Equipment fault diagnosis network model training method based on multi-source domain knowledge joint migration

Abstract

The invention relates to the technical field of computer deep learning, and relates to an equipment fault diagnosis network model training method based on multi-source domain knowledge joint migration, and the method comprises the steps: constructing a training sample set, a verification set and a test set, and training a pre-constructed equipment fault diagnosis network model through the training sample set, optimizing parameters of the equipment fault diagnosis network model to obtain a trained equipment fault diagnosis network model; verifying the trained equipment fault diagnosis network model by adopting a verification set, evaluating the diagnosis performance of the trained equipment fault diagnosis network model, and selecting an optimal equipment fault diagnosis network model; and testing by adopting the test set optimal equipment fault diagnosis network model, and evaluating the performance of the optimal equipment fault diagnosis network model. According to the method, multi-source-domain cooperative training is carried out on the model by utilizing a collected multi-source-domain data set which is composed of a plurality of labeled source-domain data sets and unlabeled source-domain data sets and contains a small number of labels, domain invariant features in fault samples are learned, and shared diagnosis knowledge is migrated to a target domain from labeled data and unlabeled data together; therefore, the trained equipment fault diagnosis network model can complete a cross-domain fault diagnosis task without obtaining a target sample in advance.

Description

Technical Field

[0001] The present application belongs to the field of computer deep learning technology, and specifically relates to a method for training a network model for equipment fault diagnosis based on joint migration of multi-source domain knowledge. Background Art

[0002] Rolling bearings are key components in twin-rotor aircraft engine systems. They usually work in high-speed, noisy environments and are prone to failure, which can affect the normal operation of the engine and even cause serious accidents. Research on fault diagnosis of aircraft engine bearings helps to accurately diagnose rolling bearing fault characteristics, which is of great significance to improving the safety and reliability of aircraft engines.

[0003] In recent years, research on data-driven bearing fault diagnosis methods has developed rapidly. Different deep learning models such as convolutional neural networks, graph convolutional networks, and long short-term memory networks have been used for equipment fault diagnosis and have achieved efficient performance. However, most methods rely on a large amount of labeled data, assuming that the training and test data are independent and identically distributed, and assuming that the target domain samples exist. However, in actual industrial application scenarios, it is extremely difficult to obtain sufficient labeled data. The working conditions of the equipment are often uncertain, and the samples and fault label information under the target domain working conditions are difficult to obtain. In addition, the data that can be used for training are usually collected from multiple working conditions, which often leads to different distributions of training data (source domain) and test data (target domain). There is an obvious domain shift problem, which greatly reduces the performance and effectiveness of traditional fault diagnosis models and makes it difficult to obtain expected results. In actual operation, the working conditions of aircraft engines will change, resulting in inconsistent data sample distribution. In addition, using only data from a single source domain to train the fault diagnosis model may not achieve ideal diagnostic performance. However, it is unrealistic to comprehensively collect and label enough training samples to make the fault diagnosis model robust to all working conditions.

[0004] It can be seen that cross-domain fault diagnosis of equipment has important practical significance. Being able to learn and classify domain-invariant features is the key to building a cross-domain fault diagnosis model. Facing the challenges brought by unknown domains in cross-domain fault diagnosis tasks, domain adaptation methods and fault diagnosis methods based on domain generalization are usually used to solve the performance degradation problem when facing unknown domain data.

[0005] When target domain data is available, domain adaptation methods can solve the domain shift problem by reducing the data distribution difference between the source domain and the target domain. Liu et al. proposed a deep adversarial domain adaptation model for rolling bearing fault diagnosis to solve the distribution inconsistency problem between the source domain and the target domain. Xiao et al. designed an improved loss function embedded with the joint maximum mean difference to achieve simultaneous alignment of marginal distribution and conditional distribution in unsupervised cross-domain scenarios. However, the transfer learning strategy based on a single source domain may not achieve ideal generalization performance due to the single source of knowledge transfer. The generalization performance of the diagnosis model can be further improved by collecting multi-source domain data in different scenarios and adopting multi-source domain-based transfer learning technology. Therefore, Gao et al. proposed a multi-source domain information fusion network, combining adversarial transfer learning technology with fine-grained information fusion of multi-source domains to achieve feature learning and knowledge transfer of source and target domains; Tian et al. used a multi-branch network structure to match the feature space distribution of each source domain and target domain respectively, and transferred multi-source domain diagnosis knowledge to the cross-domain fault diagnosis task. Domain adaptation methods use the diagnostic knowledge of known data sets to achieve fault diagnosis tasks in the target domain, but this method assumes that data samples in the target domain are accessible during the model training process. At the same time, the trained model can usually only handle fault diagnosis tasks in a single target domain, and it is difficult to effectively migrate to other target domains, and the generalization ability of the target domain is limited. Summary of the invention

[0006] In view of the problems and shortcomings in the prior art, the purpose of the present invention is to provide a method for training a network model for equipment fault diagnosis based on joint migration of multi-source domain knowledge.

