A method for equipment fault diagnosis and prediction based on transfer learning

Abstract

The present invention discloses a method for equipment fault diagnosis and prediction based on transfer learning, belonging to the field of fault diagnosis technology, comprising a prediction model and a diagnosis model; the prediction model is composed of a 4-layer CNN, a 2-layer LSTM, and a 3-layer fully connected layer; the diagnosis model is composed of a 4-layer CNN, a 2-layer LSTM, and a 3-layer fully connected layer superimposed; based on the domain adaptation technology in transfer learning using a benchmark model, a measurement method for the difference in sample features based on working conditions of two modes is proposed, and this difference is used as a term in the loss function. During the training process, the difference in the feature distribution of samples with the same working condition in the training set and the test set is continuously reduced, thereby improving the generalization of the fault prediction model; experimental results on a commercial modular aviation propulsion system simulation dataset demonstrate the effectiveness and reliability of the model, as well as the correctness of the corresponding parameter exploration strategy.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault diagnosis, and in particular relates to a method for diagnosing and predicting equipment faults based on transfer learning. Background Art

[0002] Transfer learning is a technique that uses certain means to continue using old models obtained in previous fields in new fields, taking into account certain similarities between the data or models used in the task. For an overview of transfer learning, see Domain adaptation is an important technique in transfer learning methods. It transfers learning systems from one domain to another and can effectively solve the problem of different distributions between domains. The domain with rich existing knowledge is called the source domain, and the domain where new knowledge is to be learned is called the target domain.

[0003] With the development of deep learning, some methods based on deep data mining, such as deep adaptation networks and joint adaptation networks, have been used to deal with various types of DFD problems. The key idea is to minimize the difference in marginal distribution or conditional distribution between the source domain input data and the target domain input data. In the following introduction, the methods of fault diagnosis and prediction based on domain adaptation are mainly discussed, and the related research work is reviewed, mainly involving the fault diagnosis and prediction of rolling bearings and rotating machinery.

[0004] In terms of fault diagnosis, some research has focused on using deep transfer networks to learn fault diagnosis models from one domain with rich labeled data and migrate them to another domain without labeled data. For example, a new deep transfer learning method, the deep convolutional transfer learning network, improves the accuracy of bearing fault diagnosis by migrating the invariant features of one machine to another machine without labeled data. A feature-based transfer neural network is used to migrate from bearings used in laboratory equipment to bearings used in actual equipment for intelligent fault diagnosis.

[0005] Some other fault diagnosis studies focus on transfer learning between different operating conditions; for example, a cross-domain feature fusion method based on transfer component analysis is used for gearbox fault diagnosis under different working conditions; a domain adaptation method for bearing fault diagnosis under different working conditions, in which principal component analysis is used to construct the feature subspace of spectral data, and subspace analysis is used to reduce distribution differences; a deep neural network based on domain adaptation is used for bearing fault diagnosis under different working conditions, which is trained using data in the source domain and partially labeled data in the target domain; a sparse autoencoder is used to extract spectral data features of bearings under different operating conditions, and then the maximum mean difference is minimized to adapt to the distribution of learned transferable features; a CNN based on domain adaptation is used to complete oil bearing fault diagnosis under different noise environments; a fault diagnosis model based on neural network transfer learning makes full use of data under different working conditions, and discusses the impact of similar working conditions on performance and negative transfer effects; a joint domain adaptation method for industrial fault diagnosis under different operating conditions.

[0006] For fault prediction, there are not many published related studies based on domain adaptation. By combining several existing algorithms, such as TCA, shrinkage denoising autoencoder and Hilbert-Huang transform, a new RUL prediction scheme framework is developed. This framework can well train relevant models from different but related degradation data and improve the accuracy of RUL prediction, especially when facing small-scale bearing data sets. However, this framework is not an end-to-end architecture. Feature representation learning, transfer learning and RUL prediction are separate steps. An end-to-end deep transfer learning network based on sparse autoencoder is proposed for RUL prediction of tools; three migration strategies, namely weight migration, feature transfer learning and weight update, are mainly adopted to migrate the deep network trained by the historical fault data of the machine to predict the fault and life of the newly generated data. A migration compression coding strategy for hyperplane classifier is proposed to predict the operating law of rolling bearings under different working conditions.

