Small sample fault diagnosis method based on self-supervised learning generative adversarial network

Abstract

The invention relates to a small sample fault diagnosis method based on a self-supervised learning generative adversarial network. The method comprises the following steps: collecting an acoustic emission signal of a hoisting system as real small sample fault data; constructing a meta-learning generative adversarial network model and a self-supervised classification network, and performing parameter initialization; training a meta-learning generative adversarial network model based on real small sample fault data to obtain an optimal initialization model and generate small sample data; inputting the generated small sample data into a self-supervised classification network for self-supervised training; and inputting to-be-diagnosed fault data into the optimal initialization model, outputting generated data, inputting the generated data into the trained self-supervised classification network, performing fault classification, and determining a fault diagnosis result. Diversified fault data can be generated based on a small number of data samples, and the learning ability and generalization ability of a diagnosis model are improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence and mechanical engineering technology, and in particular to a small sample fault diagnosis method based on self-supervised learning to generate adversarial networks. Background Art

[0002] As the most common and indispensable equipment in engineering construction, the lifting system plays an important role in modern industrial production. However, due to the complex application environment and high application intensity of the lifting system, its potential safety risks are gradually increasing. More specifically, due to the increased difficulty in identification, important components of the lifting system (such as balance beams, A-frames, tie rods, fly arms, and cargo arms) are prone to local damage, causing the entire large lifting system to shut down. Therefore, it is of great significance and practical value to develop an accurate and advanced intelligent fault diagnosis technology for lifting systems.

[0003] Nowadays, data-driven fault diagnosis methods have been widely used in the fields of fault identification, remaining useful life (RUL) prediction, and structural health monitoring (SHM). In particular, deep learning (DL) has achieved a series of breakthroughs in intelligent fault diagnosis based on big data, thanks to its powerful nonlinear feature self-learning and intelligent fault identification capabilities. However, in actual industrial applications, big data often has the problems of high data repetitiveness and small sample data, among which the information of small sample data is usually incomplete and insufficient, resulting in a decrease in the diagnostic accuracy of deep learning-based fault diagnosis methods, which usually require sufficient training data. On the other hand, most of the mechanical equipment is in normal operation throughout its life cycle, and most of the data does not have sufficient category labels, or the cost of obtaining category label samples is high. More importantly, once a sudden catastrophic failure occurs, the system will be shut down immediately for maintenance. In addition, the non-stationary dynamic characteristics of the lifting system under multiple working conditions (such as luffing, lifting, and rotation) greatly increase the difficulty of collecting its state data. Therefore, it is difficult to collect enough fault data to train the deep learning network, especially the insufficient number of training samples will inevitably lead to overfitting problems in the deep learning model, thereby limiting its high-precision diagnostic performance. At present, how to realize fault diagnosis with only a small amount of data and obtain satisfactory diagnostic accuracy is a difficult problem that needs to be solved urgently and a hot research direction in the field of mechanical intelligent fault diagnosis.

[0004] In recent years, a lot of research has been conducted on intelligent fault diagnosis methods to solve the problem of small sample learning, including meta-learning, data augmentation, transfer learning, semi-supervised learning and synthetic data generation. However, the existing data augmentation and synthetic data generation methods have failed to effectively solve the problem of too many model parameters and difficulty in fitting due to insufficient samples, thereby reducing the diagnostic accuracy. In addition, small sample fault diagnosis methods based on meta-learning, semi-supervised learning and transfer learning require a large number of different tasks and training samples, respectively, resulting in serious overfitting and reducing the generalization ability of such algorithms. In summary, the existing small sample analysis methods have obvious shortcomings: 1) The number of features extracted by the algorithm model is limited, and there is a lack of analysis and attention to important features of different scales; 2) The generalization performance of most algorithms is usually poor; 3) In the pre-training stage, in order to obtain the basic network, the existing methods use a data set with a large distribution difference from the new category data set, which reduces the classification effect of the new category.

[0005] In addition, the small sample learning methods based on data generation mainly include simulated sampling, stacked sparse autoencoders (SSAE), support vector machines (SVM), twin neural networks (SNN), and generative adversarial networks (GAN). Among them, simulated sampling generates data by obtaining the probability distribution of data and simulating the random sampling process, which requires accurate acquisition of the probability distribution of data and cannot guarantee the diversity of generated data. A major advantage of SSAE is that its model performance will be improved as the size of the training dataset increases. However, SSAE requires setting many hyperparameters, such as the number of hidden layers, training rounds, learning rate, and decoder function, which increases the complexity of the model and takes a long time. In addition, due to the natural imbalance problem of the collected mechanical system health dataset itself, the performance of SVM in fault classification is seriously affected. On the other hand, although SNN can learn low-level and high-level features from rich datasets and provide a high detection rate, it requires a large amount of data input to discover more correlations between features and cannot effectively detect and classify fault information trained by limited samples.

