varying operating conditions according to claim 1, characterized in that: The optimization algorithm includes but is not limited to one of a stochastic gradient descent algorithm, a root mean square transfer algorithm, and an adaptive moment estimation algorithm.

10. The multi-noise robust cross-operating condition migration fault diagnosis method under time-varying operating conditions according to claim 1, characterized in that: Step S5 is specifically as follows: the fault diagnosis test model is composed of a trained feature extractor ψ and a classifier f. The fault samples in the target domain test data only verify the accuracy of the model output prediction labels. If the accuracy reaches the preset standard, the output prediction labels correspond to the categories of each fault sample to obtain the fault diagnosis model; otherwise, training continues according to the loss function and optimization algorithm.

Images