[0007] In various methods and applications of fault diagnosis and prediction, the main focus is on data preprocessing, feature extraction, and many optimizations made at the model level. However, a phenomenon exists in many industrial datasets, that is, the ratio of the number of samples in different categories is too large. For example, in classified fault diagnosis, the ratio of the number of normal samples to the number of fault samples may be around 10:1 or 100:1. Usually, in imbalanced problems, the category with a smaller number of samples is called the minority class, and the category with a larger number of samples is called the majority class.

[0008] The processing methods at the data level can be roughly divided into oversampling, undersampling and hybrid methods. The methods at the data level can be independent of the specific type of classifier, that is, they can be applied to models composed of any classifiers, and are more versatile. Among the sampling-based algorithms, the most common is the random downsampling strategy. There are also some methods based on clustering oversampling methods, which cluster the data set and then oversample different categories. Roozbeh et al. proposed an oversampling technology based on missing value completion, but they did not pay enough attention to the importance of some minority classes, resulting in poor quality of some minority class samples generated by this method. In order to generate more effective and useful minority class samples, some scholars use algorithms to select suitable seed samples to generate new minority class samples.

[0009] Algorithm-based methods for processing imbalanced data can generally be divided into two categories: cost-sensitive algorithms and embedding-based methods. Cost-sensitive methods improve the classifier algorithm so that during training or parameter updates, the classifier's accuracy for minority class samples is improved while minimizing the loss of accuracy for majority class samples, thereby improving the overall performance of the final classifier. Embedding-based methods are relatively less widely used.

[0010] Improved Fisher discriminant analysis, weighted extreme value learning machines, and fuzzy support vector machines are widely used to solve imbalance problems in the industrial field. At the same time, deep learning has also been introduced into intelligent diagnosis of mechanical faults to achieve feature extraction and fault classification for imbalanced learning problems. To optimize classification, we usually optimize the final loss function, using a weighted softmax maximum loss function, etc.

[0011] In the existing technology, there is not much research based on domain adaptation for fault prediction. Some ensemble-based category imbalanced learning methods improve performance by combining methods based on the algorithm level or data level, including EasyEnsemble, RankCost, and improved support vector machines. In practical applications, some resampling methods can eliminate redundant information in the learning algorithm, thereby performing better than algorithm-level imbalanced learning. However, industrial scenario data is usually high in dimensionality and has fewer fault samples, making it difficult for these algorithm-level methods to predict faults. Summary of the Invention

[0012] Problem to be solved

[0013] In response to the problems of high data dimension, few fault samples and difficulty in fault prediction, the present invention provides a device fault diagnosis and prediction method based on transfer learning.

[0014] Technical Solution

[0015] To solve the above problems, the present invention adopts the following technical solutions.

[0016] A device fault diagnosis and prediction method based on transfer learning, including a prediction model and a diagnosis model;

[0017] The prediction model consists of 4 layers of CNN, 2 layers of LSTM and 3 layers of fully connected layers;

[0018] The diagnostic model is composed of 4 layers of CNN, 2 layers of LSTM and three layers of fully connected layers.

[0019] The specific steps are as follows:

[0020] S1. Preprocess the migrated data, and the processed data samples constitute the dataset D;

[0021] S2, using the CNN in the prediction model to extract feature information of the data processed in step S1;

[0022] S3, using the LSTM in the prediction model to mine the data feature information extracted in step S2;

[0023] S4, connecting the LSTM after information mining in step S3 to the fully connected layer of the prediction model to obtain the prediction result;

[0024] S5. diagnose the fault of the sample using the diagnostic model;

[0025] The data set D in step S1 is divided into a training set and a test set;

[0026] The specific diagnostic steps in step S5 are as follows:

[0027] S501, using a cost-sensitive algorithm to adjust the weight ratio of the classification in the loss function;

[0028] S502, obtaining pseudo labels of samples in the classified and adjusted data;

[0029] S503, using a conditional adaptive algorithm to calculate the feature distribution difference of the same category in the training set and the test set;

[0030] S504: Reduce the distribution difference calculated in step S503 to complete the diagnosis.