[0006] Therefore, in the related technology, there is an urgent need for a method that can solve the small sample learning problem and realize small sample fault diagnosis of the lifting system. Summary of the invention

[0007] Based on this, it is necessary to provide a small sample fault diagnosis method based on self-supervised learning generative adversarial network, which can solve the small sample learning problem and realize small sample fault diagnosis of lifting systems in response to the above technical problems.

[0008] In a first aspect, the present application provides a small sample fault diagnosis method based on self-supervised learning generative adversarial network. The method comprises:

[0009] Collect the acoustic emission signals of the lifting system as real small sample fault data;

[0010] Constructing a meta-learning generative adversarial network model and a self-supervised classification network, and initializing parameters, wherein the meta-learning generative adversarial network model is composed of a base learner and a meta-learner, and the self-supervised classification network is composed of a feature extractor, a category classifier, and a translation scale classifier;

[0011] Based on the real small sample fault data, the meta-learning generative adversarial network model is trained to obtain an optimal initialization model and generate small sample data;

[0012] Inputting the generated small sample data into the self-supervised classification network for self-supervised training;

[0013] The fault data to be diagnosed is input into the optimal initialization model, and the generated data is output. The generated data is input into the trained self-supervised classification network to perform fault classification and determine the fault diagnosis result.

[0014] Optionally, in one embodiment of the present application, the base learner and the meta learner each include a generator and a discriminator, the generator of the base learner is used to generate data, the discriminator of the base learner is used to distinguish the authenticity of the generated data, and the generator and discriminator of the meta learner are used to balance the learning effect of the base learner.

[0015] Optionally, in one embodiment of the present application, the training of the meta-learning generative adversarial network model based on the real small sample fault data includes:

[0016] Use the model parameters of the meta-learner to initialize the model parameters of each base learner, perform adversarial training updates on each base learner using the data from the respective task support set, and calculate the losses of the generator and discriminator of the base learner;

[0017] Compute the losses of the generator and discriminator of the meta-learner and update the parameters of the meta-learner.

[0018] Optionally, in one embodiment of the present application, the loss functions of the generator and the discriminator of the base learner are respectively expressed as:

[0019]

[0020]

[0021] Among them, θ G is the parameter of the generator, θ D is the parameter of the discriminator, G θG is the generator, D θD is the discriminator, z is the noise, x is the generated data, X is the real data set, and Z is the noise spatial distribution of the generator input.

[0022] Optionally, in one embodiment of the present application, the loss functions of the generator and the discriminator of the meta-learner are respectively expressed as:

[0023]

[0024]

[0025] Among them, θ G is the parameter of the generator, θ D is the parameter of the discriminator, G θG is the generator, D θD is the discriminator, and To train the parameters of the generator and discriminator of the updated base learner, z is the noise, x is the generated data, X is the real data set, Z is the noise spatial distribution of the generator input, and P(T) is the spatial distribution of this task.

[0026] Optionally, in one embodiment of the present application, inputting the generated small sample data into the self-supervised classification network for self-supervised training includes:

[0027] The waveform generating the small sample data is translated, a waveform structure information label is established, and feature extraction is performed through a feature extractor;

[0028] Perform traditional category classification and translation scale classification, and calculate category classification loss and translation scale classification loss respectively;

[0029] Parameters of the feature extractor, the category classifier, and the translation-scale classifier are adjusted based on the category classification loss and the translation-scale classification loss.

[0030] Optionally, in one embodiment of the present application, the calculation process of the category classification loss and the translation scale classification loss is expressed as:

[0031]

[0032] Among them, λ is the balance parameter between category classification loss and translation scale classification loss, L classification is the category classification loss for traditional category classification, L self-supervised Translation-scale classification loss for translation-scale classification.

[0033] In a second aspect, the present application also provides a small sample fault diagnosis device based on self-supervised learning to generate adversarial networks. The device comprises:

[0034] A data acquisition module is used to collect the acoustic emission signals of the lifting system as real small sample fault data;

[0035] A model building module, used to build a meta-learning generative adversarial network model and a self-supervised classification network, and perform parameter initialization, wherein the meta-learning generative adversarial network model is composed of a base learner and a meta-learner, and the self-supervised classification network is composed of a feature extractor, a category classifier, and a translation scale classifier;

[0036] A meta-learning generative adversarial network model training module is used to train the meta-learning generative adversarial network model based on the real small sample fault data to obtain an optimal initialization model and generate small sample data;

[0037] A self-supervised training module, used for inputting the generated small sample data into the self-supervised classification network for self-supervised training;

[0038] The fault diagnosis result determination module is used to input the fault data to be diagnosed into the optimal initialization model, output generated data, input the generated data into the trained self-supervised classification network, perform fault classification, and determine the fault diagnosis result.