[0007] To achieve the purpose of the invention, the technical solution adopted by the present invention is as follows:

[0008] A first aspect of the present invention provides a method for training a network model for equipment fault diagnosis, comprising the following steps:

[0009] S1: Collect signal data representing equipment failures during equipment operation under different working conditions, and use the signal data to construct multiple labeled source domain data sets, unlabeled source domain data sets, and target domain data sets according to the working conditions. The equipment failure categories represented by the signal data in each labeled source domain data set, unlabeled source domain data set, and target domain data set include all failure categories. Any two labeled source domain data sets correspond to different working conditions, and any two unlabeled source domain data sets correspond to different working conditions. Label each signal data in the labeled source domain data set and the target domain data set to obtain the true label information of the equipment failure category corresponding to each signal data. The signal data in the unlabeled source domain data set is not labeled.

[0010] S2: extracting the same number of signal data with the same fault category from all labeled source domain data sets and unlabeled source domain data sets in turn to form a batch of training samples, constructing multiple batches of training samples according to the construction method of the batch of training samples, and the multiple batches of training samples constitute a training sample set; extracting the same number of signal data from all target domain data sets in turn to form a batch of verification samples, constructing multiple batches of verification samples according to the construction method of the batch of verification samples, and randomly dividing the multiple batches of training samples into verification sets and test sets according to proportion;

[0011] S3: using the training sample set to train the pre-built equipment fault diagnosis network model, optimizing the parameters of the equipment fault diagnosis network model, and obtaining a trained equipment fault diagnosis network model;

[0012] S4: using the validation set to validate the trained equipment fault diagnosis network model obtained in step S3, evaluating the diagnostic performance of the trained equipment fault diagnosis network model, and selecting the optimal equipment fault diagnosis network model;

[0013] S5: Use the test set to test the optimal equipment fault diagnosis network model obtained in step S4 to evaluate the performance of the optimal equipment fault diagnosis network model.

[0014] According to the above-mentioned training method of the equipment fault diagnosis network model, preferably, the equipment fault diagnosis network model is composed of a feature extractor and a fault classifier, the feature extractor is used to extract the equipment fault features contained in the signal data, and the fault classifier is used to learn and analyze the equipment fault features extracted by the feature extractor, predict the equipment fault category corresponding to the signal data and output it; the feature extractor includes a first Cov1D layer, a second Cov1D layer, a third Cov1D layer, a first fully connected layer and a second fully connected layer connected in sequence. Among them, the first Cov1D layer is used to perform feature extraction (convolution, batch normalization, maximum pooling) on ​​the input data signal sample to obtain the first-level feature map of the data signal sample. The input of the first Cov1D layer is the data signal sample, and the output of the first Cov1D layer is the first-level feature map; the second Cov1D layer is used to perform feature extraction on the first-level feature map output by the first Cov1D layer to obtain and output the second-level feature map; the third Cov1D layer is used to perform feature extraction on the second-level feature map output by the second Cov1D layer to obtain and output the third-level feature map; the first fully connected layer of the feature extractor is used to learn the nonlinear combination between the features of the three-level feature map and integrate global information. The input of the first fully connected layer of the feature extractor is the flattened three-level feature map, and the output of the first fully connected layer of the feature extractor is the one-dimensional feature vector of the three-level feature map; the second fully connected layer of the feature extractor is used to map the one-dimensional feature vector output by the first fully connected layer to a low-dimensional feature space to generate a feature embedding vector.

[0015] According to the above-mentioned training method of the equipment fault diagnosis network model, preferably, the network structure of the first Cov1D layer, the second Cov1D layer, and the third Cov1D layer in the feature extractor is the same, and is composed of a convolutional layer, a batch normalization layer, and a maximum pooling layer connected in sequence.

[0016] According to the above-mentioned training method of the equipment fault diagnosis network model, preferably, the fault classifier is composed of a first fully connected layer, a second fully connected layer and a Softmax output layer which are connected in sequence. The first fully connected layer of the fault classifier is used to combine and transform the feature embedding vector output by the second fully connected layer of the feature extractor to extract higher-level features. The input of the first fully connected layer of the fault classifier is the feature embedding vector output by the second fully connected layer of the feature extractor, and the output of the first fully connected layer of the fault classifier is a low-dimensional feature vector. The second fully connected layer of the fault classifier is used to map the low-dimensional feature vector output by the first fully connected layer of the fault classifier to a low-dimensional feature space to obtain a more refined and discriminative feature representation. The input of the second fully connected layer of the fault classifier is the low-dimensional feature vector output by the first fully connected layer of the fault classifier, and the output of the second fully connected layer of the fault classifier is a fault discrimination feature vector. The Softmax output layer converts the fault discrimination feature vector output by the second fully connected layer of the fault classifier into a probability distribution of fault categories through a Softmax function. The output of the Softmax output layer is a probability distribution vector of fault categories. Each element in the probability distribution feature vector corresponds to a probability value of a fault category. The fault classifier selects the category with the highest probability as the predicted fault category and outputs it to achieve the fault diagnosis task.