[0031] Preferably, the preprocessing in step S1 is the segmentation of time series data; in step S1, the data set D is composed of N sub-data sets consisting of d-dimensional time series samples, and each time series is represented by X i ∈D, the length of each time series is L, where X i ∈R d×L,i∈N; the sliding window divides each time series into small time windows; the jth sliding window is represented as As the most original input, it is placed on the input side of the model.

[0032] Furthermore, the prediction model also includes a wc-DAN model, and the wc-DAN model is used to obtain better prediction and migration performance.

[0033] Furthermore, the expression of the wc-DAN model is as follows:

[0034] l (*) =l c (D s ,y s ,θ l )+λMMD(D s , D t ,θ' l ).

[0035] The first item c (D s ,y s ,θ l ) represents the loss term of the CNN-LSTM network. The test set input data does not participate in the calculation of the first term. The second term λMMD(D s , D t ,θ' l ) is the distribution difference between the source domain and target domain data under different working conditions; θ l These are the parameters that need to be optimized when minimizing the loss, including the relevant weights and biases of CNN, LSTM, and fully connected layers. The training set and test set features used in the MMD calculation process need to be output by CNN, θ' l Represents the parameters of the CNN part used, and θ l The parameters in the CNN part are the same.

[0036] Preferably, in order to improve the generalization ability of the prediction model, a domain adaptation algorithm is also used in the prediction process of the prediction model. The goal of the domain adaptation algorithm is to minimize the marginal distribution difference between the training and test data sets, that is, P(X s ) and P(X t ); P(X) is expressed as a mixture of operating condition distributions, and the specific formula is as follows:

[0037]

[0038] The α i P(X|Z=C i ) represents the prior probability of the working condition, that is, the data set belongs to working condition Ci The proportion of samples; redefine X according to the mixed marginal distribution s and X t MMD between, using MMD C (X s ,X t ) is represented by X s With X t wcMMD between;MMD C (X s ,X t ) can be defined as two modes, which are the intra-class migration mode and the weighted conditional migration mode.

[0039] Preferably, the specific steps of step S501 are as follows:

[0040] S5001. Use a cost-sensitive imbalanced strategy based on class weights, so that the misclassification results of the minority class are given greater weight in the calculation of the loss function and the update of parameters, and the weight of the majority class is reduced accordingly;

[0041] The specific steps of step S502 are as follows:

[0042] S5002. Use the conditional distribution domain adaptation algorithm to narrow the conditional distribution difference between the source domain and the target domain, and minimize the distribution difference of features under the same category in the source domain and the target domain; use an iterative pseudo-label algorithm to obtain the category information of the test set and provide pseudo-labels for the calculation of the conditional distribution domain adaptation algorithm.

[0043] Furthermore, the formula of the loss function is as follows:

[0044] l IFDM-DA =l cs (X s , Y s ,θ l )+λMMD CDA (X s , X t , Y s , Y p ,θ′ l )

[0045] l cs (X s , Y s ,θ l ) represents the loss term of the Deep CNN-LSTM network with the cost-sensitive strategy, and the second term λMMD CDA (X s , X t , Y s , Y p ,θ′l ) is the conditional distribution difference between the source domain and target domain data calculated by category; θ l is the parameter that needs to be optimized when minimizing the loss, λ is the weight parameter of the domain adaptive loss function; θ′ l ∈θ l , because MMD CDA The calculation only involves the network layer before LSTM in the framework, θ′ l Only contains relevant parameters of CNN and LSTM.

[0046] Furthermore, the weight in step S5001 is inversely proportional to the number of samples of the category in the data set.

[0047] Furthermore, the step one of data preprocessing is to segment the time series data of the big data.

