Rotating Equipment Fault Diagnosis Method Based on Partial Domain Adaptation and Knowledge Distillation

Abstract

The present invention discloses a method for fault diagnosis of rotating equipment based on partial domain adaptation and knowledge distillation, which belongs to the field of fault identification. In terms of network architecture, it is composed of two parallel migration networks with the same structure, and each migration network is composed of three basic units: feature extractor, domain discriminator and classifier. Among them, the feature extractor is constructed using the Vision Transformer network, and the domain discriminator and classifier are constructed using two independent three-layer fully connected neural networks. In terms of network parameter updating, a weighted balancing mechanism and partial adversarial training are constructed to constrain the feature alignment process of the source domain and the target domain in the two migration networks, so as to achieve edge distribution alignment in the shared label space of the source domain and the target domain; the knowledge distillation method is used to alternately update the internal parameters of the two migration networks to reduce the prediction error of the two networks, so as to achieve state distribution alignment in the shared label space of the source domain and the target domain.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a rotating equipment fault diagnosis method based on partial domain adaptation and knowledge distillation. Background Art

[0002] As modern industrial machinery becomes increasingly complex, its stability becomes increasingly important, necessitating efficient monitoring technology to confirm its health status. With the continuous advancement of sensor technology, more and more condition monitoring data can be collected, and data-driven condition monitoring technology has shown great potential.

[0003] Deep learning is a highly efficient data-driven condition monitoring method that can fully leverage multi-source condition monitoring data to adaptively assess the health status of mechanical equipment. Because it assumes that training and test data follow the same distribution, deep learning models can be trained using labeled training data and their performance evaluated using unlabeled test data. However, in most real-world scenarios, the distributions of training and test data are inconsistent, leading to significant domain shift between the two.

[0004] Diagnostic methods based on transfer learning eliminate the impact of domain shift on diagnostic performance. Domain adaptation, in particular, is a specialized method that uses moment matching or adversarial training to align the feature distributions of source and target domain data. However, existing domain adaptation methods are based on the assumption that the source and target domains share the same label space. However, in real-world scenarios, it is difficult to find source and target domain data with identical label spaces. The label information of the source domain data often far exceeds that of the target domain data, and the label space of the target domain data is often a subset of the source domain data. This scenario can be referred to as the local domain adaptation problem. It poses significant challenges to feature alignment between source and target domain data in the following two aspects.

[0005] 1. Since the shared label space between source domain data and target domain data is unknown, the marginal distribution alignment method of traditional domain adaptation methods will seriously cause negative transfer in this case. In the marginal distribution alignment method, a category weight estimation process needs to be introduced to evaluate the contribution of each source domain category to the feature alignment process.

[0006] 2. Since the subordinate relationship of label information in the source domain data and the target domain data is relatively chaotic, the conditional distribution alignment method in the traditional domain adaptation method is not available in the local domain adaptation method, and it is impossible to achieve the category feature alignment between the source domain data and the target domain data under the shared category label. Summary of the Invention

[0007] The present invention provides a rotating equipment fault diagnosis method based on partial domain adaptation and knowledge distillation, which can improve the fault identification accuracy of the existing partial migration diagnosis method for rotating equipment.

[0008] An embodiment of the present invention provides a method for fault diagnosis of rotating equipment based on partial domain adaptation and knowledge distillation, comprising the following steps: obtaining operating data of the rotating equipment; inputting the operating data into a pre-trained fault identification network for identification to obtain the fault type of the rotating equipment, wherein the fault identification network is two parallel migration networks with the same structure, inputting source domain data and target domain data into the two parallel migration networks at the same time to update the parameters of the fault identification network, constructing a weighted balancing mechanism and partial adversarial training to constrain the source domain and target domain feature alignment process in the two parallel migration networks, and alternately updating the internal parameters of the two parallel migration networks using the knowledge distillation method, so that the edge distributions of the source domain and target domain in the two parallel migration networks in the shared label space are aligned and the state distributions of the source domain and target domain in the two parallel migration networks are aligned.

