A migratable multi-scale rotating machinery fault diagnosis method and system

Abstract

The present application relates to the technical field of fault identification, in particular to a kind of migratory multi-scale rotating machinery fault diagnosis method and system, method includes: obtaining training dataset;The training dataset includes source domain sample set and target domain sample set;Based on the training dataset, training fault diagnosis model is carried out, and the trained fault diagnosis model is obtained;The fault diagnosis model includes multi-scale feature extraction network, fully connected neural network classifier and feature alignment network based on deconvolution;The real-time vibration signal of rotating machinery is obtained, and the real-time vibration signal is input into the trained fault diagnosis model, and fault diagnosis result is obtained.The present application improves the accuracy of fault diagnosis, can adapt to the data distribution of target domain, avoid the waste of time and experience caused by training from scratch.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault identification, in particular to a transferable multi-scale rotating machinery fault diagnosis method and system. BACKGROUND

[0002] With the development of artificial intelligence, data-driven intelligent fault diagnosis technology is continuously applied in various aspects of industry. For example, health state monitoring, mechanical life prediction and maintenance decision, etc. The traditional manual diagnosis method based on rich expert knowledge is gradually evolving into a model diagnosis method driven by data. When the working condition of a rotating machinery changes, the fault diagnosis model trained in the source domain will have a model drift phenomenon in the target domain, resulting in a decrease in diagnosis accuracy. SUMMARY

[0003] Therefore, the present application provides a transferable multi-scale rotating machinery fault diagnosis method and system, which improves the accuracy of fault diagnosis, can adapt to the data distribution of the target domain, and avoids the waste of time and experience caused by training from scratch.

[0004] To achieve the above object, the present application provides the following scheme:

[0005] A transferable multi-scale rotating machinery fault diagnosis method, comprising:

[0006] obtaining a training data set; the training data set comprises a source domain sample set and a target domain sample set;

[0007] training a fault diagnosis model based on the training data set to obtain a trained fault diagnosis model; the fault diagnosis model comprises a multi-scale feature extraction network, a fully connected neural network classifier and a feature alignment network based on deconvolution;

[0008] obtaining a real-time vibration signal of a rotating machinery, inputting the real-time vibration signal into the trained fault diagnosis model, and obtaining a fault diagnosis result.

[0009] Preferably, the source domain sample set and the target domain sample set have different distributions.

[0010] Preferably, the training of the fault diagnosis model based on the training data set to obtain the trained fault diagnosis model comprises:

[0011] training the multi-scale feature extraction network and the fully connected neural network classifier based on the source domain sample set to obtain the trained multi-scale feature extraction network and the trained fully connected neural network classifier;

[0012] training the deconvolution-based feature alignment network based on the source domain sample set and the target domain sample set to obtain the trained deconvolution-based feature alignment network;

[0013] The multi-scale feature extraction network is used for multi-scale feature extraction and fusion of an input sample to obtain fused features; the input sample is a target domain sample or a source domain sample.

[0014] The fully connected neural network classifier is used for classifying the input fused features to obtain a fault category prediction result.

[0015] The deconvolution-based feature alignment network is used for discriminating the domain label of the fault category prediction result to distinguish whether the fault category prediction result comes from the target domain or the source domain; the multi-scale feature extraction network and the fully connected neural network classifier confuse the discrimination result of the deconvolution-based feature alignment network, and are mutually antagonistic to realize alignment of the source domain and the target domain.

[0016] Preferably, the multi-scale feature extraction network comprises a feature fusion network and a plurality of sub-feature extraction networks.

[0017] The plurality of sub-feature extraction networks perform feature extraction on the input sample to obtain a plurality of features of different scales.

[0018] The feature fusion network fuses the plurality of features of different scales to obtain the fused features.

[0019] Preferably, the loss function of the deconvolution-based feature alignment network is selected as a cross-entropy loss function with label smoothing.

[0020] The application further provides a transferable multi-scale rotating machinery fault diagnosis system, comprising:

[0021] A data acquisition module is configured to acquire a training data set; the training data set comprises a source domain sample set and a target domain sample set.