[0048] Beneficial effects

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] (1) In this invention, a method for measuring the difference in sample characteristics between two modes based on working conditions is proposed based on the domain adaptation technology in transfer learning through a benchmark model. This difference is used as a term in the loss function. During the training process, the difference in the feature distribution of samples with the same working conditions in the training set and the test set is continuously reduced, thereby improving the generalization of the fault prediction model. Experimental results on a commercial modular aviation propulsion system simulation dataset demonstrate the effectiveness and reliability of the model, as well as the correctness of the corresponding parameter exploration strategy.

[0051] (2) The present invention uses a cost-sensitive algorithm through an unbalanced fault diagnosis model to adjust the weights of different categories in the loss function, thereby improving the classifier's emphasis on misclassification of fault samples; uses an unsupervised method to obtain pseudo labels for samples in the test set, and uses the conditional distribution adaptation algorithm in domain adaptation to calculate the feature differences of the same category in the training set and the test set. The size of the difference is used as an item in the loss function, and the difference in features of the same category between the training set and the test set is continuously reduced during the training process, thereby improving the generalization of the model and the classification accuracy of minority classes. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments or exemplary embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments or exemplary descriptions. Obviously, the drawings described below are only some embodiments of the present application and should not be regarded as limiting the scope. For ordinary technicians in this field, other drawings can be obtained according to the drawings without paying creative work.

[0053] Figure 1 Schematic diagram of the process of the present invention Figure 1 ;

[0054] Figure 2 Schematic diagram of the process of the present invention Figure 2 ;

[0055] Figure 3 This is the model diagram of wc-DAN in the present invention;

[0056] Figure 4 This is a model diagram of IFDM-DA in the present invention. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Generally, the components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations.

[0058] Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the present application for protection, but merely represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.

[0059] Example 1:

[0060] like Figure 1-2 As shown, a device fault diagnosis and prediction method based on transfer learning includes a prediction model and a diagnosis model;

[0061] The prediction model consists of 4 CNN layers, 2 LSTM layers, and 3 fully connected layers;

[0062] The diagnostic model is composed of 4 layers of CNN, 2 layers of LSTM and three layers of fully connected layers.

[0063] The specific steps are as follows:

[0064] S1. Preprocess the migrated data, and the processed data samples constitute the dataset D;

[0065] S2, using the CNN in the prediction model to extract feature information of the data processed in step S1;

[0066] S3, using the LSTM in the prediction model to mine the data feature information extracted in step S2;

[0067] S4, connecting the LSTM after information mining in step S3 to the fully connected layer of the prediction model to obtain the prediction result;

[0068] S5. diagnose the fault of the sample using the diagnostic model;

[0069] The dataset D in step S1 is divided into a training set and a test set;

[0070] The specific diagnostic steps in step S5 are as follows:

[0071] S501, using a cost-sensitive algorithm to adjust the weight ratio of the classification in the loss function;

[0072] S502, obtaining pseudo labels of samples in the classified and adjusted data;

[0073] S503, using a conditional adaptive algorithm to calculate the feature distribution difference of the same category in the training set and the test set;

[0074] S504: Reduce the distribution difference calculated in step S503 to complete the diagnosis.

[0075] When the diagnostic model is constructed, it needs to be built before the diagnostic model.

[0076] The preprocessing in step S1 is the segmentation of time series data; in step S1, the dataset D is composed of N sub-datasets consisting of d-dimensional time series samples, and each time series is represented by X i ∈D, the length of each time series is L, where X i ∈R d×L ,i∈N; the sliding window divides each time series into small time windows; the jth sliding window is represented as As the most original input, it is placed on the input side of the model.

[0077] The prediction model also includes the wc-DAN model, which is used to obtain better prediction and migration performance.

[0078] The expression of the wc-DAN model is as follows:

[0079] l (*) =l c (D s ,y s ,θ l )+λMMD(D s , D t ,θ' l ).