[0009] Optionally, in one embodiment of the present invention, data containing a number of fault categories greater than a first preset number in any working condition is collected as the source domain data, and the source domain data are all labeled data. The labeled source domain data set is represented as in represents the i-th sample in the source domain, represents the fault label of the i-th sample, n t Indicates the number of source domain samples, source domain dataset The label space is Y s , the corresponding edge state distribution is P s .

[0010] Optionally, in one embodiment of the present invention, data containing a number of fault categories less than a second preset number in other working conditions other than any working condition is collected as the target domain data, and the target domain data are all unlabeled data, and the target domain data are a subset of the source domain data. The unlabeled target domain data set is represented as in, represents the i-th sample in the target domain, n t Indicates the number of target domain samples, target domain dataset The label space is Y t , the corresponding edge state distribution is P t , the target domain label information is a subset of the source domain label information

[0011] Optionally, in one embodiment of the present invention, each migration network structure in the fault identification network is a feature extractor, a domain discriminator and a classifier; wherein the feature extractor is constructed using a ViT network, and the domain discriminator and the classifier are both three-layer fully connected neural networks.

[0012] Optionally, in one embodiment of the present invention, the feature extractor includes: an input processing module, a Transformer encoding module, and a dimensionality reduction module. In the input processing module, first, the collected one-dimensional vibration signal is truncated into N equally spaced tokens to obtain input data for the ViT network; then, the N segments of the intercepted vibration signal are brought into a fully connected neural network for processing, and then spliced ​​with the trainable parameter category token to obtain a combined feature; finally, the trainable parameter position embedding code and the combined feature are added to obtain the input feature of the transformer encoding layer;

[0013] The Transformer encoding module uses a multi-head self-attention mechanism to mine global feature information in monitoring data. The output features processed by the Transformer encoding module have the same dimension as the input features.

[0014] In the dimensionality reduction module, the element of the first row of the feature matrix extracted from the Transformer encoding module is a token, and the feature dimensionality reduction is performed through the fully connected network to obtain the output features of the ViT network. The output features of the ViT network are further used as the input of the classifier and discriminator.

[0015] Optionally, in one embodiment of the present invention, the marginal distributions of the source domain and the target domain in the shared label space in two parallel migration networks are aligned, including: using a weighted balancing mechanism and a partial adversarial training method to constrain the feature distributions of the source domain and the target domain in the shared label space, so that the shared category marginal distributions of the source domain and the target domain in the two parallel migration networks are aligned; in the weighted balancing mechanism, the output probability of the trained classifier for the target domain data belongs to the shared label space, and the output probability distribution of the target domain data given by the classifier is used to evaluate the shared label space and the outlier label space, so that the classifier output probabilities of the two migration networks establish a weighted balancing mechanism.

[0016] Optionally, in one embodiment of the present invention, the edge distribution alignment process includes the following steps:

[0017] The first step of the weighted balancing mechanism is to use two parallel migration networks to predict the labels of all target domain data. and The average value is used to formulate the initial weight of each category. The formula is as follows:

[0018]

[0019] in, is the probability distribution of the target domain data label given by the classifier softmax layer, K is the number of source domain data label categories, γ=[γ 1, …,γ k ,…,γ K ]; represents the initial weight of the K-th category;

[0020] The second step of the weighted balancing mechanism is to normalize the class weight coefficients to expand the weight difference between the shared label space and the outlier label space:

[0021]

[0022] Among them, max γ =max(γ)=max([γ1,…,γ k ,…,γ K ]) is the maximum value of the category weight coefficient, γ k is the category weight coefficient of the kth category, γ n is the category weight coefficient after regularization;