[0017] According to the above-mentioned training method of the equipment fault diagnosis network model, preferably, in step S3, the pre-built equipment fault diagnosis network model is trained using the training sample set, and the specific operation of optimizing the parameters of the equipment fault diagnosis network model is:

[0018] S301: Input the batch training samples in the training sample set into the feature extractor of the pre-built equipment fault diagnosis network model, use the feature extractor to extract features from the batch training samples, obtain the equipment fault features of each signal data in the batch training samples, use the maximum mean difference to calculate the marginal distribution difference between the equipment fault features of signal data from different source domain data sets in a batch training sample, use the maximum mean difference to calculate the conditional distribution difference between signal data samples with the same equipment fault category from different labeled source domain data sets in the training sample, and construct the multi-source domain joint loss function L of the feature extractor according to the marginal distribution difference and the conditional distribution difference f ;

[0019] S302: Input the device fault features of the signal data samples from the labeled source domain data set in the batch training samples extracted by the feature extractor into the fault classifier for detection, obtain the predicted device fault category of the signal data samples from the labeled source domain data set in the batch training samples, and construct the supervised classification loss function L of the fault classifier using cross entropy according to the difference between the predicted device fault category and the true label. y ; Then according to the multi-source domain joint loss function L f And supervised classification loss function L y Construct the total loss function L of the equipment fault diagnosis network model all , the total loss function L all The calculation formula is as follows:

[0020] L all =L y +λ 2 L f

[0021] Where, L all Represents the total loss function, L y represents the supervised classification loss function, L f represents the joint loss function of multiple source domains, λ 2 It is the empirical parameter used in model training;

[0022] S303: Based on the total loss function L all The network parameters of the feature extractor are optimized and updated by the back propagation algorithm until the total loss function L all Convergence; based on supervised classification loss function L y The network parameters of the fault classifier are optimized and updated by the back propagation algorithm until the supervised classification loss function L y After convergence, the trained equipment fault diagnosis network model is obtained.

[0023] According to the above-mentioned training method of the equipment fault diagnosis network model, preferably, in step S301, the multi-source domain joint loss function L f The calculation formula is as follows:

[0024]

[0025] Where N is the number of source domain datasets involved in the model training process; M is the number of labeled source domain datasets; N c is the number of fault categories in the labeled source domain; MMD (m) is the marginal distribution difference measure, MMD (c) is the conditional distribution difference measure corresponding to the c-th fault category; Z k and Z pThey represent the deep feature distribution fault feature sets corresponding to the kth labeled source domain or unlabeled source domain and the pth labeled source domain or unlabeled source domain respectively; and denote the fault category feature set corresponding to the cth fault category in the kth labeled source domain and the pth labeled source domain respectively; λ 1 is the empirical parameter corresponding to the conditional distribution difference metric; X and Y correspond to different source domain data sets in the batch training samples, respectively. x i is the i-th sample of the source domain dataset X; y j is the j-th sample of the source domain dataset Y; m represents the number of samples in the source domain dataset X; n represents the number of samples in the source domain dataset Y; k(·,·) is the kernel function.

[0026] According to the above-mentioned training method of the equipment fault diagnosis network model, preferably, in step S302, the supervised classification loss function L y The calculation formula is as follows:

[0027]

[0028] Where, L y represents the supervised classification loss function; o s represents the device fault features extracted from the labeled source domain data samples; y s represents the true label of the fault category corresponding to the signal data sample from the labeled source domain dataset; n s represents the number of samples in a batch of labeled source domain datasets; C represents the number of equipment fault categories; I[·] represents the probability score of the fault category of the signal data sample output by the SoftMax output layer.

[0029] According to the above-mentioned training method of the equipment fault diagnosis network model, preferably, in step S303, the network parameter θ of the feature extractor is optimized and updated. f The formula is as follows:

[0030]

[0031] In the formula, θ f is the parameter of the feature extractor, L y represents the supervised classification loss function, ε is the learning rate of the network parameters, L f represents the joint loss function of multiple source domains, λ 2 It is the empirical parameter used in model training.

[0032] According to the above-mentioned training method of the equipment fault diagnosis network model, preferably, in step S303, the network parameter θ of the fault classifier is optimized and updated. yThe formula is as follows:

[0033]

[0034] In the formula, θ y is the parameter of the fault classifier, ε is the learning rate of the network parameters, L y represents the supervised classification loss function.

[0035] The second aspect of the present invention provides a method for diagnosing equipment faults, specifically: obtaining signal data characterizing equipment faults during equipment operation, inputting the signal data into an equipment fault diagnosis network model for detection, and obtaining fault category information of the equipment; wherein the equipment fault diagnosis network model is an equipment fault diagnosis network model trained using the training method described in the first aspect above.

[0036] The third aspect of the present invention provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the training method as described in the first aspect or the device fault diagnosis method as described in the second aspect.

[0037] The fourth aspect of the present invention provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the training method described in the first aspect or the equipment fault diagnosis method described in the second aspect.

[0038] Compared with the prior art, the present invention has the following positive effects:

[0039] (1) Based on the multi-source domain transfer learning strategy, the present invention utilizes the collected multi-source domain datasets with a small amount of labels composed of several labeled source domain datasets and unlabeled source domain datasets to perform multi-source domain collaborative training on the model; moreover, in order to learn domain-invariant features, a multi-source domain joint difference loss function based on the maximum mean difference is designed, which reduces the marginal distribution differences and conditional distribution differences between different domains, minimizes domain-related information, enhances the domain invariance of learning features, and effectively improves the domain generalization performance of the model. At the same time, the multi-source domain joint difference loss is combined with the supervised classification loss to jointly update the network parameters during the model training process; the domain-invariant features in the fault samples are learned, and the shared diagnostic knowledge is jointly transferred from the labeled data and the unlabeled data to the target domain. Therefore, the trained equipment fault diagnosis network model does not need to obtain the target samples in advance to complete the cross-domain fault diagnosis task.