[0080] The first item c (D s ,y s ,θl ) represents the loss term of the CNN-LSTM network. The test set input data does not participate in the calculation of the first term. The second term λMMD(D s , D t ,θ' l ) is the distribution difference between the source domain and target domain data under different working conditions; θ l These are the parameters that need to be optimized when minimizing the loss, including the relevant weights and biases of CNN, LSTM, and fully connected layers. The training set and test set features used in the MMD calculation process need to be output by CNN, θ' l Represents the parameters of the CNN part used, and θ l The parameters in the CNN part are the same.

[0081] In order to improve the generalization ability of the prediction model, a domain adaptation algorithm is also used in the prediction process of the prediction model. The goal of the domain adaptation algorithm is to minimize the marginal distribution difference between the training and test data sets, that is, P(X s ) and P(X t ); P(X) is expressed as a mixture of operating condition distributions, and the specific formula is as follows:

[0082]

[0083] α i P(X|Z=C i ) represents the prior probability of the working condition, that is, the data set belongs to working condition C i The proportion of samples; redefine X according to the mixed marginal distribution s and X t MMD between, using MMD C (X s ,X t ) is represented by X s With X t wcMMD between;MMD C (X s ,X t ) can be defined as two modes, the two modes are intra-class migration mode and weighted conditional migration mode.

[0084] The specific steps of step S501 are as follows:

[0085] S5001. Use a cost-sensitive imbalanced strategy based on class weights, so that the misclassification results of the minority class are given greater weight in the calculation of the loss function and the update of parameters, and the weight of the majority class is reduced accordingly;

[0086] The specific steps of step S502 are as follows:

[0087] S5002. Use the conditional distribution domain adaptation algorithm to narrow the conditional distribution difference between the source domain and the target domain, and minimize the distribution difference of features under the same category in the source domain and the target domain; use an iterative pseudo-label algorithm to obtain the category information of the test set and provide pseudo-labels for the calculation of the conditional distribution domain adaptation algorithm.

[0088] The formula of the loss function is as follows:

[0089] l IFDM-DA =l cs (X s , Y s ,θ l )+λMMD CDA (X s , X t , Y s , Y p ,θ′ l )

[0090] l cs (X s , Y s ,θ l ) represents the loss term of the Deep CNN-LSTM network with the cost-sensitive strategy, and the second term λMMD CDA (X s , X t , Y s , Y p ,θ′ l ) is the conditional distribution difference between the source domain and target domain data calculated by category; θ l is the parameter that needs to be optimized when minimizing the loss, λ is the weight parameter of the domain adaptive loss function; θ′ l ∈θ l , because MMD CDA The calculation only involves the network layer before LSTM in the framework, θ′ l Only contains relevant parameters of CNN and LSTM.

[0091] The weight in step S5001 is inversely proportional to the number of samples of the category in the data set.

[0092] By studying the application and optimization of transfer learning and domain distribution adaptation technology under complex working conditions, a related fault prediction model was constructed to solve the problem of poor generalization ability of the training model under the influence of data distribution differences caused by complex working conditions;

[0093] By optimizing the fault diagnosis model's ability to classify fault samples and overcoming the distribution differences caused by different working conditions, the accuracy and reliability of the model can be improved.

[0094] Example 2:

[0095] like Figure 1-4 As shown, a device fault diagnosis and prediction method based on transfer learning includes a prediction model and a diagnosis model;

[0096] The prediction model consists of 4 CNN layers, 2 LSTM layers, and 3 fully connected layers;

[0097] The diagnostic model is composed of 4 layers of CNN, 2 layers of LSTM and three layers of fully connected layers.

[0098] The specific steps are as follows:

[0099] S1. Preprocess the migrated data, and the processed data samples constitute the dataset D;

[0100] S2, using the CNN in the prediction model to extract feature information of the data processed in step S1;

[0101] S3, using the LSTM in the prediction model to mine the data feature information extracted in step S2;

[0102] S4, connecting the LSTM after information mining in step S3 to the fully connected layer of the prediction model to obtain the prediction result;

[0103] S5. diagnose the fault of the sample using the diagnostic model;

[0104] The dataset D in step S1 is divided into a training set and a test set;

[0105] The specific diagnostic steps in step S5 are as follows:

[0106] S501, using a cost-sensitive algorithm to adjust the weight ratio of the classification in the loss function;

[0107] S502, obtaining pseudo labels of samples in the classified and adjusted data;

[0108] S503, using a conditional adaptive algorithm to calculate the feature distribution difference of the same category in the training set and the test set;

[0109] S504: Reduce the distribution difference calculated in step S503 to complete the diagnosis.