[0022] A training module is configured to train a fault diagnosis model based on the training data set to obtain a trained fault diagnosis model; the fault diagnosis model comprises a multi-scale feature extraction network, a fully connected neural network classifier, and a deconvolution-based feature alignment network.

[0023] A diagnosis module is configured to acquire a real-time vibration signal of a rotating machinery, input the real-time vibration signal into the trained fault diagnosis model, and obtain a fault diagnosis result.

[0024] Preferably, the source domain sample set and the target domain sample set are different in distribution.

[0025] Preferably, the training module comprises:

[0026] a first training unit configured to train the multi-scale feature extraction network and the fully connected neural network classifier based on the source domain sample set, to obtain the trained multi-scale feature extraction network and the trained fully connected neural network classifier;

[0027] a second training unit configured to train the deconvolution-based feature alignment network based on the source domain sample set and the target domain sample set, to obtain the trained deconvolution-based feature alignment network;

[0028] The multi-scale feature extraction network is configured to perform multi-scale feature extraction and fusion on an input sample to obtain fused features, wherein the input sample is a target domain sample or a source domain sample.

[0029] The fully connected neural network classifier is configured to classify the input fused features to obtain a fault category prediction result.

[0030] The deconvolution-based feature alignment network is configured to identify the domain label of the fault category prediction result, and distinguish whether the fault category prediction result comes from the target domain or the source domain. The multi-scale feature extraction network and the fully connected neural network classifier confuse the identification result of the deconvolution-based feature alignment network, and are mutually antagonistic to realize alignment of the source domain and the target domain.

[0031] Preferably, the multi-scale feature extraction network comprises a feature fusion network and a plurality of sub-feature extraction networks.

[0032] The plurality of sub-feature extraction networks perform feature extraction on the input sample to obtain a plurality of features of different scales.

[0033] The feature fusion network fuses the plurality of features of different scales to obtain the fused features.

[0034] Preferably, the loss function of the deconvolution-based feature alignment network is a cross-entropy loss function with label smoothing.

[0035] According to the embodiments of the present application, the following technical effects are provided:

[0036] The application relates to a migratable multi-scale rotating machinery fault diagnosis method and system, and the method comprises the following steps: acquiring a training data set; the training data set comprises a source domain sample set and a target domain sample set; a fault diagnosis model is trained based on the training data set to obtain a trained fault diagnosis model; the fault diagnosis model comprises a multi-scale feature extraction network, a full-connection neural network classifier and a feature alignment network based on deconvolution; real-time vibration signals of rotating machinery are acquired, the real-time vibration signals are input into the trained fault diagnosis model, and a fault diagnosis result is obtained. BRIEF DESCRIPTION OF DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description only represent some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort on the basis of these drawings.

[0038] Figure 1 The flow chart of the migratable multi-scale rotating machinery fault diagnosis method of the present application is shown in the figure.

[0039] Figure 2 The structure diagram of the migratable multi-scale rotating machinery fault diagnosis system of the present application is shown in the figure.

[0040] Symbol explanation: 1-data acquisition module, 2-training module, 3-diagnosis module. DETAILED DESCRIPTION

[0041] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments only represent some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without any creative effort fall within the protection scope of the present application.

[0042] The present application aims to provide a migratable multi-scale rotating machinery fault diagnosis method and system, which improves the accuracy of fault diagnosis, can adapt to the data distribution of the target domain, and avoids the waste of time and experience caused by training from scratch.

[0043] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail in combination with the drawings and specific embodiments.

[0044] Figure 1 The flow chart of the migratable multi-scale rotating machinery fault diagnosis method of the present application is shown in the figure.Figure 1 As shown, the present invention provides a transferable multi-scale rotating machinery fault diagnosis method, comprising:

[0045] Step S1: Obtain the training dataset; the training dataset includes a source domain sample set and a target domain sample set.