[0040] (2) The equipment fault diagnosis method of the present invention ensures good diagnostic performance and domain generalization performance when facing unknown operating condition samples; moreover, in a specific embodiment, the present invention uses an aircraft engine inter-shaft bearing fault data set to establish multiple groups of experiments under different cross-operating condition scenarios, and verifies the effectiveness of the diagnostic method of the present invention in the cross-operating condition fault diagnosis performance of aviation components through comparison with other methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A flow chart of a method for training a network model for equipment fault diagnosis according to the present invention;

[0042] Figure 2 A schematic diagram of the network structure of the equipment fault diagnosis network model of the present invention;

[0043] Figure 3 A schematic diagram of the network structure of a feature extractor in a network model for equipment fault diagnosis of the present invention;

[0044] Figure 4 The diagram is a schematic diagram of the network structure of the fault classifier in the equipment fault diagnosis network model of the present invention. DETAILED DESCRIPTION

[0045] The present invention is further described in detail below through specific examples, but the scope of the present invention is not limited thereto.

[0046] Embodiment 1:

[0047] A training method for a network model of equipment fault diagnosis (such as Figure 1 As shown), comprising the following steps:

[0048] S1: Collect signal data representing equipment failures during equipment operation under different working conditions, and use the signal data to construct multiple labeled source domain data sets, unlabeled source domain data sets, and target domain data sets according to the working conditions. The equipment failure categories represented by the signal data in each labeled source domain data set, unlabeled source domain data set, and target domain data set include all fault classifications. Any two labeled source domain data sets correspond to different working conditions, and any two unlabeled source domain data sets correspond to different working conditions. Label each signal data in the labeled source domain data set and the target domain data set to obtain the true label information of the equipment failure category corresponding to each signal data. The signal data in the unlabeled source domain data set are not labeled.

[0049] S2: Extract the same number of signal data with the same fault category from all labeled source domain data sets and unlabeled source domain data sets in turn to form a batch of training samples, and construct multiple batches of training samples according to the construction method of the batch of training samples, and the multiple batches of training samples constitute a training sample set; extract the same number of signal data from all target domain data sets in turn to form a batch of verification samples, and construct multiple batches of verification samples according to the construction method of the batch of verification samples, and randomly divide the multiple batches of training samples into verification sets and test sets according to proportion.

[0050] S3: Using the training sample set to train the pre-built equipment fault diagnosis network model, optimizing the parameters of the equipment fault diagnosis network model, and obtaining the trained equipment fault diagnosis network model.

[0051] S4: Use the validation set to validate the trained equipment fault diagnosis network model obtained in step S3, evaluate the diagnostic performance of the trained equipment fault diagnosis network model, and select the optimal equipment fault diagnosis network model.

[0052] S5: Use the test set to test the optimal equipment fault diagnosis network model obtained in step S4 to evaluate the performance of the optimal equipment fault diagnosis network model.

[0053] As a preferred implementation, in the above step S1, when constructing multiple labeled source domain data sets, unlabeled source domain data sets, and target domain data sets using signal data according to working conditions, sample segmentation and frequency domain conversion can be performed on the signal data samples in the labeled source domain data sets, the unlabeled source domain data sets, and the target domain data sets to obtain more signal data samples.

[0054] As a preferred implementation, the equipment fault diagnosis network model (such as Figure 2 The feature extractor (as shown) is composed of a feature extractor and a fault classifier. The feature extractor is used to extract the equipment fault features contained in the signal data. The fault classifier is used to learn and analyze the equipment fault features extracted by the feature extractor, predict the equipment fault category corresponding to the signal data and output it. Figure 3The network structure of the first Cov1D layer, the second Cov1D layer, and the third Cov1D layer in the feature extractor is the same, and they are all composed of a convolutional layer, a batch normalization layer, and a maximum pooling layer connected in sequence. Among them, the first Cov1D layer is used to perform preliminary feature extraction on the input data signal sample, extract short-term features in the signal data sample and suppress high-frequency noise, so as to obtain a first-level feature map of the data signal sample. The input of the first Cov1D layer is the data signal sample, and the output of the first Cov1D layer is the first-level feature map; the second Cov1D layer is used to perform feature extraction on the first-level feature map output by the first Cov1D layer, extract deeper local features, obtain a second-level feature map and output it; the third Cov1D layer is used to perform feature extraction on the second-level feature map output by the second Cov1D layer, so as to further extract fault features in the signal data sample, obtain a third-level feature map and output it; the first fully connected layer of the feature extractor is used to learn the nonlinear combination between the features of the three-level feature map and integrate global information. The input of the first fully connected layer of the feature extractor is the flattened three-level feature map, and the output of the first fully connected layer of the feature extractor is the one-dimensional feature vector of the three-level feature map; the second fully connected layer of the feature extractor is used to map the one-dimensional feature vector output by the first fully connected layer to a low-dimensional feature space to generate a feature embedding vector.