[0110] The preprocessing in step S1 is the segmentation of time series data; in step S1, the dataset D is composed of N sub-datasets consisting of d-dimensional time series samples, and each time series is represented by X i ∈D, the length of each time series is L, where X i ∈R d×L ,i∈N; the sliding window divides each time series into small time windows; the jth sliding window is represented as As the most original input, it is placed on the input side of the model.

[0111] The prediction model also includes the wc-DAN model, which is used to obtain better prediction and migration performance.

[0112] The expression of the wc-DAN model is as follows:

[0113] l (*) =l c (D s ,y s ,θ l )+λMMD(D s , D t ,θ' l ).

[0114] The first item c (D s ,y s ,θ l ) represents the loss term of the CNN-LSTM network. The test set input data does not participate in the calculation of the first term. The second term λMMD(D s , D t ,θ' l ) is the distribution difference between the source domain and target domain data under different working conditions; θ l These are the parameters that need to be optimized when minimizing the loss, including the relevant weights and biases of CNN, LSTM, and fully connected layers. The training set and test set features used in the MMD calculation process need to be output by CNN, θ' l Represents the parameters of the CNN part used, and θ l The parameters in the CNN part are the same.

[0115] In order to improve the generalization ability of the prediction model, a domain adaptation algorithm is also used in the prediction process of the prediction model. The goal of the domain adaptation algorithm is to minimize the marginal distribution difference between the training and test data sets, that is, P(X s ) and P(X t ); P(X) is expressed as a mixture of operating condition distributions, and the specific formula is as follows:

[0116]

[0117] α i P(X|Z=C i ) represents the prior probability of the working condition, that is, the data set belongs to working condition C i The proportion of samples; redefine X according to the mixed marginal distribution s and X t MMD between, using MMDC (X s ,X t ) is represented by X s With X t wcMMD between;MMD C (X s ,X t ) can be defined as two modes, the two modes are intra-class migration mode and weighted conditional migration mode.

[0118] The specific steps of step S501 are as follows:

[0119] S5001. Use a cost-sensitive imbalanced strategy based on class weights, so that the misclassification results of the minority class are given greater weight in the calculation of the loss function and the update of parameters, and the weight of the majority class is reduced accordingly;

[0120] The specific steps of step S502 are as follows:

[0121] S5002. Use the conditional distribution domain adaptation algorithm to narrow the conditional distribution difference between the source domain and the target domain, and minimize the distribution difference of features under the same category in the source domain and the target domain; use an iterative pseudo-label algorithm to obtain the category information of the test set and provide pseudo-labels for the calculation of the conditional distribution domain adaptation algorithm.

[0122] The formula of the loss function is as follows:

[0123] l IFDM-DA =l cs (X s , Y s ,θ l )+λMMD CDA (X s , X t , Y s , Y p ,θ′ l )

[0124] l cs (X s , Y s ,θ l ) represents the loss term of the Deep CNN-LSTM network with the cost-sensitive strategy, and the second term λMMD CDA (X s , X t , Y s , Y p ,θ′ l ) is the conditional distribution difference between the source domain and target domain data calculated by category; θ l is the parameter that needs to be optimized when minimizing the loss, λ is the weight parameter of the domain adaptive loss function; θ′ l∈θ l , because MMD CDA The calculation only involves the network layer before LSTM in the framework, θ′ l Only contains relevant parameters of CNN and LSTM.

[0125] The weight in step S5001 is inversely proportional to the number of samples of the category in the data set.

[0126] like Figure 3 As shown in the figure, Deep CNN-LSTM consists of 4 convolutional layers and 2 LSTM layers, and is finally input into 3 fully connected layers.

