An intelligent fault diagnosis method based on unsupervised domain adaptation

Abstract

The present invention discloses an intelligent fault diagnosis method based on unsupervised domain adaptation, which relates to the technical field of fault detection. A target part signal annotated with a fault label is obtained as a source domain signal, and an unlabeled target part signal is obtained as a target domain signal; the target part signal is sensor data used to determine whether a target part fails; the fault label is the type of target part failure; the source domain signal and the target domain signal are preprocessed to obtain multiple source domain samples and multiple target domain samples; the source domain samples and the target domain samples are used to train a preset domain adaptation model, the optimization objective function of the domain adaptation model is optimized, and the training of the domain adaptation model is completed; the trained domain adaptation model is used to perform fault diagnosis of the target part. In this way, the fault type of the bearing monitoring data of the parts with similar distribution can be diagnosed based on the known part data, thereby improving the diagnostic effect of the model in practical applications.

Description

Technical Field

[0001] The present invention relates to the technical field of fault detection, and in particular to an intelligent fault diagnosis method based on unsupervised domain adaptation. Background Art

[0002] In the increasingly developed modern industrial society, mechanical equipment is increasingly widely used and is constantly developing in the direction of complexity, intelligence and systematization. Large-scale mechanical equipment is composed of many small parts. Once a part is damaged, it will affect other parts, which is likely to reduce production efficiency and cause economic losses, and even cause casualties. In order to ensure the benefits of industrial production and improve the safety performance of mechanical equipment, it is very necessary to monitor the faults of mechanical equipment parts.

[0003] In the prior art, machine learning methods such as support vector machines (SVM) and artificial neural networks (ANN) are usually used for part fault diagnosis, or deep learning methods such as convolutional neural networks (CNN) and autoencoders (AE) are used for part fault diagnosis. However, in the prior art, in the diagnostic model training stage, in order to ensure the robustness and generalization performance of the test data, the part fault diagnosis method needs to meet two prerequisites: (1) there is a large amount of labeled data available; (2) the test data needs to maintain the same distribution as the training data. However, these two conditions are difficult to meet in actual application scenarios, resulting in unsatisfactory results of the trained model in actual applications. Summary of the invention

[0004] The purpose of the embodiments of the present invention is to provide an intelligent fault diagnosis method based on unsupervised domain adaptation, so as to diagnose the fault types of part monitoring data with similar distribution according to known part data, and improve the diagnostic effect of the model in practical applications.

[0005] The specific technical solutions are as follows:

[0006] The present invention provides an intelligent fault diagnosis method based on unsupervised domain adaptation, the method comprising:

[0007] A target part signal with a fault label is obtained as a source domain signal, and an unlabeled target part signal is obtained as a target domain signal; the target part signal is sensor data used to determine whether a target part fails; the fault label is the type of the target part failure;

[0008] Preprocessing the source domain signal and the target domain signal to obtain a plurality of source domain samples and a plurality of target domain samples;

[0009] Using the source domain samples and the target domain samples to train a preset domain adaptation model, optimizing an optimization objective function of the domain adaptation model, and completing the training of the domain adaptation model;

[0010] The trained domain adaptation model is used to perform fault diagnosis on the target part.

[0011] Optionally, the preset domain adaptation model includes a feature extractor, a classifier and three domain discriminators; wherein the feature extractor includes three convolutional layers and one fully connected layer, and each convolutional layer also includes a BN layer, a pooling layer and a ReLU layer; wherein the pooling layers all adopt a maximum pooling method, and the ReLU layer adopts a linear correction function; the classifier includes two fully connected layers, which are used to predict fault labels using a softmax activation function; each domain discriminator includes two fully connected layers, which are used to predict domain labels using a sigmoid activation function.