[0046] The source domain sample set is D s ={(X s ,Y s )|(X s ,Y s )-P s (X s ,Y s )};

[0047] in: N* represents the number of samples in the source domain sample set, and s represents the source domain. This represents the i-th source domain sample in the source domain sample set. for The label, i.e., the type of rotating machinery failure, P s (X s ,Y s ) is (X s ,Y s The distribution of ).

[0048] The target domain sample set is D. t ={(X t ,Y t )|(X t ,Y t )-P t (X t ,Y t )};

[0049] in: Let t be the k-th target domain sample in the target domain sample set, where K is the number of target domain samples in the target domain sample set, and t represents the target domain. for The label, i.e., the type of rotating machinery failure, P t (X t ,Y t ) is (X t ,Y t The distribution of ).

[0050] P t (X t ,Y t ) and P s (X s ,Ys ) are different, i.e. P t (X t , Y t )≠ P s (X s , Y s ).

[0051] And the distribution of each target sample in the target domain sample set is also different, i.e. l∈K.

[0052] Step S2, training a fault diagnosis model based on the training data set to obtain a trained fault diagnosis model; the fault diagnosis model includes a multi-scale feature extraction network, a fully connected neural network classifier and a feature alignment network based on deconvolution.

[0053] Training the multi-scale feature extraction network and the fully connected neural network classifier based on the source domain sample set to obtain a trained multi-scale feature extraction network and a trained fully connected neural network classifier.

[0054] Training the feature alignment network based on deconvolution based on the source domain sample set and the target domain sample set to obtain a trained feature alignment network based on deconvolution.

[0055] The multi-scale feature extraction network is used for multi-scale feature extraction and fusion of input samples to obtain fusion features; the input samples are target domain samples or source domain samples.

[0056] The multi-scale feature extraction network includes a plurality of sub-feature extraction networks and a feature fusion network.

[0057] The plurality of sub-feature extraction networks extract features from input samples to obtain features of different scales.

[0058] The feature fusion network fuses features of different scales to obtain the fusion features.

[0059] In this embodiment, the number of sub-feature extraction networks is three, and the three sub-feature extraction networks are and The feature fusion network is The three sub-feature extraction networks and use different sizes of convolution kernels to extract features of different granularities, and the feature fusion network fuses corresponding elements of features of different granularities to obtain fusion features.

[0060]

[0061]

[0062]

[0063]

[0064] wherein: f1 is the output of the first sub-feature extraction network, f2 is the output of the second sub-feature extraction network, f3 is the output of the third sub-feature extraction network, and f4 is the fusion feature. 1,3 2,3 3,3

[0065] The three sub-feature extraction networks are respectively The specific parameters of the feature fusion network are shown in Table 1.

[0066] Table 1 Parameters of each network in the multi-scale feature extraction network

[0067]

[0068]

[0069] wherein: Input represents input, CNN is a convolutional neural network (Convolutional Neural Network), BN is a batch sample normalization network (Batch Normalization), Tanh is an activation function, MaxPool is maximum pooling, Stride is a step length, padding is padding, Kernel is a kernel, i.e. a convolution kernel, and Shape is a shape.

[0070] The fully connected neural network classifier is configured to classify the input fusion feature to obtain a fault category prediction result.

[0071] The fully connected neural network classifier includes an adaptive max-pooling layer, a fully connected layer, and a softmax layer. The adaptive max-pooling layer is configured to perform adaptive max-pooling on the input fusion feature to obtain a feature vector.

[0072] The fully connected layer is configured to predict the fault type, and the softmax layer is configured to perform normalization processing to obtain a fault prediction probability and a result of the rotating machinery, i.e. the fault category prediction result. The parameters of the fully connected neural network classifier are shown in Table 2.

[0073] Table 2 Parameters of the fully connected neural network classifier are shown in Table 2.

[0074]

[0075] ​​​​Wherein, AdapMaxPool is an adaptive maximum pooling layer, flatten is a flattening operation layer, FC is a full connection layer, and outsize is an output size.