[0127] Use the convolutional layer to extract features from the original input of the model, which is the original input from the time series X i The time window cut in ∈d

[0128] During training, mini-batch stochastic gradient descent was used, with a batch size of 128. This means that 128 sliding time windows (length l = 32) were fed into the CNN for feature extraction each time. Each node within the time window had 17-dimensional sensor features (the original number of dimensions was 26, after removing the 2-dimensional timestamp indicator and the 3-dimensional working condition; in addition, four sensor features were removed because they were found to remain constant throughout the process). The number of filters in each layer of the stacked multi-layer CNN was 16, 32, 64, and 64, respectively, with a kernel size of 3×1. To ensure the speed of CNN training, faster convergence, and smaller errors, the Xavier weight initialization strategy was used to initialize the CNN parameters. Each CNN layer was followed by a maximum pooling layer and a dropout layer to reduce the number of parameters and improve generalization.

[0129] Original time window data After passing through CNN, it will be mapped to a new feature space and input as a new feature to the next layer of CNN in series. RNNs, especially RNNs based on LSTM units or GRU units, have been successfully applied to the most advanced sequence modeling tasks, such as machine translation. As done in the CNNs part, two layers of LSTM are stacked to mine the time series features of the features, and the number of nodes in each layer is represented as n. r , n in the model r = 32, the output of each unit in the LSTM layer will be connected to the input of the corresponding unit in the next LSTM layer. Finally, the LSTM layer is connected to three fully connected layers to obtain the final prediction result.

[0130] In this network, each layer uses the rectified linear unit as the activation function, and the Deep-CNN-LSTM network uses the root mean square error as the loss function.

[0131] The architecture of IFDM-DA is as follows Figure 2 As shown in the figure, first, the Deep-CNN-LSTM benchmark model is used to diagnose the fault of the sample. The benchmark model is composed of 4 layers of CNN, 2 layers of LSTM and 3 layers of fully connected layers.

[0132] It's important to note that the calculation of the difference in feature distribution between the source and target domains is placed after the LSTM layer. This is because, during the experiment, the boundaries between feature categories extracted by the CNN were unclear, and the differences were too large, resulting in poor transfer performance. Therefore, preliminary experiments have shown that placing the calculation module after the LSTM layer on the dataset used in this chapter achieves more significant results, facilitating experimental progress.

[0133] By studying the application and optimization of transfer learning and domain distribution adaptation technology under complex working conditions, a related fault prediction model was constructed to solve the problem of poor generalization ability of the training model under the influence of data distribution differences caused by complex working conditions;

[0134] By optimizing the fault diagnosis model's ability to classify fault samples and overcoming the distribution differences caused by different working conditions, the accuracy and reliability of the model can be improved.

[0135] By leveraging the domain adaptation technology of transfer learning in a baseline model, a method for measuring the difference in sample characteristics between two operating conditions is proposed. This difference is used as a term in the loss function. During the training process, the difference in the feature distribution of samples under the same operating condition in the training and test sets is continuously narrowed, improving the generalization of the fault prediction model. Experimental results on a commercial modular aviation propulsion system simulation dataset demonstrate the effectiveness and reliability of the model, as well as the correctness of the corresponding parameter exploration strategy.

[0136] Through the unbalanced fault diagnosis model, the cost-sensitive algorithm is used to adjust the weights of different categories in the loss function, thereby increasing the classifier's emphasis on misclassification of fault samples; the pseudo-labels of samples in the test set are obtained using an unsupervised method, and the conditional distribution adaptation algorithm in domain adaptation is used to calculate the feature differences of the same category in the training set and the test set. The size of the difference is used as an item in the loss function. During the training process, the difference between the features of the same category in the training set and the test set is continuously narrowed, thereby improving the generalization of the model and the classification accuracy of minority classes.

[0137] The above embodiments merely represent preferred embodiments of the present invention. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous modifications, improvements, and substitutions without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention.