[0023] In partial adversarial training, the normalized weight constraint parameter update process is adopted for the classifier and domain discriminator to reduce the contribution of the source domain data to the adversarial training process in the outlier labeled space. The optimization objective formula of the first migration network is:

[0024]

[0025] Among them, θ 1e ,θ 1d ,θ 1c represents the network parameters of feature extractor E, domain discriminator D and classifier C; d i represents the domain label of the i-th sample; L y and L d represents the cross entropy loss function of the classifier and the domain discriminator; λ represents the weight parameter for measuring the two loss functions; the partial transfer network introduces a gradient reversal layer between the feature extractor and the domain discriminator. The gradient corresponding to the domain classification loss in the domain discriminator is automatically reversed before being back-propagated to the parameters of the feature extractor, thus achieving adversarial training of network parameters in an end-to-end manner;

[0026] The optimization objective formula of the second migration network is:

[0027]

[0028] Among them, θ 2e ,θ 2d ,θ 2cRepresents the network parameters of the feature extractor E, domain discriminator D and classifier C.

[0029] Optionally, in one embodiment of the present invention, aligning the state distributions of the source domain and the target domain in two parallel migration networks includes:

[0030] The two migration networks act as the teacher network and the student network in turn in each iteration. The optimization objective formula of the first migration network in each iteration is as follows:

[0031]

[0032] Among them, KLdiv is the Kullback-Leibler divergence, and is the output probability of the two migration networks for the target domain data, T is the hyperparameter of the distillation temperature, and Grd_Fix represents the gradient fixing operation. During the training process of the first migration network, the gradient of the second migration network is fixed and the prediction result of the second migration network is used as the benchmark;

[0033] The optimization objective formula of the second migration network at each iteration is as follows:

[0034]

[0035] During the parameter update process of the second migration network, the gradient of the first migration network is fixed, and the prediction result of the first migration network is used as a benchmark.

[0036] The beneficial effects of the rotating equipment fault diagnosis method based on partial domain adaptation and knowledge distillation in the embodiment of the present invention are:

[0037] (1) Using a parallel ViT network structure to extract global feature information from monitoring data. The ViT network uses a self-attention mechanism to extract global feature information, and the network has excellent classification performance. The method of the present invention uses two ViT networks to construct a parallel network structure, and iteratively updates the internal parameters of the two parallel networks to achieve the feature alignment task in the shared label space between the source and target domains.

[0038] (2) Partial domain adaptation is used to achieve marginal distribution alignment between shared category data in the source and target domains. By constraining the feature alignment process between the source and target domains through a weighted balancing mechanism and partial adversarial training, the influence of source domain data on the feature alignment process in the outlier label space in the first and second migration networks is weakened, thereby promoting marginal distribution alignment in the shared label space between the source and target domains.

[0039] (3) Using knowledge distillation, we train two migration networks with the same structure to achieve conditional distribution alignment between the shared category data in the source and target domains. In each iteration, the two migration networks act as the teacher network and the student network respectively to improve the label consistency of the target domain data of the two migration networks, thus achieving the category feature alignment between the source and target domains of the two migration networks in the shared label space.

[0040] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0042] Figure 1 A flowchart of a rotating equipment fault diagnosis method based on partial domain adaptation and knowledge distillation according to an embodiment of the present invention;

[0043] Figure 2 A schematic diagram of the execution process of a rotating equipment fault diagnosis method based on partial domain adaptation and knowledge distillation according to an embodiment of the present invention;

[0044] Figure 3 This is a diagram of the network structure of the knowledge distillation part migration provided according to an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The following describes embodiments of the present invention in detail, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and are not to be construed as limiting the present invention.

[0046] Figure 1 The present invention provides a flowchart of a rotating equipment fault diagnosis method based on partial domain adaptation and knowledge distillation according to an embodiment of the present invention.

[0047] like Figure 1 As shown in FIG, the rotating equipment fault diagnosis method based on partial domain adaptation and knowledge distillation includes the following steps:

[0048] In step S101 , operating data of the rotating equipment is acquired.