[0055] The fault classifier (such as Figure 4 The fault classifier is composed of a first fully connected layer, a second fully connected layer and a Softmax output layer connected in sequence; the first fully connected layer of the fault classifier is used to combine and transform the feature embedding vector output by the second fully connected layer of the feature extractor to extract higher-level features, the input of the first fully connected layer of the fault classifier is the feature embedding vector output by the second fully connected layer of the feature extractor, and the output of the first fully connected layer of the fault classifier is a low-dimensional feature vector; the second fully connected layer of the fault classifier is used to map the low-dimensional feature vector output by the first fully connected layer of the fault classifier to a low-dimensional feature space to obtain a more refined and discriminative feature. The input of the second fully connected layer of the fault classifier is the low-dimensional feature vector output by the first fully connected layer of the fault classifier, and the output of the second fully connected layer of the fault classifier is the fault discrimination feature vector; the Softmax output layer converts the fault discrimination feature vector output by the second fully connected layer of the fault classifier into the probability distribution of the fault category through the Softmax function, and the output of the Softmax output layer is the probability distribution vector of the fault category, each element in the probability distribution feature vector corresponds to the probability value of a fault category, and the fault classifier will select the category with the highest probability as the predicted fault category and output it to achieve the fault diagnosis task.

[0056] As a preferred implementation, in the above step S3, the specific operation of training the pre-built equipment fault diagnosis network model using the training sample set is:

[0057] S301: Input a batch of training samples in the training sample set into the feature extractor of the pre-built equipment fault diagnosis network model, use the feature extractor to extract features from the batch of training samples, obtain the equipment fault features of each signal data in the batch of training samples, use the maximum mean difference to calculate the marginal distribution difference between the equipment fault features of signal data from different source domain data sets in a batch of training samples, use the maximum mean difference to calculate the conditional distribution difference between signal data samples with the same equipment fault category from different labeled source domain data sets in the training samples, and construct a multi-source domain joint loss function L of the feature extractor according to the marginal distribution difference and the conditional distribution difference f .

[0058] The multi-source domain joint loss function L f The calculation formula is as follows:

[0059]

[0060] Where N is the number of source domain datasets involved in the model training process; M is the number of labeled source domain datasets; N c is the number of fault categories in the labeled source domain; MMD (m) is the marginal distribution difference measure, MMD (c) is the conditional distribution difference measure corresponding to the c-th fault category; Z k and Z p They represent the deep feature distribution fault feature sets corresponding to the k-th labeled source domain or unlabeled source domain and the p-th labeled source domain or unlabeled source domain respectively; and denote the fault category feature set corresponding to the cth fault category in the kth labeled source domain and the pth labeled source domain respectively; λ 1 is the empirical parameter corresponding to the conditional distribution difference metric; X and Y correspond to different source domain data sets in the batch training samples, respectively. x i is the i-th sample of the source domain dataset X; y j is the j-th sample of the source domain dataset Y; m represents the number of samples in the source domain dataset X; n represents the number of samples in the source domain dataset Y; k(·,·) is the kernel function.

[0061] S302: Input the device fault features of the signal data samples from the labeled source domain data set in the batch training samples extracted by the feature extractor into the fault classifier for detection, obtain the predicted device fault category of the signal data samples from the labeled source domain data set in the batch training samples, and construct the supervised classification loss function L of the fault classifier using cross entropy according to the difference between the predicted device fault category and the true label. y ; Then according to the multi-source domain joint loss function L f And supervised classification loss function L y Construct the total loss function L of the equipment fault diagnosis network model all .

[0062] The supervised classification loss function L y The calculation formula is as follows:

[0063]

[0064] Where, L y represents the supervised classification loss function; o s represents the device fault features extracted from the labeled source domain data samples; y s represents the true label of the fault category corresponding to the signal data sample from the labeled source domain dataset; n s represents the number of samples in a batch of labeled source domain datasets; C represents the number of equipment fault categories; I[·] represents the probability score of the fault category of the signal data sample output by the SoftMax output layer.

[0065] The total loss function L all The calculation formula is as follows:

[0066] L all =L y +λ 2 L f

[0067] Where, L all Represents the total loss function, L y represents the supervised classification loss function, L f represents the joint loss function of multiple source domains, λ 2 It is the empirical parameter used in model training.

[0068] S303: Based on the total loss function L all The network parameters of the feature extractor are optimized and updated by the back propagation algorithm until the total loss function L all Convergence; based on supervised classification loss function L y The network parameters of the fault classifier are optimized and updated by the back propagation algorithm until the supervised classification loss function Ly Converge, and obtain a trained equipment fault diagnosis network model. Among them, the network parameters θ of the feature extractor are optimized and updated. f , network parameters θ of the fault classifier y The formula is as follows:

[0069]

[0070] In the formula, θ f is the parameter of the feature extractor, L y represents the supervised classification loss function, ε is the learning rate of the network parameters, L f represents the joint loss function of multiple source domains, λ 2 is the empirical parameter used in model training, θ y is the parameter of the fault classifier, ε is the learning rate of the network parameters, L y represents the supervised classification loss function.

[0071] S304: Repeat the above steps S301-S303 until all batches of samples in the training sample set are traversed to obtain multiple trained equipment fault diagnosis network models.

[0072] Embodiment 2:

[0073] The content of Example 2 is basically the same as that of Example 1, except that:

[0074] In this embodiment, the equipment targeted for fault diagnosis is the rolling bearing of an aircraft engine. The signal data for the training sample set, validation set, and test set are all collected from the vibration signal data of a real aircraft engine prototype under variable speed conditions. In order to obtain more training samples, the vibration signal data collected under different speed conditions can be further divided into samples and converted into the frequency domain by equal window division. The rolling bearing fault categories corresponding to the collected vibration signal data samples include three types: 0-healthy; 1-inner ring fault, 2-outer ring fault. The above three fault categories are used as fault labels to annotate each vibration signal data in the labeled source domain data set and the target domain data set, and the real label information of each vibration signal data in the labeled source domain data set and the target domain data set is obtained.