Claims

1. A device fault diagnosis and prediction method based on transfer learning, characterized in that: Includes predictive and diagnostic models; The prediction model consists of 4 layers of CNN, 2 layers of LSTM and 3 layers of fully connected layers; The diagnostic model is composed of a stack of 4 CNN layers, 2 LSTM layers, and three fully connected layers; The specific steps are as follows: S1. Preprocess the migrated data, and the processed data samples constitute the dataset D; S2, using the CNN in the prediction model to extract feature information of the data processed in step S1; S3, using the LSTM in the prediction model to mine the data feature information extracted in step S2; S4, connecting the LSTM after information mining in step S3 to the fully connected layer of the prediction model to obtain the prediction result; S5. diagnose the fault of the sample using the diagnostic model; The data set D in step S1 is divided into a training set and a test set; The specific diagnostic steps in step S5 are as follows: S501, using a cost-sensitive algorithm to adjust the weight ratio of the classification in the loss function; S502, obtaining pseudo labels of samples in the classified and adjusted data; S503, using a conditional adaptive algorithm to calculate the feature distribution difference of the same category in the training set and the test set; S504, reducing the distribution difference calculated in step S503, and completing the diagnosis; The specific steps of step S501 are as follows: S5001. Use a cost-sensitive imbalanced strategy based on class weights, so that the misclassification results of the minority class are given greater weight in the calculation of the loss function and the update of parameters, and the weight of the majority class is reduced accordingly; The specific steps of step S502 are as follows: S5002: Use the conditional distribution domain adaptation algorithm to narrow the conditional distribution difference between the source domain and the target domain, minimizing the distribution difference of features under the same category in the source and target domains. Use an iterative pseudo-labeling algorithm to obtain category information of the test set and provide pseudo-labels for the calculation of the conditional distribution domain adaptation algorithm. The formula of the loss function is as follows: ； Represents the loss term of the Deep CNN-LSTM network with the cost-sensitive strategy. The second term It is the conditional distribution difference between the source domain and target domain data calculated by category; is the parameter that needs to be optimized when minimizing the loss, and λ is the weight parameter of the domain adaptive loss function; ,because The calculation only involves the network layers before LSTM in the framework. Only contains relevant parameters of CNN and LSTM.

2. The device fault diagnosis and prediction method based on transfer learning according to claim 1, characterized in that: The preprocessing in step S1 is the segmentation of time series data; the data set D in step S1 is composed of N sub-data sets consisting of d-dimensional time series samples, and each time series is represented by , the length of each time series is L, where , ; The sliding window divides each time series into small time windows; the jth sliding window is represented as ; As the most original input, it is placed on the input side of the model.

3. The device fault diagnosis and prediction method based on transfer learning according to claim 2, characterized in that: The prediction model also includes a wc-DAN model, and the wc-DAN model is used to obtain better prediction and migration performance.

4. The method for equipment fault diagnosis and prediction based on transfer learning according to claim 2, characterized in that: The expression of the wc-DAN model is as follows: ； Item 1 Represents the loss term of the CNN-LSTM network. The test set input data does not participate in the calculation of the first term; the second term is the distribution difference between different working conditions between the source domain and target domain data; These are the parameters that need to be optimized when minimizing the loss, including the relevant weights and biases of CNN, LSTM, and fully connected layers. The training set and test set features used in the MMD calculation process need to be output by CNN. Represents the parameters of the CNN part used, and The parameters in the CNN part are the same.

5. The method for equipment fault diagnosis and prediction based on transfer learning according to claim 2, characterized in that: In order to improve the generalization ability of the prediction model, a domain adaptation algorithm is also used in the prediction process of the prediction model. The goal of the domain adaptation algorithm is to minimize the marginal distribution difference between the training and test data sets, that is, and ; It is expressed as a mixture of operating condition distributions, and the specific formula is as follows: ； ； described Represents the prior probability of the working condition, that is, the data set belongs to the working condition The proportion of samples; redefine according to the mixed marginal distribution and MMD between, use Expressed as wcMMD between ; It can be defined as two modes, which are the intra-class migration mode and the weighted conditional migration mode.

6. The method for equipment fault diagnosis and prediction based on transfer learning according to claim 5, characterized in that: The weight in step S5001 is inversely proportional to the number of samples of the category in the data set.

Images