[0076] The feature alignment network based on deconvolution is used for identifying the domain label of the fault category prediction result, distinguishing whether the fault category prediction result comes from the target domain or the source domain, and confusing the identification result of the feature alignment network based on deconvolution by the multi-scale feature extraction network and the full connection neural network classifier, so as to realize the alignment of the source domain and the target domain. The loss function of the feature alignment network based on deconvolution is selected as a cross-entropy loss function with label smoothing. The parameters of the feature alignment network based on deconvolution are shown in Table 3.

[0077] Table 3 Parameters of the feature alignment network based on deconvolution

[0078]

[0079]

[0080]

[0081] Step S3, obtaining a real-time vibration signal of the rotating machine, inputting the real-time vibration signal into the trained fault diagnosis model, and obtaining a fault diagnosis result.

[0082] Figure 2 The structure diagram of the transferable multi-scale rotating machine fault diagnosis system is shown in the figure. Figure 2 As shown in the figure, the application provides a transferable multi-scale rotating machine fault diagnosis system, which comprises:

[0083] A data acquisition module 1 is configured to acquire a training data set, wherein the training data set comprises a source domain sample set and a target domain sample set.

[0084] A training module 2 is configured to train a fault diagnosis model based on the training data set, and obtain a trained fault diagnosis model, wherein the fault diagnosis model comprises a multi-scale feature extraction network, a full connection neural network classifier and a feature alignment network based on deconvolution.

[0085] A diagnosis module 3 is configured to acquire a real-time vibration signal of the rotating machine, input the real-time vibration signal into the trained fault diagnosis model, and obtain a fault diagnosis result.

[0086] Optionally, the source domain sample set and the target domain sample set are different in distribution.

[0087] Optionally, the training module 2 comprises:

[0088] The first training unit is configured to train the multi-scale feature extraction network and the fully connected neural network classifier based on the source domain sample set, and obtain the trained multi-scale feature extraction network and the trained fully connected neural network classifier.

[0089] The second training unit is configured to train the deconvolution-based feature alignment network based on the source domain sample set and the target domain sample set, and obtain the trained deconvolution-based feature alignment network.

[0090] The multi-scale feature extraction network is configured to perform multi-scale feature extraction and fusion on an input sample, and obtain fused features, wherein the input sample is a target domain sample or a source domain sample.

[0091] The fully connected neural network classifier is configured to classify the input fused features, and obtain a fault category prediction result.

[0092] The deconvolution-based feature alignment network is configured to identify the domain label of the fault category prediction result, and distinguish whether the fault category prediction result comes from the target domain or the source domain, wherein the multi-scale feature extraction network and the fully connected neural network classifier confuse the identification result of the deconvolution-based feature alignment network, and are mutually antagonistic to realize alignment of the source domain and the target domain.

[0093] Optionally, the multi-scale feature extraction network comprises a feature fusion network and a plurality of sub-feature extraction networks.

[0094] The plurality of sub-feature extraction networks perform feature extraction on the input sample, and obtain a plurality of features of different scales.

[0095] The feature fusion network fuses the plurality of features of different scales, and obtains the fused features.

[0096] Optionally, the loss function of the deconvolution-based feature alignment network is a cross-entropy loss function with label smoothing.

[0097] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts of each embodiment can be referred to each other. For the system disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the related parts can be referred to the method part.

[0098] The principles and implementation manners of the present application are described by using specific examples in the present application, and the above examples are only used to help understand the method of the present application and its core idea; meanwhile, for the general technical personnel in the art, the specific implementation manners and application ranges will be changed according to the idea of the present application. In conclusion, the content of the present specification should not be understood as the limitation of the present application.

Claims

1. A transferable multi-scale rotating machinery fault diagnosis method, characterized in that, include: Obtain the training dataset; The training dataset includes a source domain sample set and a target domain sample set; The fault diagnosis model is trained based on the training dataset to obtain the trained fault diagnosis model; the fault diagnosis model includes a multi-scale feature extraction network, a fully connected neural network classifier, and a deconvolution-based feature alignment network; The deconvolution-based feature alignment network is used to identify the domain label of the fault category prediction result and distinguish whether the fault category prediction result comes from the target domain or the source domain. The loss function of the deconvolution-based feature alignment network is the labeled smooth cross-entropy loss function. The parameters of the deconvolution-based feature alignment network are shown in the table below: The real-time vibration signal of the rotating machinery is acquired, and the real-time vibration signal is input into the trained fault diagnosis model to obtain the fault diagnosis result.