[0039] In a third aspect, the present application further provides a computer device, wherein the computer device comprises a memory and a processor, wherein the memory stores a computer program, and the processor executes the steps of the methods described in the above embodiments.

[0040] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the method described in each of the above embodiments are implemented.

[0041] The above-mentioned small sample fault diagnosis method based on self-supervised learning generative adversarial network, first, collects the acoustic emission signal of the lifting system as the real small sample fault data; then, constructs a meta-learning generative adversarial network model and a self-supervised classification network, and performs parameter initialization, wherein the meta-learning generative adversarial network model is composed of a base learner and a meta-learner, and the self-supervised classification network is composed of a feature extractor, a category classifier and a translation scale classifier; then, trains the meta-learning generative adversarial network model based on the real small sample fault data to obtain the optimal initialization model and generate small sample data; then, inputs the generated small sample data into the self-supervised classification network for self-supervised training; finally, inputs the fault data to be diagnosed into the optimal initialization model, outputs the generated data, and inputs the generated data into the trained self-supervised classification network to perform fault classification and determine the fault diagnosis result. In other words, it is possible to generate new data using a small amount of training data, and effectively combine real data with generated data for lifting system fault diagnosis, and the designed self-supervised strategy can deeply extract information from the generated data and improve the generalization performance of the algorithm. With only a small amount of data samples, diversified fault data can be generated, improving the learning and generalization capabilities of the diagnostic model, thereby achieving high-precision fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 An application environment diagram of a small sample fault diagnosis method based on self-supervised learning to generate adversarial networks in one embodiment;

[0043] Figure 2 A schematic diagram of a process of a small sample fault diagnosis method based on self-supervised learning to generate adversarial networks in one embodiment;

[0044] Figure 3 A schematic diagram of the structure of a meta-learning generative adversarial network model and a self-supervised classification network in one embodiment;

[0045] Figure 4 A schematic diagram of fault classification accuracy rates obtained by different methods under different ratios of real data to generated data in one embodiment;

[0046] Figure 5 A schematic diagram of classification accuracy of three self-supervisory strategies under different task settings in one embodiment;

[0047] Figure 6 A structural block diagram of a small sample fault diagnosis device based on self-supervised learning to generate adversarial networks in one embodiment;

[0048] Figure 7 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0050] The small sample fault diagnosis method based on self-supervised learning generative adversarial network provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown. Among them, the terminal communicates with the server through the network. The data storage system can store the data that the server needs to process. The data storage system can be integrated on the server, or it can be placed on the cloud or other network servers. Among them, the terminal can be but not limited to various personal computers, laptops, smart phones, tablets, Internet of Things devices and portable wearable devices. The Internet of Things devices can be smart speakers, smart TVs, smart air conditioners, smart car-mounted devices, etc. Portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. The server can be implemented as an independent server or a server cluster consisting of multiple servers.

[0051] In one embodiment, Figure 2 As shown in the figure, a small sample fault diagnosis method based on self-supervised learning generative adversarial network is provided. Figure 1 The server in the example is used to illustrate the following steps:

[0052] S201: Collect the acoustic emission signal of the lifting system as real small sample fault data.

[0053] In the embodiment of the present application, first, the acoustic emission signal of the lifting system is collected as the real small sample fault data. Specifically, the acoustic emission data set is collected using the MQ1260-45 gantry crane, and the digitized waveform at a sampling frequency of 1MHz is obtained by the SAMOS data acquisition system (equipped with 40 channels). The R15I-AST piezoelectric acoustic emission sensor with a resonant frequency of 150kHz is used and attached to the crane with vaseline as an acoustic coupling agent. In addition, the environmental noise can be effectively eliminated by setting the bandpass filter to a frequency range of 10kHz to 500kHz and a threshold of 40dB. A set of acoustic emission data of the lifting system containing 8 fault modes is collected. Specifically, these 8 fault modes are composed of acoustic emission signals generated by 7 key components with certain damage and 1 healthy component in the lifting system under lifting conditions. The 8 fault modes obtained are defined as slewing column fault (RC), A-frame fault (AF), boom fault (CB), gantry leg fault (GL), tower body fault (T), rotating platform beam fault (RP), counterweight fault (BW) and no fault (N). The acoustic emission data samples of different fault modes contain a total of 160,000 data points, and the data points corresponding to each fault mode are counted respectively. According to the analysis results of the original time domain signal. Among them, the number of data sample points of the 8 fault modes (RC, AF, CB, GL, T, RP, BW and N) are 5600, 3360, 2720, 4480, 3840, 4960, 3040, and 132000 respectively, and the specific percentages are shown in Table 1. It can be seen from Table 1 that the monitoring data of the hoisting system under fault state is far less than the data collected under healthy state, which further verifies the necessity and significance of small sample fault diagnosis of the hoisting system. At the same time, compared with the 1024 fault data points usually contained in the vibration signal, these acoustic emission data points can be fully used as a verification data set to study the effectiveness of the proposed method.