[0075] In a specific embodiment, in order to collect vibration signal data under different speed conditions of an aircraft engine, the present invention sets 6 detection points at different positions of the aircraft engine, and the corresponding sensors are numbered #1 to #6, of which 2 displacement sensors (#1, #2) are used to measure the vibration displacement response, and 4 acceleration sensors (#3, #4, #5, #6) are used to measure the vibration acceleration response, and the sampling frequency of the sensor is 25000Hz. By artificially adding crack faults to the outer ring and inner ring of the bearing, including outer ring faults (depth and length of 0.5mm) and bearing inner ring faults (depth of 0.5mm, length of 0.5mm or 1.0mm). Tests were carried out and data samples were collected under 28 different low and high pressure rotor speed conditions. The information of the vibration signal data sample set obtained after data preprocessing is shown in Table 1 below. According to the collected vibration signal data sample set, a labeled source domain data set, an unlabeled source domain data set and a target domain data set are divided, and the specific division results are shown in Table 2.

[0076] Table 1 Detailed information of the sample set of aircraft engine vibration signal data collected in the experiment

[0077]

[0078] Table 2 Division results of labeled source domain dataset, unlabeled source domain dataset and target domain dataset

[0079]

[0080] Moreover, when used for the fault diagnosis of rolling bearings of aircraft engines, the network layer parameter settings of the feature extractor and fault classifier of the equipment fault diagnosis network model are specifically shown in Table 3. When training, the relevant parameters selected are as follows: the sample length of the input data is 1024, the number of fault categories N c =3, the maximum number of iterations of network training N tra =50,λ 1 =0.8,λ 2 =2,∈=0.0005.

[0081] Table 3 Network layer parameter settings of feature extractor and fault classifier of equipment fault diagnosis network model

[0082]

[0083] In order to evaluate the effectiveness of the training method of the equipment fault diagnosis network model in Example 2 of the present invention, the following methods are used to compare with the training method of the present invention:

[0084] (1) Baseline method: In this method, only a single labeled source domain dataset is used to train the model, and the training process adopts the supervised classification loss Ly As the loss function, the network parameters of the feature extractor and the fault classifier are optimized and updated by the back propagation algorithm until the supervised classification loss function L y After convergence, the trained equipment fault diagnosis network model is obtained; this training method does not include any transfer learning strategy, and then the target task is directly diagnosed to obtain the final diagnosis result.

[0085] (2) Comparative Method 1: Multi-source domain learning method without transfer knowledge. Considering the superiority of multi-source domain training, multiple labeled source domain datasets are used to co-train the model. Comparative Method 1 deletes the multi-source domain joint difference loss used to measure the distribution difference between domains, which is used to verify the effectiveness of the multi-source domain joint difference loss in the proposed method. Specifically, multiple labeled source domain datasets are used to co-train the model, and the training process only uses the supervised classification loss L y As the loss function, the network parameters of the fault classifier are optimized and updated through the back propagation algorithm until the supervised classification loss function L y After convergence, the trained equipment fault diagnosis network model is obtained.

[0086] (3) Comparative Method 2: Multi-source Domain Transfer Learning Method. Based on Comparative Method 1, the multi-source domain joint difference loss and supervised classification loss are used as the total loss to optimize the model parameters during the training process. The transfer learning strategy is used to solve the domain difference problem to handle unknown target tasks. Specifically, multiple labeled source domain datasets are used to train the model collaboratively. The multi-source domain joint loss L is used in the training process. f and supervised classification loss L y Construct the total loss function L all , the network parameters of the fault classifier are optimized and updated through the back propagation algorithm until the total loss function L all After convergence, the trained equipment fault diagnosis network model is obtained.

[0087] The equipment fault diagnosis network model obtained by each training method was tested in each experimental scenario of the target task cases corresponding to the four target domain datasets listed in Table 2. The diagnostic accuracy of these equipment fault diagnosis network models is shown in Table 4. For the baseline method, each labeled source domain dataset is used alone for network training. Table 4 shows the accuracy of the models obtained after different single source domain training on the target task (two: respectively represent the accuracy obtained on the target task after the model is trained alone with labeled source domain dataset 1; the accuracy obtained on the target task after the model is trained alone with labeled source domain dataset 2).

[0088] Table 4 Accuracy of diagnostic results of equipment fault diagnosis network models obtained by different training methods (%)

[0089]

[0090] As can be seen from Table 4, compared with other methods, the model training method of the present invention has higher diagnostic accuracy and smaller fluctuation range in all four cases, and can more effectively diagnose unknown target tasks. Specifically, in Case 1, when only a single labeled source domain dataset B is used for training in the baseline method, the diagnostic accuracy of the two target tasks is 74.44% and 91.11%, and when only the labeled source domain dataset C is used for training, the accuracy is 84.44% and 66.67%, which is low and the diagnostic accuracy between the two target tasks is very different. Although comparison method 1 and comparison method 2 obtain the same diagnostic accuracy as the proposed method on target task 1, both are 100%. However, the diagnostic accuracy for target task 2 is only 77.78% and 69.44%, respectively, which is significantly lower than that of the model training method of the present invention, and there is a significant difference in the diagnostic accuracy between different target tasks. The accuracy of the model training method of the present invention in both target tasks reached 100%, exceeding other comparison methods.