[0049] In step S102, the operating data is input into a pre-trained fault identification network for identification to obtain the fault type of the rotating equipment, wherein the fault identification network is two parallel migration networks with the same structure, the source domain data and the target domain data are simultaneously input into the two parallel migration networks to update the parameters of the fault identification network, a weighted balance mechanism and partial adversarial training are constructed to constrain the source domain and target domain feature alignment process in the two parallel migration networks, and the knowledge distillation method is used to alternately update the internal parameters of the two parallel migration networks, so that the edge distribution of the source domain and the target domain in the shared label space in the two parallel migration networks is aligned and the state distribution of the source domain and the target domain in the two parallel migration networks is aligned.

[0050] Part of the data with the same working conditions as the target domain data is selected as test data, and the test data is input into the trained two-level alignment partial transfer network to predict the category information of the test data.

[0051] In order to train the fault recognition network, Figure 2 As shown in the figure, we first collect data from a certain working condition containing a large number of fault categories as source domain data. The source domain data is all labeled data. We then collect data from other working conditions containing a smaller number of fault categories as target domain data. The target domain data is all unlabeled data and is a subset of the source domain data. The source and target domain data are simultaneously input into two constructed migration networks with identical structures. By updating the internal parameters of the networks, we achieve alignment of the edge distributions in the shared label space of the source and target domains in the first and second migration networks, as well as alignment of the state distributions of the source and target domains in the two migration networks.

[0052] Specifically, the labeled source domain dataset is represented as in represents the i-th sample in the source domain, represents the fault label of the i-th sample, n s Indicates the number of source domain samples; source domain dataset The label space is Y s , the corresponding edge state distribution is P s .

[0053] The unlabeled target domain dataset is represented as in represents the i-th sample in the target domain, n t Indicates the number of target domain samples; target domain dataset The label space is Y t , the corresponding edge state distribution is P t ; The target domain label information is a subset of the source domain label information, that is

[0054] like Figure 3 As shown in the figure, the fault recognition network structure based on partial domain adaptation and knowledge distillation includes three basic units, namely feature extractor E, domain discriminator D and classifier C; among them, the feature extractor E is constructed using the ViT (Vision Transformer) network, and the domain discriminator D and classifier C are both three-layer fully connected neural networks.

[0055] The ViT network utilizes a self-attention mechanism to effectively extract global information and has superior recognition capabilities compared to convolutional neural networks. Therefore, the ViT network is used as a feature extractor. The constructed ViT network includes an input processing module, a Transformer encoding module, and a dimensionality reduction module. In the input processing module, the collected one-dimensional vibration signal is first truncated into N equally spaced tokens to obtain the input data for the ViT network. Then, the N segments of the intercepted vibration signal are brought into a fully connected neural network for processing and then concatenated with the trainable parameter category tokens to obtain the combined features. Finally, the trainable parameter position embedding code and the combined features are added together to obtain the input features of the transformer encoding layer.

[0056] The constructed Transformer encoding module and classification module maintain the same structure as those in the traditional ViT network. The Transformer encoding module utilizes a multi-head self-attention mechanism to mine global feature information from monitoring data. The output features processed by the Transformer encoding module have the same dimension as the input features. In the dimensionality reduction module, the elements of the first row of the feature matrix extracted from the Transformer encoding module can be called a token. Feature dimensionality reduction is performed through a fully connected network, resulting in the output features of the ViT network. The output features of the ViT network are further used as input to the classifier and discriminator.

[0057] The specific steps of the edge distribution alignment process are as follows:

[0058] A weighted balancing mechanism and partial adversarial training are used to constrain the feature distributions of the source and target domains in a shared label space, aligning the shared category margins of the source and target domains. In the weighted balancing mechanism, since the output probabilities of a well-trained classifier for target domain data largely fall into the shared label space, the output probability distribution of the target domain data given by the classifier can be used to evaluate the shared label space and the outlier label space. Therefore, the classifier output probabilities of the two transferred networks are used to establish a weighted balancing mechanism.