[0054] Table 1. Data percentage of different failure modes

[0055]

[0056] S203: constructing a meta-learning generative adversarial network model and a self-supervised classification network, and initializing parameters, wherein the meta-learning generative adversarial network model is composed of a base learner and a meta-learner, and the self-supervised classification network is composed of a feature extractor, a category classifier, and a translation scale classifier.

[0057] In the embodiments of the present application, the traditional data generation model based on generative adversarial network (GAN) requires a large amount of training data, while the existing few-sample learning method based on meta-learning requires a large number of different tasks and faces serious overfitting problems. Therefore, a small-sample fault diagnosis algorithm based on self-supervised meta-learning generative adversarial network (SSMLGAN) is proposed, which specifically includes a meta-learning generative adversarial network model and a self-supervised classification network, wherein the meta-learning generative adversarial network model is composed of a base learner and a meta-learner, and the self-supervised classification network is composed of a feature extractor, a category classifier, and a translation scale classifier. First, the meta-learner is initialized, and the model parameters of each base learner are initialized using the model parameters of the meta-learner. At the same time, the Resnet12 feature extractor that implements the self-supervised task is initialized. lightGBM category classifier Translational Scale Classifier Perform random initialization.

[0058] Specifically, in one embodiment of the present application, the base learner and the meta learner each include a generator and a discriminator, the generator of the base learner is used to generate data, the discriminator of the base learner is used to distinguish the authenticity of the generated data, and the generator and discriminator of the meta learner are used to balance the learning effect of the base learner.

[0059] In one embodiment of the present application, the meta-learning generative adversarial network model is composed of multiple base learners and meta-learners, and both the base learner and the meta-learner include a generator G and a discriminator D. Among them, the model of the base learner is composed of a generator G and a discriminator D, where the generator represents the mapping of noise z to data x, which is used to generate data, and the discriminator represents the mapping of data x to true or false categories, which is used to distinguish the true and false of the generated data. The purpose of the meta-learner is to balance the learning effect of each base learner and find the optimal initialization model for all tasks, so that only a small amount of data is needed to obtain good data generation results when learning new tasks.

[0060] S205: Training the meta-learning generative adversarial network model based on the real small sample fault data to obtain an optimal initialization model and generate small sample data.

[0061] In the embodiment of the present application, the meta-learning generative adversarial network model is composed of multiple base learners and meta-learners, each of which contains a generator G and a discriminator D. The algorithm is trained by alternating between base learners and meta-learners, and the optimal initialization parameters are searched by continuously optimizing the meta-learning generative adversarial network (MLGAN) to achieve rapid convergence of new tasks (i.e., small sample data generation), and obtain a specific GAN suitable for the new task, thereby achieving the generation and expansion of small sample data. The training process is to separate each group of GAN tasks T i (consisting of a support set and a query set) as training data, which can be expressed as follows:

[0062]

[0063] Among them, both the support set and the query set are composed of actual data and generated data, that is, real small sample fault data and generated small sample data.

[0064] In one embodiment of the present application, training the meta-learning generative adversarial network model based on the real small sample fault data includes:

[0065] S301: Use the model parameters of the meta-learner to initialize the model parameters of each base learner, perform adversarial training updates on each base learner using the data of the respective task support set, and calculate the losses of the generator and discriminator of the base learner.

[0066] S303: Calculate the losses of the generator and the discriminator of the meta-learner, and update the parameters of the meta-learner.

[0067] In one embodiment of the present application, Figure 3 As shown in Figure 1, first, the meta-learner is initialized, and the model parameters of each base learner are initialized using the model parameters of the meta-learner. The adversarial training of each base learner is updated using the data of the respective task support set, and the losses of the generator and discriminator of the base learner are calculated. Specifically, the generator and discriminator of the base learner are updated on a set of meta-tasks. The goal of the discriminator is to distinguish whether the input data is real or fake, and the goal of the generator is to generate real data within a limited number of training samples and iterations.