[0091] The baseline method only uses a single labeled source domain dataset for training. Due to the lack of diverse knowledge, this method has poor diagnostic performance when there are obvious differences between the target domain and the source domain, and leads to an increase in the difference in diagnostic accuracy between the two target tasks. Comparative method 1 uses multiple labeled source domain datasets to train the model and obtains better diagnostic performance. Comparative method 2 integrates feature information and diagnostic knowledge through a transfer learning strategy based on multi-source domain data to further improve the generalization performance of the model. In some diagnostic tasks, it shows similar diagnostic accuracy to the proposed method, but there are significant differences in diagnostic accuracy between different target tasks, indicating that the generalization performance of the model is poor. The model training method proposed in the present invention updates the network parameters through the multi-source domain joint difference loss and the supervised classification loss, reducing the marginal distribution difference and conditional distribution difference between domains. This enhances the learning of domain-invariant features, shares diagnostic knowledge in multiple source domain datasets, and obtains good diagnostic performance in different target tasks.

[0092] In Case 4, the source and target domain datasets contain data samples collected from sensors at different measurement points. In the baseline method, when only a single labeled source domain dataset A is used for training, the diagnostic accuracy of the two target tasks is only 33.33%. The diagnostic accuracy of the comparison method 1 for the two target tasks is 100% and 69.44%, respectively. The comparison method 2 optimizes the domain boundary, and the accuracy rates for the same tasks are 100% and 78.33%, respectively, showing better performance. The model training method of the present invention has significant advantages over other comparison methods with poor performance, with accuracy rates of 96.67% and 100%, respectively, both greater than 96%, and the accuracy of the diagnostic results on different target tasks is not much different. The results show that the model training method of the present invention has good domain-invariant feature extraction capabilities for fault samples through training with limited and partially labeled multi-source domain datasets (including labeled source domain datasets and unlabeled source domain datasets), and can be extended to unknown target domains, with good diagnostic accuracy and domain generalization performance.

[0093] In the four groups of experiments, the diagnostic accuracy of the model training method of the present invention reached more than 96%, which met the expected results of the target task, and fully proved that the model training method of the present invention can still accurately and effectively diagnose samples when the target domain samples cannot be obtained for network training. In summary, the model training method of the present invention has a good ability to extract domain-invariant features, effectively solves the problem of difficulty in diagnosing target task fault samples due to the lack of complete training samples, is more suitable for fault diagnosis in complex cross-operating scenarios, and has broad application prospects in actual industrial fault diagnosis.

[0094] Embodiment 3:

[0095] A method for diagnosing equipment faults, specifically comprising: obtaining signal data characterizing equipment faults during equipment operation, inputting the signal data into an equipment fault diagnosis network model for detection, and obtaining fault category information of the equipment; wherein the equipment fault diagnosis network model is an equipment fault diagnosis network model trained using the training method described in Example 1.

[0096] Embodiment 4:

[0097] A method for diagnosing aircraft engine bearing faults, specifically comprising: obtaining vibration signal data of the engine during operation of the aircraft engine, inputting the vibration signal data into an equipment fault diagnosis network model for detection, and obtaining fault category information of the aircraft engine bearing; wherein the equipment fault diagnosis network model is an equipment fault diagnosis network model trained using the training method described in Example 2.

[0098] Embodiment 5:

[0099] An electronic device comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the training method described in Example 1 or Example 2, or the device fault diagnosis method described in Example 3 or Example 4.

[0100] Embodiment 6:

[0101] A computer-readable storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, the training method described in Example 1 or Example 2, or the equipment fault diagnosis method described in Example 3 or Example 4 is implemented.

[0102] Finally, it should be noted that the above embodiments are only preferred embodiments of the present invention, and are not intended to limit the present invention in other forms. Any technician familiar with the profession may use the above technical content as inspiration to make changes or modifications. This is an equivalent embodiment of equivalent changes. However, any simple modification, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the technical concept of the present invention still fall within the scope of protection of the claims of the present invention.

Claims

1. A training method for a device fault diagnosis network model, characterized in that: The following steps are involved: S1: Collect signal data representing equipment failures during equipment operation under different working conditions, and use the signal data to construct multiple labeled source domain data sets, unlabeled source domain data sets, and target domain data sets according to the working conditions. The equipment failure categories represented by the signal data in each labeled source domain data set, unlabeled source domain data set, and target domain data set include all fault classifications. Any two labeled source domain data sets correspond to different working conditions, and any two unlabeled source domain data sets correspond to different working conditions. Label each signal data in the labeled source domain data set and the target domain data set to obtain the true label information of the equipment failure category corresponding to each signal data. The signal data in the unlabeled source domain dataset is not labeled; S2: extracting the same number of signal data with the same fault category from all labeled source domain data sets and unlabeled source domain data sets in sequence to form a batch of training samples, and constructing multiple batches of training samples according to the construction method of the batch of training samples, and the multiple batches of training samples constitute a training sample set; Extracting the same amount of signal data from all target domain data sets in turn to form a batch of validation samples, constructing multiple batches of validation samples according to the construction method of the batch of validation samples, and randomly dividing the multiple batches of training samples into validation sets and test sets in proportion; S3: using the training sample set to train the pre-built equipment fault diagnosis network model, optimizing the parameters of the equipment fault diagnosis network model, and obtaining a trained equipment fault diagnosis network model; S4: using the validation set to validate the trained equipment fault diagnosis network model obtained in step S3, evaluating the diagnostic performance of the trained equipment fault diagnosis network model, and selecting the optimal equipment fault diagnosis network model; S5: Use the test set to test the optimal equipment fault diagnosis network model obtained in step S4 to evaluate the performance of the optimal equipment fault diagnosis network model.