[0059] First, the weighted balancing mechanism consists of two steps. In the first step, the two migration networks predict the labels of all target domain data. and The average value is used to determine the initial weight of each category. This averaging operation can effectively eliminate the influence of small errors during training. The formula is as follows:

[0060]

[0061] in, is the probability distribution of the target domain data label given by the classifier softmax layer; K is the number of source domain data label categories, γ=[γ 1, …,γ k ,…,γ K ]; represents the initial weight of the K-th category.

[0062] Then, in the second step, the class weight coefficients are normalized to expand the weight difference between the shared label space and the outlier label space:

[0063]

[0064] Among them, max γ =max(γ)=max([γ1,…,γ k ,…,γ K ]) is the maximum value of the category weight coefficient, γ k is the category weight coefficient of the kth category, γ n is the category weight coefficient after regularization; a small weight means that the corresponding category largely belongs to the outlier label space. On the contrary, a large weight means that the corresponding category has a high probability of being an element of the shared label space.

[0065] Finally, in partial adversarial training, normalized weights are used to constrain parameter updates for both the classifier and the domain discriminator to reduce the contribution of source domain data to the adversarial training process in the outlier labeled space. The first transfer network optimization objective formula is as follows:

[0066]

[0067] Among them, θ 1e ,θ 1d ,θ 1c represents the network parameters of feature extractor E, domain discriminator D and classifier C; d i represents the domain label of the i-th sample; L y and L d Represents the cross entropy loss function of the classifier and the domain discriminator; λ represents the weight parameter for measuring the two loss functions; the partial transfer network introduces a gradient reversal layer between the feature extractor and the domain discriminator. The gradient corresponding to the domain classification loss in the domain discriminator is automatically reversed before backpropagating to the parameters of the feature extractor, thereby realizing network parameter adversarial training in an end-to-end manner.

[0068] In the same way, the optimization goal of the second migration network is also given as follows:

[0069]

[0070] Among them, θ 2e ,θ 2d ,θ 2c Represents the network parameters of the feature extractor E, domain discriminator D and classifier C.

[0071] The state distribution alignment process specifically includes the following steps:

[0072] Since label consistency of target domain data between two well-trained transfer networks can achieve categorical feature alignment of source and target domain data in a shared label space, an effective method is needed to determine the label consistency of the two parallel networks. Knowledge distillation is a successful method to distill knowledge extracted from the teacher network into the student network, thereby ensuring label consistency between the teacher and student networks.

[0073] In this study, in order to improve the label consistency between the two migration networks, the two networks act as the teacher network and the student network in turn in each iteration. The optimization objective formula of the first migration network in each iteration is as follows:

[0074]

[0075] Among them, KLdiv means Kullback-Leibler divergence, and are the output probabilities of the two transfer networks for the target domain data. T is the hyperparameter for the distillation temperature, and Grd_Fix represents the gradient fixing operation. During the training of the first transfer network, the gradient of the second transfer network is fixed, and the prediction results of the second transfer network are used as the benchmark.

[0076] In the same way, the optimization objective formula of the second migration network at each iteration is given as follows:

[0077]

[0078] During the parameter update process of the second migration network, the gradient of the first migration network is fixed, and the prediction result of the first migration network is used as a benchmark.

[0079] Bearing failure data at two rotational speeds was collected using a rotating equipment test bench to verify the effectiveness of the proposed method. The test bench consists of a motor, rotor system, load block, and supporting bearings. The experimental bearing is located in the bearing housing on the right side of the test bench. An acceleration sensor attached to the front of the bearing housing was used to collect bearing failure data.