[0068] In one embodiment of the present application, the loss functions of the generator and the discriminator of the base learner are respectively expressed as:

[0069]

[0070]

[0071] Among them, θ Gis the parameter of the generator, θ D are the parameters of the discriminator, For the generator, is the discriminator, z is the noise, x is the generated data, X is the real data set, and Z is the noise spatial distribution of the generator input.

[0072] Since the base learner iteratively updates the generator and discriminator according to the current task loss, the model parameters of the generator and discriminator can be updated as and The model performs a gradient update on a new task, which can be expressed as follows:

[0073]

[0074] Among them, α D and α G are learning rates respectively. The parameters learned by the base learner cannot be guaranteed to be applicable to all tasks. Therefore, a meta-learner is needed to update the initialization parameters according to the gradient corresponding to the optimal parameters of each task, so as to search for the optimal initialization parameters of all tasks.

[0075] The meta-learner updates the initialization parameters on the query set through the gradient corresponding to the optimal parameters of each meta-task. In one embodiment of the present application, the loss functions of the generator and the discriminator of the meta-learner are respectively expressed as:

[0076]

[0077]

[0078] Among them, θ G is the parameter of the generator, θ D are the parameters of the discriminator, For the generator, is the discriminator, and To train the parameters of the generator and discriminator of the updated base learner, z is the noise, x is the generated data, X is the real data set, Z is the noise spatial distribution of the generator input, and P(T) is the spatial distribution of this task.

[0079] In addition, the gradient is updated by evaluating the performance of the objective function on all task query sets to obtain a set of globally optimal model parameters, which is expressed as:

[0080]

[0081] Among them, β D is the control factor for the discriminator optimization process.

[0082] S207: Inputting the generated small sample data into the self-supervised classification network for self-supervised training.

[0083] In the embodiment of the present application, in the process of pre-training the basic network obtained by meta-learning, as the distribution difference between the training data set used and the new class data set increases, the classification ability of the new class also becomes worse. In order to maximize the generalization performance of the algorithm, a self-supervised training strategy is designed, such as Figure 3 As shown in the figure, by translating the generated waveform data along the X-axis, the supervision information in the training task is increased, the structural information of the generated data is fully mined, and fault classification is performed.

[0084] Specifically, in one embodiment of the present application, inputting the generated small sample data into the self-supervised classification network for self-supervised training includes:

[0085] S401: translating the waveform for generating the small sample data, establishing a waveform structure information label, and performing feature extraction through a feature extractor.

[0086] S403: Perform traditional category classification and translation scale classification, and calculate category classification loss and translation scale classification loss respectively.

[0087] S405: Adjust the parameters of the feature extractor, the category classifier, and the translation-scale classifier based on the category classification loss and the translation-scale classification loss.

[0088] In one embodiment of the present application, the generated data waveform is translated along the x-axis to establish a label of the waveform structure information to increase the supervision information in the training task, thereby making full use of the structural information of the waveform itself and improving the generalization performance of the algorithm. π and ) and use these translation scales as supervision information to create signal translation dimension labels, thereby obtaining more supervision information. After that, feature extraction is performed through the feature extractor, traditional category classification and translation scale classification are performed, and the category classification loss and translation scale classification loss are calculated respectively. The gradient descent method is used to optimize the parameters of the feature extractor, category classifier, and translation scale classifier. The specific process is as follows:

[0089]

[0090] in, is the cross entropy loss function used, including category classification loss and translation scale classification loss, and α is the learning rate.

[0091] In one embodiment of the present application, the calculation process of the category classification loss and the translation scale classification loss is expressed as:

[0092]

[0093] Among them, λ is the balance parameter between category classification loss and translation scale classification loss, L classification is the category classification loss for traditional category classification, L self-supervised Translation-scale classification loss for translation-scale classification.

[0094] Specific experiments show that when λ<0.6, the classification effect of the category classifier will decrease, and the feature extractor cannot fully learn the classification information generated by the translation dimension classifier. Therefore, λ is in the range of 0.6 to 0.9.

[0095] S209: inputting the fault data to be diagnosed into the optimal initialization model, outputting generated data, inputting the generated data into the trained self-supervised classification network, and performing fault classification in combination with the fault data to be diagnosed to determine the fault diagnosis result.

[0096] In an embodiment of the present application, the fault data to be diagnosed is input into the optimal initialization model, the generated data is output, and then input into the self-supervised classification network to obtain the classification result, complete the small sample fault diagnosis of the lifting system, and output the fault diagnosis result.