[0012] Optionally, using the source domain samples and the target domain samples to train a preset domain adaptation model, optimizing an optimization objective function of the domain adaptation model, and completing the training of the domain adaptation model includes:

[0013] Using the feature extractor to extract features of the source domain samples as source domain features, and using the feature extractor to extract features of the target domain samples as target domain features;

[0014] Inputting the source domain features and the target domain features into the classifier to predict the fault labels of the source domain samples and the target domain samples;

[0015] The source domain features and the target domain features with the fault labels output by the classifier are input into the domain discriminator after passing through a gradient reversal layer through multiple linear mapping, and the domain labels of the source domain samples and the target domain samples are predicted by the domain discriminator; the domain labels include label 1 and label 0, wherein label 1 indicates that the samples belong to the source domain, and label 0 indicates that the samples belong to the target domain;

[0016] The optimization objective function of the domain adaptation model is optimized to complete the training of the domain adaptation model.

[0017] Optionally, the optimization objective function is:

[0018]

[0019]

[0020]

[0021] Among them, L cis the loss function of the classifier, L d1 , L d2 and L d3 are the loss functions of the three domain discriminators, L mmd is the loss function of the feature extractor, λ1, λ2, λ3, λ4 are trade-off parameters for adjusting the contribution of the domain discriminator and the feature extractor; by minimizing L c Optimize the parameters of the feature extractor and the classifier; minimize L d1 , L d2 and L d3 Optimize the parameters of the three domain discriminators; minimize L mmd Optimize the parameters of the feature extractor; maximize L d1 , L d2 and L d3 Optimizing parameters of the feature extractor.

[0022] Optionally, the fully connected layer of the feature extractor uses a multi-core maximum mean difference MK-MMD function as a loss function; the MK-MMD loss function is:

[0023]

[0024]

[0025] Among them, n s Indicates the number of source domain samples, n t represents the number of samples in the target domain, h s and h t They represent the source domain features and target domain features extracted by the fully connected layer respectively, k(,) represents the Gaussian kernel function, σ is the bandwidth, and the multi-kernel maximum mean difference MK-MMD is obtained by taking different σ.

[0026] Optionally, the classifier uses a cross entropy loss function as a loss function; the cross entropy loss function is:

[0027]

[0028] Wherein, C() is the operation of the classifier, F() is the feature extraction operation of the feature extractor, and X s represents the source domain data, Y s For its corresponding label space, n c Indicates the number of preset fault types. If but

[0029] Optionally, the loss function for the domain discriminator is:

[0030] Wherein, D() is the operation of the domain identifier;

[0031] By entropy criterion Calculate prediction uncertainty by using entropy-aware weight ω(H(g))=1+e -H(g) , reweight the samples that are easy to transfer; where g c Indicates the probability that the sample belongs to the Cth class;

[0032] For the mth domain discriminator, the loss function is:

[0033]

[0034] in, represents the prediction result of the i-th source domain sample classifier, Denotes the prediction result of the classifier for the jth target domain sample, D m represents the operation of the mth domain discriminator, represents the features extracted by the mth convolution layer for the i-th source domain sample, Represents the features extracted by the m-th convolution layer for the j-th target domain sample.

[0035] Optionally, the optimization objective function is optimized using a stochastic gradient descent (SGD) algorithm.

[0036] Optionally, after using the source domain samples and the target domain samples to train a preset domain adaptation model, optimizing an optimization objective function of the domain adaptation model, and completing the training of the domain adaptation model, the method further includes:

[0037] Put the unlabeled target domain samples into the trained domain adaptation model for classification, and calculate the fault diagnosis accuracy of the domain adaptation model;

[0038] The training steps are repeated until the maximum preset number of training times is reached.

[0039] Based on the intelligent fault diagnosis method based on unsupervised domain adaptation provided by the embodiment of the present invention, a target part signal with a fault label is obtained as a source domain signal, and an unlabeled target part signal is obtained as a target domain signal; the target part signal is sensor data used to determine whether a target part fails; the fault label is the type of target part failure; the source domain signal and the target domain signal are preprocessed to obtain multiple source domain samples and multiple target domain samples; the source domain samples and the target domain samples are used to train a preset domain adaptation model, optimize the optimization objective function of the domain adaptation model, and complete the training of the domain adaptation model; the trained domain adaptation model is used to perform fault diagnosis of the target part. The source domain signal with existing labels and the target domain signal without labels are put into the neural network model for training. Based on the similarity between the data, the fault type of the part monitoring data with similar distribution can be diagnosed based on the known part data, thereby improving the diagnostic effect of the model in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The present invention will be further described below in conjunction with the accompanying drawings.