[0080] In a fault experiment, the effectiveness of the proposed method was verified by adjusting the motor speed to obtain bearing fault data at two speeds, 900 rpm and 1500 rpm. The bearing fault data at each speed included four types of bearing health states: normal (N), inner race fault (IF), ball fault (RF), and outer race fault (OF). There were 200 bearing fault samples for each type, and the data length of each sample was 10 kHz. The recorded state detection data was sampled at a frequency of 10 kHz and a duration of 1 second. There were 200 bearing fault samples for each type, and the data length of each sample was 10 kHz. Using the collected bearing data, four partial domain adaptation tasks (as shown in Table 1) were set to verify the performance of the proposed method. The 900 rpm fault data was used as the source domain data, and the 1500 rpm fault data of some categories was used as the target domain data.

[0081] During training, the model updates were set to 100 times, with a batch size of 20. The learning rate for both networks was set to 0.0001. Backpropagation was used to update the parameters of both networks, and Adam was used to minimize the optimization objective for both networks. After training, the diagnostic results of the first network with consistent structure were analyzed. The network was implemented on a computer with an Intel Core i9-12900K CPU processor at 3.60GHz, 32GB of RAM, and an NVIDIA GeForce RTX 3060 GPU, using the PyTorch 1.7.1 deep learning framework.

[0082] Table 1 Experimental data of bearing failure under different working conditions

[0083]

[0084] In order to verify the effectiveness of the proposed method, five comparison methods were constructed. These comparison methods were subjected to five experiments in the following analysis. The details of these five comparison methods are as follows:

[0085] Comparative Method 1 (Single ViT Network + Partial Adversarial Training): In this comparative method, the ViT-based feature extractor and the classifier in the proposed method are combined into a new network. Only partial adversarial training in the proposed method is used to narrow the distribution difference of marginal features between the source and target domain data. The comparison results between this method and the proposed method demonstrate the effectiveness of the proposed method's conditional distribution alignment measure.

[0086] Comparative Method 2 (Structurally Consistent Dual ViT Network + Knowledge Distillation): This method uses the same Dual ViT network as the proposed method as the basic network architecture. It only uses the knowledge distillation from the proposed method to achieve conditional distribution alignment of source and target domain data in a shared label space. The comparison between this method and the proposed method demonstrates the effectiveness of the proposed method's marginal distribution alignment measure.

[0087] Comparison Method 3 (Structurally Consistent Dual ViT Network + L1 Loss): This comparison method uses a dual ViT network with the same structure as the proposed method as the base network architecture. L1 loss is used to evaluate the inconsistency of target domain features extracted from two structurally consistent networks. This comparison method aims to highlight the performance of the proposed method's conditional distribution alignment and partial domain adaptation methods.

[0088] Comparison Method 4 (Structurally Consistent Dual CNN Network + Proposed Loss Function): In this comparison method, a CNN is used as the feature extractor and combined with the classifier and discriminator from the proposed method to form a CNN-based network structure. The two basic CNN-based networks are integrated to form a structurally consistent dual-CNN network architecture. The comparison results of this comparison method with the proposed method highlight the feature extraction capabilities of the ViT network.

[0089] Comparison Method 5 (Single ViT Network + Cross-Entropy Loss): In this comparison method, the ViT-based feature extractor is combined with the classifier from the proposed method to form a base network. Only the cross-entropy loss is used to update the parameters of the base network. The results of this comparison method serve as a benchmark for comparison with other methods.

[0090] Table 2 shows the diagnostic results of the proposed method and other comparative methods, and the following conclusions can be drawn:

[0091] 1. Under the four working conditions, the diagnostic results of the proposed method are better than those of other comparison methods, which is reflected in the accuracy and volatility of the statistical diagnostic results.

[0092] 2. Comparing the proposed method with the comparative method 1, it can be concluded that the conditional distribution alignment measure in the proposed method helps to improve the performance of feature alignment.