[0097] In one embodiment of the present application, the generated fault data is mixed with the real data to verify the effectiveness of the method proposed in the present application for small sample fault diagnosis of the lifting system. Table 2 shows the fault classification accuracy under different mixing ratios. As shown in Table 2, when the ratio of real data to generated data is 6:4, the classification accuracy reaches the highest, which is 99.76%. This shows that the generated data significantly learns the data characteristics of each fault mode. In addition, when the ratio of real data to generated data is greater than 6:4, the classification accuracy begins to decline, which shows that the classification accuracy can be improved by using a small amount of generated data. The results show that the proposed method has good diagnostic capabilities for small sample fault diagnosis of the lifting system.

[0098] Table 2. Fault classification accuracy obtained using the proposed method for the lifting system dataset at different ratios of real data to generated data

[0099]

[0100] In one embodiment of the present application, it is compared with five meta-learning-based methods (MAML, MAML++, MTL, T-NAS, and TSML) and three representative small sample learning methods (TADAM, TransMatch, and DKT). It should be noted that the selected small sample fault diagnosis method has the same experimental setting as the method proposed in this application. In addition, in order to better demonstrate the superiority of the proposed method, all selected meta-learning-based comparison methods use the same initialization parameter optimization strategy as the proposed method to obtain a specific GAN for the new task, thereby achieving the generation and expansion of a small amount of data. On the other hand, according to the ratio of actual data to generated data shown in Table 2, the data set generated by this method is imported into the above-mentioned three small sample learning methods to achieve the classification of different fault modes. In addition, for a fair comparison, these comparison methods adopt different task settings. As Figure 4 As shown in Figure 2, the fault classification accuracy obtained by different methods under different ratios of real data to generated data is shown. Figure 4 It can be seen that when the ratio of real data to generated data is 6:4, the classification accuracy obtained by each comparison method is the largest, which is consistent with the above analysis results. Specifically, the classification accuracy of the proposed method is higher than that of all the comparison methods, which shows the superiority of the proposed method in the small sample fault diagnosis of the lifting system. When the ratio of real data to generated data is 6:4, the detailed diagnosis results are shown in Table 3. As shown in Table 3, the classification accuracy of the method proposed in this application under different task settings is 98.89%, 99.46%, 98.43% and 99.16%, respectively, which is significantly higher than other comparison methods.

[0101] Table 3. Comparison of fault classification results obtained using different methods when the ratio of real data to generated data is 6:4

[0102]

[0103] In one embodiment of the present application, in order to further verify the superiority of the proposed method, it is compared with four existing hoisting system fault diagnosis methods (CIELMD, MPCA-CNN, MCNN-MSCF and MHMFDE-MCFS). And MATLAB R2018a is run on an Intel Core i7-8550U 3.80GHz CPU, 16.00GB RAM computer to analyze the fault recognition performance of various methods. Among them, Table 4 lists the detailed diagnosis results when the ratio of real data to generated data is 6:4, including recognition accuracy, standard deviation and CPU running time. It can be seen from Table 4 that the classification accuracy of this method is higher than that of the existing four methods, which are 73.19%, 85.34%, 89.26% and 92.16% respectively. In addition, the standard deviation and CPU running time of the proposed method are lower than those of all the compared methods, indicating that the proposed method has good stability in the small sample fault diagnosis of the hoisting system.

[0104] Table 4. Comparative analysis results with existing hoisting system fault diagnosis methods

[0105]

[0106] In one embodiment of the present application, in order to verify the effectiveness of the proposed self-supervision strategy, two types of self-supervision, four types of self-supervision and no self-supervision are used to examine the classification accuracy of the proposed algorithm under different task settings. Among them, the two types of self-supervision indicate that the translation scale of the generated signal is The 4-category self-supervision indicates that the translation scale is π and like Figure 5 As shown in Figure 2, the classification accuracy of the three self-supervised strategies under different task settings is shown. Figure 5 As shown in the figure, the addition of the self-supervision strategy can significantly improve the classification accuracy, and the proposed 4-category self-supervision method can obtain better diagnosis results compared with the 2-category self-supervision strategy.

[0107] In the above-mentioned small sample fault diagnosis method based on self-supervised learning generative adversarial network, first, the acoustic emission signal of the lifting system is collected as the real small sample fault data; then, a meta-learning generative adversarial network model and a self-supervised classification network are constructed, and the parameters are initialized, wherein the meta-learning generative adversarial network model is composed of a base learner and a meta-learner, and the self-supervised classification network is composed of a feature extractor, a category classifier and a translation scale classifier; then, the meta-learning generative adversarial network model is trained based on the real small sample fault data to obtain the optimal initialization model and generate small sample data; then, the generated small sample data is input into the self-supervised classification network for self-supervised training; finally, the fault data to be diagnosed is input into the optimal initialization model, the generated data is output, and the generated data is input into the trained self-supervised classification network to perform fault classification and determine the fault diagnosis result. In other words, it is possible to generate new data using a small amount of training data, and to effectively combine real data with generated data for lifting system fault diagnosis. The designed self-supervised strategy can deeply extract information from the generated data and improve the generalization performance of the algorithm. With only a small amount of data samples, diversified fault data can be generated, improving the learning and generalization capabilities of the diagnostic model, thereby achieving high-precision fault diagnosis.