[0041] Figure 1 A flowchart of an intelligent fault diagnosis method based on unsupervised domain adaptation provided by an embodiment of the present invention;

[0042] Figure 2 A structural diagram of a domain adaptation network model provided by an embodiment of the present invention;

[0043] Figure 3 A flowchart of another intelligent fault diagnosis method based on unsupervised domain adaptation provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] The embodiment of the present invention provides an intelligent fault diagnosis method based on unsupervised domain adaptation. Figure 1 , Figure 1 A flowchart of an intelligent fault diagnosis method based on unsupervised domain adaptation provided by an embodiment of the present invention, the method may include the following steps:

[0046] S101, obtaining a target part signal annotated with a fault label as a source domain signal, and obtaining an unannotated target part signal as a target domain signal.

[0047] S102, preprocessing the source domain signal and the target domain signal to obtain a plurality of source domain samples and a plurality of target domain samples.

[0048] S103, using the source domain samples and the target domain samples to train a preset domain adaptation model, optimizing the optimization objective function of the domain adaptation model, and completing the training of the domain adaptation model;

[0049] S104: Use the trained domain adaptation model to perform fault diagnosis on the target part.

[0050] The target part signal is sensor data used to determine whether the target part fails; the fault label is the type of fault of the target part.

[0051] In one implementation, the target part may be a core part of a mechanical device, such as a rolling bearing, a gear, etc. The present invention uses an unsupervised domain adaptation model to train a diagnostic network, and the training samples are source domain signals with fault labels. and target domain signal X s is the source domain signal, Y s For X s Corresponding fault label. D s and D T Shared label space Y = {1, 2, ···, N}. The source domain signal is the sensor data of the target part, and the target domain signal is the sensor data of the target part. The fault label of the source domain signal indicates the type of fault of the target part when the target part signal is the source domain signal. By combining the prior knowledge of the source domain signal with the target domain signal, better model training can be achieved.

[0052] In one implementation, preprocessing the source domain signal and the target domain signal may include the following steps: converting the original part signal to the frequency domain by FFT, and then segmenting the part signal to form fixed-length samples, each sample having a fixed length of 600 data points.

[0053] In one embodiment, a preset domain adaptation model includes a feature extractor, a classifier and three domain discriminators; wherein the feature extractor includes three convolutional layers and one fully connected layer, and each convolutional layer also includes a BN layer, a pooling layer and a ReLU layer; wherein the pooling layers all use a maximum pooling method, and the ReLU layer uses a linear correction function; the classifier includes two fully connected layers for predicting fault labels through a softmax activation function; each domain discriminator includes two fully connected layers for predicting domain labels through a sigmoid activation function.

[0054] In one implementation, the domain adaptation model includes multiple domain discriminators, which can be used to learn domain information representation in lower convolution blocks and domain information representation in higher convolution blocks to prevent excessive information loss. In order to fully capture the interaction between feature representation and category information, the idea of ​​CGANs (Conditional Generative Adversarial Networks) is borrowed and combined with multiple domain discriminators. By adding the discriminative information passed by the classifier to each domain discriminator, the cross-domain distribution difference is reduced.

[0055] See Table 1, which shows the backbone network structure of the domain adaptation network model provided by an embodiment of the present invention.

[0056] Table 1

[0057]

[0058] See also Figure 2 , Figure 2 A structural diagram of a domain adaptation network model provided in an embodiment of the present invention.