[0093] 3. Comparing comparison method 2 with comparison method 3, it can be concluded that knowledge distillation has much stronger constraints on conditional feature alignment than the L1 loss function.

[0094] 4. Comparing the proposed method with the comparative method 2, it can be found that both marginal distribution alignment and conditional distribution alignment help improve the performance of fault diagnosis.

[0095] 5. The main difference between the comparison method 4 and the proposed method lies in the network structure. The diagnostic results of the proposed method are significantly higher than those of the comparison method based on the CNN network architecture. This shows that the ViT network can extract distinguishable features better than the CNN network.

[0096] 6. Since the proposed method and the first three comparison methods are built by adding some different elements to the original ViT network structure, these methods have higher fault diagnosis results compared with the single ViT network structure.

[0097] Table 2 Diagnosis results of different methods for the first migration network

[0098]

[0099] According to the rotating equipment fault diagnosis method based on partial domain adaptation and knowledge distillation proposed in an embodiment of the present invention, in terms of network architecture, it is composed of two migration networks with the same structure, and each migration network is composed of three basic units: feature extractor, domain discriminator and classifier. Among them, the feature extractor is constructed using the Vision Transformer network, and the domain discriminator and classifier are constructed using two independent three-layer fully connected neural networks. In terms of network parameter updating, 1) a weighted balancing mechanism and partial adversarial training are constructed to constrain the feature alignment process of the source domain and target domain in the two migration networks, so as to achieve edge distribution alignment in the shared label space of the source domain and the target domain; 2) the knowledge distillation method is used to alternately update the internal parameters of the two migration networks to reduce the prediction error of the two networks and achieve state distribution alignment in the shared label space of the source domain and the target domain.

[0100] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and the features of different embodiments or examples without contradiction.

[0101] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of technical features indicated. Thus, a feature specified as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "N" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0102] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code comprising one or N executable instructions for implementing a custom logical function or step of a process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may be performed out of the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present invention pertain.

Claims

1. A rotating equipment fault diagnosis method based on partial domain adaptation and knowledge distillation, characterized in that: The following steps are involved: Obtain operating data of rotating equipment; The operating data is input into a pre-trained fault identification network for identification to obtain the fault type of the rotating equipment, wherein the fault identification network is two parallel migration networks with the same structure, the source domain data and the target domain data are simultaneously input into the two parallel migration networks to update the parameters of the fault identification network, a weighted balance mechanism and partial adversarial training are constructed to constrain the source domain and target domain feature alignment process in the two parallel migration networks, and the internal parameters of the two parallel migration networks are alternately updated using the knowledge distillation method, so that the edge distributions of the source domain and the target domain in the two parallel migration networks in the shared label space are aligned and the state distributions of the source domain and the target domain in the two parallel migration networks are aligned; wherein the edge distribution alignment of the source domain and the target domain in the shared label space in the two parallel migration networks includes: A weighted balancing mechanism and partial adversarial training methods are used to constrain the feature distributions of the source and target domains in a shared label space, aligning the marginal distributions of shared categories between the source and target domains in two parallel transfer networks. In the weighted balancing mechanism, the output probability of the trained classifier for target domain data belongs to the shared label space. The output probability distribution of the target domain data given by the classifier is used to evaluate the shared label space and the outlier label space, thus establishing a weighted balancing mechanism for the classifier output probabilities of the two transfer networks. Aligning the state distributions of the source and target domains in two parallel transfer networks involves: The two migration networks act as the teacher network and the student network in turn in each iteration. The optimization objective formula of the first migration network in each iteration is as follows: Among them, KLdiv is the Kullback-Leibler divergence, and is the output probability of the two migration networks for the target domain data, T is the hyperparameter of the distillation temperature, and Grd_Fix represents the gradient fixing operation. During the training process of the first migration network, the gradient of the second migration network is fixed and the prediction result of the second migration network is used as the benchmark; The optimization objective formula of the second migration network at each iteration is as follows: During the parameter update process of the second migration network, the gradient of the first migration network is fixed, and the prediction result of the first migration network is used as a benchmark.