[0108] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0109] Based on the same inventive concept, the embodiment of the present application also provides a small sample fault diagnosis device based on self-supervised learning generative adversarial network for implementing the small sample fault diagnosis method based on self-supervised learning generative adversarial network involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more embodiments of the small sample fault diagnosis device based on self-supervised learning generative adversarial network provided below can refer to the limitations of the small sample fault diagnosis method based on self-supervised learning generative adversarial network above, and will not be repeated here.

[0110] In one embodiment, Figure 6As shown, a small sample fault diagnosis device 600 based on self-supervised learning generative adversarial network is provided, comprising: a data acquisition module 601, a model construction module 603, a meta-learning generative adversarial network model training module 605, a self-supervised training module 607 and a fault diagnosis result determination module 609, wherein:

[0111] The data acquisition module 601 is used to collect the acoustic emission signal of the lifting system as real small sample fault data.

[0112] The model building module 603 is used to build a meta-learning generative adversarial network model and a self-supervised classification network, and perform parameter initialization, wherein the meta-learning generative adversarial network model is composed of a base learner and a meta-learner, and the self-supervised classification network is composed of a feature extractor, a category classifier and a translation scale classifier.

[0113] The meta-learning generative adversarial network model training module 605 is used to train the meta-learning generative adversarial network model based on the real small sample fault data to obtain the optimal initialization model and generate small sample data.

[0114] The self-supervised training module 607 is used to input the generated small sample data into the self-supervised classification network for self-supervised training.

[0115] The fault diagnosis result determination module 609 is used to input the fault data to be diagnosed into the optimal initialization model, output generated data, input the generated data into the trained self-supervised classification network, perform fault classification, and determine the fault diagnosis result.

[0116] In one embodiment of the present application, the base learner and the meta learner each include a generator and a discriminator, the generator of the base learner is used to generate data, the discriminator of the base learner is used to distinguish the authenticity of the generated data, and the generator and discriminator of the meta learner are used to balance the learning effect of the base learner.

[0117] In one embodiment of the present application, the meta-learning generative adversarial network model training module is also used to:

[0118] Use the model parameters of the meta-learner to initialize the model parameters of each base learner, perform adversarial training updates on each base learner using the data from the respective task support set, and calculate the losses of the generator and discriminator of the base learner;

[0119] Compute the losses of the generator and discriminator of the meta-learner and update the parameters of the meta-learner.

[0120] In one embodiment of the present application, the loss functions of the generator and the discriminator of the base learner are respectively expressed as:

[0121]

[0122]

[0123] Among them, θ G is the parameter of the generator, θ D are the parameters of the discriminator, For the generator, is the discriminator, z is the noise, x is the generated data, X is the real data set, and Z is the noise spatial distribution of the generator input.

[0124] In one embodiment of the present application, the loss functions of the generator and the discriminator of the meta-learner are respectively expressed as:

[0125]

[0126]

[0127] Among them, θ G is the parameter of the generator, θ D are the parameters of the discriminator, For the generator, is the discriminator, and To train the parameters of the generator and discriminator of the updated base learner, z is the noise, x is the generated data, X is the real data set, Z is the noise spatial distribution of the generator input, and P(T) is the spatial distribution of this task.

[0128] In one embodiment of the present application, the self-supervised training module is further used to:

[0129] The waveform generating the small sample data is translated, a waveform structure information label is established, and feature extraction is performed through a feature extractor;

[0130] Perform traditional category classification and translation scale classification, and calculate the category classification loss and translation scale classification loss respectively.

[0131] In one embodiment of the present application, the calculation process of the category classification loss and the translation scale classification loss is expressed as:

[0132]

[0133] Among them, λ is the balance parameter between category classification loss and translation scale classification loss, L classification is the category classification loss for traditional category classification, L self-supervised Translation-scale classification loss for translation-scale classification.

[0134] Each module in the above-mentioned small sample fault diagnosis device based on self-supervised learning generative adversarial network can be implemented in whole or in part by software, hardware and their combination. Each of the above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.

[0135] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 7 As shown. The computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, a mobile cellular network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, a small sample fault diagnosis method based on self-supervised learning to generate an adversarial network is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a key, trackball or touchpad set on the computer device shell, or an external keyboard, touchpad or mouse.