[0059] Feature extractor is the feature extractor (F), classifier (C) is the classifier (the three discriminators are D1, D2 and D3), and Domain classifier is the domain discriminator. and They represent the parameters of the feature extractor F, F1, F2 and F3 represent the first, second and third convolutional layers respectively, and h s and h t They represent the source domain features and target domain features extracted by the fully connected layer. mmd is the loss function of F, L c is the loss function of C, L d1 , L d2 and L d3 are the loss functions of D1, D2 and D3 respectively. g represents the label prediction result of C for the sample, ω represents the entropy perception weight. θ represents the parameters of classifier C, and denote the parameters of F1, F2 and F3 respectively, and φ1, φ2 and φ3 denote the parameters of domain discriminators D1, D2 and D3 respectively.

[0060] In one embodiment, see Figure 3 ,exist Figure 1 On the basis of, step S103 comprises:

[0061] S1031, using a feature extractor to extract features of source domain samples as source domain features, and using a feature extractor to extract features of target domain samples as target domain features.

[0062] S1032, input the source domain features and the target domain features into the classifier to predict the fault labels of the source domain signal and the target domain signal.

[0063] S1033, the source domain features and target domain features with fault labels output by the classifier are input into the domain discriminator through multiple linear mapping and gradient reversal layers, and the domain labels of the source domain samples and the target domain samples are predicted by the domain discriminator.

[0064] S1034, optimizing the optimization objective function of the domain adaptation model to complete the training of the domain adaptation model.

[0065] The domain labels include label 1 and label 0. Label 1 indicates that it belongs to the source domain, and label 0 indicates that it belongs to the target domain.

[0066] In one implementation, the algorithm for optimizing the objective function may be selected by a worker based on experience.

[0067] In one embodiment, the objective function is optimized using a stochastic gradient descent (SGD) algorithm.

[0068] In one embodiment, the optimization objective function of the domain adaptation model is:

[0069]

[0070]

[0071]

[0072] Among them, L c is the loss function of the classifier, L d1 , L d2 and L d3 are the loss functions of the three domain discriminators, L mmd is the loss function of the feature extractor, λ1, λ2, λ3, λ4 are trade-off parameters used to adjust the contribution of the domain discriminator and feature extractor; by minimizing L c Optimize the parameters of the feature extractor and classifier; minimize L d1 , L d2 and L d3 Optimize the parameters of the three domain discriminators; minimize L mmd Optimize the parameters of the feature extractor; maximize L d1 , L d2 and L d3 Optimize the parameters of the feature extractor.

[0073] In one embodiment, the fully connected layer of the feature extractor uses the multi-kernel maximum mean difference MK-MMD function as the loss function; the MK-MMD loss function is:

[0074]

[0075]

[0076] Among them, h s and h t They represent the source domain features and target domain features extracted by the fully connected layer respectively, k(,) represents the Gaussian kernel function, σ is the bandwidth, and different σ are used to obtain the multi-kernel maximum mean difference MK-MMD.

[0077] In one implementation, by introducing MMD in the fully connected layer of the feature extractor to measure the distance between the source domain features and the target domain features, the invariant features of the domains can be better extracted and the differences between data distributions can be reduced.

[0078] In one embodiment, the classifier uses a cross entropy loss function as a loss function; the cross entropy loss function is:

[0079]

[0080] Wherein, C() is the operation of the feature extractor, F() is the operation of the classifier, and X s represents the source domain data, Y s For its corresponding label space, n c represents the number of categories, If but

[0081] In one implementation, the classifier uses a cross entropy loss function as a loss function, which can make the predicted label closer to the true label.

[0082] In one embodiment, the loss function of the domain discriminator is:

[0083] Where D() is the operation of the domain discriminator;

[0084] By entropy criterion Calculate prediction uncertainty by using entropy-aware weight ω(H(g))=1+e -H(g) , re-weighting samples that are easy to transfer;

[0085] For the mth domain discriminator, the loss function is:

[0086]

[0087] Among them, ns Indicates the number of source domain samples, n t represents the number of samples in the target domain, represents the prediction result of the i-th source domain sample classifier, Denotes the prediction result of the classifier for the jth target domain sample, D m represents the operation of the mth domain discriminator, represents the features extracted by the mth convolution layer for the i-th source domain sample, Represents the features extracted by the m-th convolution layer for the j-th target domain sample.