2. The training method according to claim 1, characterized in that: The equipment fault diagnosis network model is composed of a feature extractor and a fault classifier. The feature extractor is used to extract the equipment fault features contained in the signal data, and the fault classifier is used to learn and analyze the equipment fault features extracted by the feature extractor, predict the equipment fault category corresponding to the signal data and output it; wherein the feature extractor includes a first Cov1D layer, a second Cov1D layer, a third Cov1D layer, a first fully connected layer and a second fully connected layer connected in sequence.

3. The training method according to claim 2, characterized in that: The network structures of the first Cov1D layer, the second Cov1D layer, and the third Cov1D layer in the feature extractor are the same, and are all composed of a convolutional layer, a batch normalization layer, and a maximum pooling layer connected in sequence.

4. The training method according to claim 3, characterized in that: The fault classifier consists of a first fully connected layer, a second fully connected layer and a Softmax output layer which are connected in sequence.

5. The training method according to claim 1, characterized in that: In step S3, the pre-built equipment fault diagnosis network model is trained using the training sample set, and the specific operation of optimizing the parameters of the equipment fault diagnosis network model is: S301: Input the batch training samples in the training sample set into the feature extractor of the pre-built equipment fault diagnosis network model, use the feature extractor to extract features from the batch training samples, obtain the equipment fault features of each signal data in the batch training samples, use the maximum mean difference to calculate the marginal distribution difference between the equipment fault features of signal data from different source domain data sets in a batch training sample, use the maximum mean difference to calculate the conditional distribution difference between signal data samples with the same equipment fault category from different labeled source domain data sets in the training sample, and construct the multi-source domain joint loss function L of the feature extractor according to the marginal distribution difference and the conditional distribution difference f ; S302: Input the device fault features of the signal data samples from the labeled source domain data set in the batch training samples extracted by the feature extractor into the fault classifier for detection, obtain the predicted device fault category of the signal data samples from the labeled source domain data set in the batch training samples, and construct the supervised classification loss function L of the fault classifier using cross entropy according to the difference between the predicted device fault category and the true label. y ; Then according to the multi-source domain joint loss function L f And supervised classification loss function L y Construct the total loss function L of the equipment fault diagnosis network model all , the total loss function L all The calculation formula is as follows: L all =L y +λ2L f Where, L all Represents the total loss function, L y represents the supervised classification loss function, L f represents the multi-source domain joint loss function, λ2 is the empirical parameter used in model training; S303: Based on the total loss function L all The network parameters of the feature extractor are optimized and updated by the back propagation algorithm until the total loss function L all Convergence; based on supervised classification loss function L y The network parameters of the fault classifier are optimized and updated by the back propagation algorithm until the supervised classification loss function L y After convergence, the trained equipment fault diagnosis network model is obtained.

6. The training method according to claim 5, characterized in that: In step S301, the multi-source domain joint loss function L f The calculation formula is as follows: Where N is the number of source domain datasets involved in the model training process; M is the number of labeled source domain datasets; N c is the number of fault categories in the labeled source domain; MMD (m) is the marginal distribution difference, MMD (c) is the conditional distribution difference corresponding to the c-th fault category; Z k and Z p They represent the deep feature distribution fault feature sets corresponding to the kth labeled source domain or unlabeled source domain and the pth labeled source domain or unlabeled source domain respectively; and denote the fault category feature sets corresponding to the cth fault category in the kth labeled source domain and the pth labeled source domain respectively; λ1 is the empirical parameter corresponding to the conditional distribution difference measure; X and Y correspond to different source domain data sets in the batch training samples respectively, x i is the i-th sample of the source domain dataset X; y j is the j-th sample of the source domain dataset Y; m represents the number of samples in the source domain dataset X; n represents the number of samples in the source domain dataset Y; k(·,·) is the kernel function.

7. The training method according to claim 5, characterized in that: In step S302, the supervised classification loss function L y The calculation formula is as follows: Where, L y represents the supervised classification loss function; o s represents the device fault features extracted from the labeled source domain data samples; y s represents the true label of the fault category corresponding to the signal data sample from the labeled source domain dataset; n s represents the number of samples in a batch of labeled source domain datasets; C represents the number of equipment fault categories; I[·] represents the probability score of the fault category of the signal data sample output by the SoftMax output layer.

8. A method for diagnosing equipment faults, characterized in that: Acquire signal data characterizing equipment faults during equipment operation, input the signal data into an equipment fault diagnosis network model for detection, and obtain fault category information of the equipment; wherein the equipment fault diagnosis network model is an equipment fault diagnosis network model trained using any training method described in claims 1 to 7.

9. An electronic device, comprising a memory and a processor, characterized in that: The memory stores a computer program, and when the processor executes the computer program, the training method according to any one of claims 1 to 7 or the equipment fault diagnosis method according to claim 8 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the training method according to any one of claims 1 to 7 or the equipment fault diagnosis method according to claim 8 is implemented.

Images