2. The transferable multi-scale rotating machinery fault diagnosis method according to claim 1, characterized in that, The source domain sample set and the target domain sample set have different distributions.

3. The transferable multi-scale rotating machinery fault diagnosis method according to claim 1, characterized in that, The fault diagnosis model is trained based on the training dataset to obtain the trained fault diagnosis model; The fault diagnosis model includes a multi-scale feature extraction network, a fully connected neural network classifier, and a deconvolution-based feature alignment network, including: The multi-scale feature extraction network and the fully connected neural network classifier are trained based on the source domain sample set to obtain the trained multi-scale feature extraction network and the trained fully connected neural network classifier. The deconvolution-based feature alignment network is trained based on the source domain sample set and the target domain sample set to obtain the trained deconvolution-based feature alignment network. The multi-scale feature extraction network is used to extract and fuse multi-scale features from the input samples to obtain fused features; the input samples are either target domain samples or source domain samples. The fully connected neural network classifier is used to classify the input fused features to obtain the fault category prediction result; The multi-scale feature extraction network and the fully connected neural network classifier confuse the discrimination results of the deconvolution-based feature alignment network, thus opposing each other and achieving alignment between the source and target domains.

4. The transferable multi-scale rotating machinery fault diagnosis method according to claim 3, characterized in that, The multi-scale feature extraction network includes a feature fusion network and several sub-feature extraction networks; Several sub-feature extraction networks extract features from the input sample to obtain features at several different scales; The feature fusion network fuses features at several different scales to obtain the fused features.

5. A transferable multi-scale rotating machinery fault diagnosis system, characterized in that, include: The data acquisition module is used to acquire the training dataset; The training dataset includes a source domain sample set and a target domain sample set; The training module is used to train the fault diagnosis model based on the training dataset to obtain the trained fault diagnosis model; the fault diagnosis model includes a multi-scale feature extraction network, a fully connected neural network classifier, and a deconvolution-based feature alignment network; The deconvolution-based feature alignment network is used to identify the domain label of the fault category prediction result and distinguish whether the fault category prediction result comes from the target domain or the source domain. The loss function of the deconvolution-based feature alignment network is the labeled smooth cross-entropy loss function. The parameters of the deconvolution-based feature alignment network are shown in the table below: The diagnostic module is used to acquire real-time vibration signals of rotating machinery, input the real-time vibration signals into the trained fault diagnosis model, and obtain fault diagnosis results.

6. The transferable multi-scale rotating machinery fault diagnosis system according to claim 5, characterized in that, The source domain sample set and the target domain sample set have different distributions.

7. The transferable multi-scale rotating machinery fault diagnosis system according to claim 5, characterized in that, The training module includes: The first training unit is used to train the multi-scale feature extraction network and the fully connected neural network classifier based on the source domain sample set, so as to obtain the trained multi-scale feature extraction network and the trained fully connected neural network classifier. The second training unit is used to train the deconvolution-based feature alignment network based on the source domain sample set and the target domain sample set to obtain the trained deconvolution-based feature alignment network. The multi-scale feature extraction network is used to extract and fuse multi-scale features from the input samples to obtain fused features; the input samples are either target domain samples or source domain samples. The fully connected neural network classifier is used to classify the input fused features to obtain the fault category prediction result; The multi-scale feature extraction network and the fully connected neural network classifier confuse the discrimination results of the deconvolution-based feature alignment network, thus opposing each other and achieving alignment between the source and target domains.

8. The transferable multi-scale rotating machinery fault diagnosis system according to claim 7, characterized in that, The multi-scale feature extraction network includes a feature fusion network and several sub-feature extraction networks; Several sub-feature extraction networks extract features from the input sample to obtain features at several different scales; The feature fusion network fuses features at several different scales to obtain the fused features.

Images