[0088] In one implementation, the feature extractor confuses the extracted source domain features and target domain features to deceive the domain discriminator due to the effect of the gradient reversal layer. Through adversarial learning between the domain discriminator and the feature extractor, the distribution difference between the source domain and the target domain is further reduced.

[0089] In one embodiment, after step S103, the method further comprises the following steps:

[0090] Step 1: Put the unlabeled target domain samples into the trained domain adaptation model for classification, and calculate the fault diagnosis accuracy of the domain adaptation model.

[0091] Step 2: Repeat the training steps until the maximum preset number of training times is reached.

[0092] In one implementation, the preset number of training times can be set by a technician based on experience and is not limited here.

[0093] In one embodiment, in order to verify the effectiveness of the method of the present invention, bearing data from Case Western Reserve University was selected for verification, and bearing data under load conditions of 1HP to 3HP were taken. For specific division, see Table 2, which is a bearing data division table provided in an embodiment of the present invention.

[0094] Table 2

[0095]

[0096] AB indicates that Dataset-A is used as the source domain and Dataset-B is used as the target domain for migration experiments. Some typical bearing diagnosis methods in recent years are selected for comparison to prove the superiority of the method of the present invention.

[0097] Comparison method 1: CNN (Convolutional neural network) does not use domain adaptive transfer technology, and the network structure is the same as the convolutional network of the present invention. It is only trained with source domain data and tested with target domain data.

[0098] Comparative Method 2: DANN (Domain Adversarial training of Neural Networks) uses a domain discriminator to perform adversarial training on the model to learn domain-invariant features across source and target domains.

[0099] Comparison method 3: MK-MMD (Multi-kernel MMD) minimizes the feature distribution between the source domain and the target domain.

[0100] Two types of migration tasks were performed: (1) migration under different loads on the same end; (2) migration under the same load on different ends. Each experiment was performed 10 times and the average result was taken.

[0101] See Table 3, which shows the results of migration tasks under different loads on the same end provided by an embodiment of the present invention.

[0102] See Table 4, which shows the results of migration tasks under the same load on different ends provided by an embodiment of the present invention.

[0103] Table 3

[0104]

[0105] Table 4

[0106]

[0107] From the experimental results, for the task (1) where the data distribution difference is not big, migration under different loads at the same end; the intelligent fault diagnosis method based on unsupervised domain adaptation provided by the embodiment of the present invention has good effect. From the task (2) where the data distribution difference is very large, because the bearing data is collected from sensors in different positions, the accuracy of CNN is only 0.374, while the network using the domain adaptation method has reached more than 0.6 overall. The intelligent fault diagnosis method based on unsupervised domain adaptation provided by the embodiment of the present invention has an accuracy of 0.792. The above results fully demonstrate the effectiveness and superiority of the intelligent fault diagnosis method based on unsupervised domain adaptation provided by the embodiment of the present invention.