[0136] Those skilled in the art will understand that Figure 7 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0137] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0138] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0139] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0140] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0141] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0142] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0143] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A small sample fault diagnosis method based on self-supervised learning generative adversarial network, characterized in that: The method comprises: Collect the acoustic emission signals of the lifting system as real small sample fault data; Constructing a meta-learning generative adversarial network model and a self-supervised classification network, and initializing parameters, wherein the meta-learning generative adversarial network model is composed of a base learner and a meta-learner, and the self-supervised classification network is composed of a feature extractor, a category classifier, and a translation scale classifier; Based on the real small sample fault data, the meta-learning generative adversarial network model is trained to obtain an optimal initialization model and generate small sample data; Inputting the generated small sample data into the self-supervised classification network for self-supervised training; The fault data to be diagnosed is input into the optimal initialization model, and the generated data is output. The generated data is input into the trained self-supervised classification network to perform fault classification and determine the fault diagnosis result.

2. The small sample fault diagnosis method based on self-supervised learning generative adversarial network according to claim 1 is characterized in that: The base learner and the meta learner each include a generator and a discriminator, the generator of the base learner is used to generate data, the discriminator of the base learner is used to distinguish the authenticity of the generated data, and the generator and discriminator of the meta learner are used to balance the learning effect of the base learner.

3. The small sample fault diagnosis method based on self-supervised learning generative adversarial network according to claim 2 is characterized in that: The training of the meta-learning generative adversarial network model based on the real small sample fault data includes: Use the model parameters of the meta-learner to initialize the model parameters of each base learner, perform adversarial training updates on each base learner using the data from the respective task support set, and calculate the losses of the generator and discriminator of the base learner; Compute the losses of the generator and discriminator of the meta-learner and update the parameters of the meta-learner.

4. The small sample fault diagnosis method based on self-supervised learning generative adversarial network according to claim 3 is characterized in that: The loss functions of the generator and discriminator of the base learner are respectively expressed as: Among them, θ G is the parameter of the generator, θ D are the parameters of the discriminator, For the generator, is the discriminator, z is the noise, x is the generated data, X is the real data set, and Z is the noise spatial distribution of the generator input.

5. The small sample fault diagnosis method based on self-supervised learning generative adversarial network according to claim 3 is characterized in that: The loss functions of the generator and discriminator of the meta-learner are expressed as: Among them, θ G is the parameter of the generator, θ D are the parameters of the discriminator, For the generator, is the discriminator, and To train the parameters of the generator and discriminator of the updated base learner, z is the noise, x is the generated data, X is the real data set, Z is the noise spatial distribution of the generator input, and P(T) is the spatial distribution of this task.

6. The small sample fault diagnosis method based on self-supervised learning generative adversarial network according to claim 1 is characterized in that: The step of inputting the generated small sample data into the self-supervised classification network for self-supervised training comprises: The waveform generating the small sample data is translated, a waveform structure information label is established, and feature extraction is performed through a feature extractor; Perform traditional category classification and translation scale classification, and calculate category classification loss and translation scale classification loss respectively; Parameters of the feature extractor, the category classifier, and the translation-scale classifier are adjusted based on the category classification loss and the translation-scale classification loss.

7. The small sample fault diagnosis method based on self-supervised learning generative adversarial network according to claim 1 is characterized in that: The calculation process of the category classification loss and the translation scale classification loss is expressed as: Among them, λ is the balance parameter between category classification loss and translation scale classification loss, L classification is the category classification loss for traditional category classification, L self-supervised Translation-scale classification loss for translation-scale classification.

8. A small sample fault diagnosis device based on self-supervised learning generative adversarial network, characterized in that: The device comprises: A data acquisition module is used to collect the acoustic emission signals of the lifting system as real small sample fault data; A model building module, used to build a meta-learning generative adversarial network model and a self-supervised classification network, and perform parameter initialization, wherein the meta-learning generative adversarial network model is composed of a base learner and a meta-learner, and the self-supervised classification network is composed of a feature extractor, a category classifier, and a translation scale classifier; A meta-learning generative adversarial network model training module is used to train the meta-learning generative adversarial network model based on the real small sample fault data to obtain an optimal initialization model and generate small sample data; A self-supervised training module, used for inputting the generated small sample data into the self-supervised classification network for self-supervised training; The fault diagnosis result determination module is used to input the fault data to be diagnosed into the optimal initialization model, output generated data, input the generated data into the trained self-supervised classification network, perform fault classification, and determine the fault diagnosis result.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are 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 steps of the method according to any one of claims 1 to 7 are implemented.

Images