2. The method according to claim 1, characterized in that Collect data containing fault categories greater than a first preset number in any working condition as the source domain data. The source domain data are all labeled data. The labeled source domain data set is represented as in represents the i-th sample in the source domain, represents the fault label of the i-th sample, n s Indicates the number of source domain samples, source domain dataset The label space is Y s , the corresponding edge state distribution is P s .

3. The method according to claim 2, characterized in that Collect data containing fault categories less than a second preset number in other working conditions other than any working condition as the target domain data. The target domain data are all unlabeled data. The target domain data is a subset of the source domain data. The unlabeled target domain data set is represented as in, represents the i-th sample in the target domain, n t Indicates the number of target domain samples, target domain dataset The label space is Y t , the corresponding edge state distribution is P t , the target domain label information is a subset of the source domain label information 4. The method according to claim 1, wherein Each migration network structure in the fault identification network is a feature extractor, a domain discriminator and a classifier; wherein the feature extractor is constructed using the ViT network, and the domain discriminator and the classifier are both three-layer fully connected neural networks.

5. The method according to claim 4, characterized in that The feature extractor includes: an input processing module, a Transformer encoding module, and a dimensionality reduction module. In the input processing module, the collected one-dimensional vibration signal is first truncated into N equally spaced tokens to obtain input data for the ViT network; then, the N segments of the truncated vibration signal are brought into a fully connected neural network for processing and then concatenated with trainable parameter category tokens to obtain combined features; finally, the trainable parameter position embedding code and the combined features are added to obtain the input features of the transformer encoding layer; The Transformer encoding module uses a multi-head self-attention mechanism to mine the global feature information in the monitoring data. The output features processed by the Transformer encoding module have the same dimension as the input features. In the dimensionality reduction module, the element of the first row of the feature matrix extracted from the Transformer encoding module is a token. The feature dimensionality reduction is performed through a fully connected network to obtain the output features of the ViT network. The output features of the ViT network are further used as the input of the classifier and discriminator.

6. The method according to claim 1, characterized in that The marginal distribution alignment process includes the following steps: The first step of the weighted balancing mechanism is to use two parallel migration networks to predict the labels of all target domain data. and The average value is used to formulate the initial weight of each category. The formula is as follows: in, is the probability distribution of the target domain data label given by the classifier softmax layer, K is the number of source domain data label categories, γ=[γ 1, …,γ k ,…,γ K ]; γ k represents the initial weight coefficient of the kth category; The second step of the weighted balancing mechanism is to normalize the class weight coefficients to expand the weight difference between the shared label space and the outlier label space: Among them, max γ =max(γ)=max([γ1,…,γ k ,…,γ K ]) is the maximum value of the category weight coefficient, γ n is the category weight coefficient after regularization; In partial adversarial training, the normalized weight constraint parameter update process is adopted for the classifier and domain discriminator to reduce the contribution of the source domain data to the adversarial training process in the outlier labeled space. The optimization objective formula of the first migration network is: Among them, θ 1e ,θ 1d ,θ 1c represents the network parameters of feature extractor E, domain discriminator D and classifier C; d i represents the domain label of the i-th sample; L y and L d represents the cross entropy loss function of the classifier and the domain discriminator; λ represents the weight parameter for measuring the two loss functions; the partial transfer network introduces a gradient reversal layer between the feature extractor and the domain discriminator. The gradient corresponding to the domain classification loss in the domain discriminator is automatically reversed before being back-propagated to the parameters of the feature extractor, thus achieving adversarial training of network parameters in an end-to-end manner; The optimization objective formula of the second migration network is: Among them, θ 2e ,θ 2d ,θ 2c Represents the network parameters of the feature extractor E, domain discriminator D and classifier C.

Images