[0108] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0109] Each embodiment in this specification is described in a related manner, and the same or similar parts between the embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0110] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. An intelligent fault diagnosis method based on unsupervised domain adaptation, characterized in that: The following steps are involved: A target part signal with a fault label is obtained as a source domain signal, and an unlabeled target part signal is obtained as a target domain signal; the target part signal is sensor data used to determine whether a target part fails; The fault label is the type of fault of the target part; Preprocessing the source domain signal and the target domain signal to obtain a plurality of source domain samples and a plurality of target domain samples; Using the source domain samples and the target domain samples to train a preset domain adaptation model, optimizing an optimization objective function of the domain adaptation model, and completing the training of the domain adaptation model; Using the trained domain adaptation model to perform fault diagnosis on the target part; The preset domain adaptation model includes a feature extractor, a classifier and three domain discriminators; wherein the feature extractor includes three convolutional layers and one fully connected layer, and each convolutional layer also includes a BN layer, a pooling layer and a ReLU layer; wherein the pooling layer adopts a maximum pooling method, and the ReLU layer adopts a linear correction function; the classifier includes two fully connected layers for predicting fault labels using a softmax activation function; each domain discriminator includes two fully connected layers for predicting domain labels using a sigmoid activation function; Using the source domain samples and the target domain samples to train a preset domain adaptation model, optimizing an optimization objective function of the domain adaptation model, and completing the training of the domain adaptation model, includes: Using the feature extractor to extract features of the source domain samples as source domain features, and using the feature extractor to extract features of the target domain samples as target domain features; Inputting the source domain features and the target domain features into the classifier to predict fault labels of the source domain samples and the target domain samples; The source domain features and target domain features with fault labels output by the classifier are input into the domain discriminator after passing through a gradient reversal layer through multiple linear mapping, and the domain labels of the source domain samples and the target domain samples are predicted by the domain discriminator; the domain labels include label 1 and label 0, wherein label 1 indicates that the samples belong to the source domain, and label 0 indicates that the samples belong to the target domain; Optimizing the optimization objective function of the domain adaptation model to complete the training of the domain adaptation model; The optimization objective function is: Among them, L c is the loss function of the classifier, L d1 , L d2 and L d3 are the loss functions of the three domain discriminators, L mmd is the loss function of the feature extractor, λ1, λ2, λ3, λ4 are trade-off parameters for adjusting the contribution of the domain discriminator and the feature extractor; by minimizing L c Optimize the parameters of the feature extractor and the classifier; minimize L d1 , L d2 and L d3 Optimize the parameters of the three domain discriminators; minimize L mmd Optimize the parameters of the feature extractor; maximize L d1 , L d2 and L d3 Optimizing parameters of the feature extractor; The loss function of the domain discriminator is: Wherein, D() is the operation of the domain identifier; By entropy criterion Calculate prediction uncertainty by using entropy-aware weight ω(H(g))=1+e -H(g) , reweight the samples that are easy to transfer; where g c Indicates the probability that the sample belongs to the Cth class; For the mth domain discriminator, the loss function is: in, represents the prediction result of the i-th source domain sample classifier, Denotes the prediction result of the classifier for the jth target domain sample, D m represents the operation of the mth domain discriminator, represents the features extracted by the mth convolution layer for the i-th source domain sample, Represents the features extracted by the m-th convolution layer for the j-th target domain sample.

2. The intelligent fault diagnosis method based on unsupervised domain adaptation according to claim 1 is characterized in that: The fully connected layer of the feature extractor uses the multi-core maximum mean difference MK-MMD function as the loss function; the MK-MMD loss function is: Among them, n s Represents the number of source domain samples, n t represents the number of samples in the target domain, h s and h t They represent the source domain features and target domain features extracted by the fully connected layer respectively, k(,) represents the Gaussian kernel function, σ is the bandwidth, and the multi-kernel maximum mean difference MK-MMD is obtained by taking different σ.

3. The intelligent fault diagnosis method based on unsupervised domain adaptation according to claim 1 is characterized in that: The classifier uses a cross entropy loss function as a loss function; the cross entropy loss function is: Wherein, C() is the operation of the classifier, F() is the feature extraction operation of the feature extractor, and X s represents the source domain data, Y s For its corresponding label space, n c Indicates the number of preset fault types. If but 4. The intelligent fault diagnosis method based on unsupervised domain adaptation according to claim 1 is characterized in that: The optimization objective function is optimized using the stochastic gradient descent (SGD) algorithm.

5. The intelligent fault diagnosis method based on unsupervised domain adaptation according to claim 1 is characterized in that: After using the source domain samples and the target domain samples to train a preset domain adaptation model, optimizing the optimization objective function of the domain adaptation model, and completing the training of the domain adaptation model, the method further includes: Put the unlabeled target domain samples into the trained domain adaptation model for classification, and calculate the fault diagnosis accuracy of the domain adaptation model; The training steps are repeated until the maximum preset number of training times is reached.

Images