A mechanical intelligent fault diagnosis method, system, device, medium and product for data imbalance

Abstract

The application discloses a mechanical intelligent fault diagnosis method and system for data imbalance, equipment, medium and product, and relates to the field of mechanical fault diagnosis. First, the application constructs a generalization classification regularization generative adversarial network composed of a random sampling module, a generation module and a generalization module; second, the original vibration data is used as the generator input to accelerate the generator training process; in addition, a regularization loss term is introduced into the discriminator loss function to stabilize the training process of the generative adversarial network under the input of little data; finally, a classifier module is developed to distinguish the original samples from the generated samples to facilitate fault classification. The application can efficiently generate strong generalization samples by using only a small amount of data, and is suitable for intelligent fault diagnosis scenes with insufficient training data or unbalanced training data, so as to improve the performance, efficiency and adaptability of the intelligent fault diagnosis method.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of mechanical fault diagnosis, and in particular to a mechanical intelligent fault diagnosis method and system for data imbalance, equipment, medium and product. BACKGROUND

[0002] Fault diagnosis is considered as one of the key technologies to ensure the safe and reliable operation of rotating machinery. In recent years, with the continuous development of artificial intelligence, intelligent fault diagnosis technology has been quite deeply researched. Deep neural networks represented by convolutional neural networks can effectively extract fault information from massive data and have been widely applied in the field of intelligent fault diagnosis. Although many intelligent diagnosis methods have been developed, their application in actual engineering scenarios is still challenging due to the limited labeled training data. Deep neural networks require rich data to ensure excellent diagnostic performance. However, in actual engineering scenarios, rotating machinery usually does not allow to run with faults, and even if it runs, the duration is very short, which leads to a large number of healthy samples but very limited fault samples. This causes data imbalance between healthy samples and fault samples, which can easily cause misdiagnosis when using deep neural networks for fault diagnosis. Therefore, it is of great practical significance to study mechanical fault diagnosis under data imbalance in actual engineering application scenarios.

[0003] To address the problem of data imbalance, data augmentation to expand limited fault data is one of the effective solutions. Data augmentation methods can be mainly divided into sampling-based methods, feature diversity-based methods, and synthetic-based methods. Sampling-based methods include adding noise, overlapping segmentation strategies, etc. These methods aim to maximize the use of limited data, however, this method may cause the loss of some key fault information. Feature diversity-based methods extract different feature representations from limited data, but the performance improvement space is limited, especially in the case of extremely few samples. Synthetic-based methods include synthetic minority over-sampling technique (SMOTE) and generative adversarial network (GAN). However, in the case of very limited data, the samples generated by the SMOTE-based method are of low quality.

[0004] Due to the advantages of feature self-extraction and deep network architecture, GAN can generate samples through the adversarial training of the designed generator and discriminator. Many improved GAN variants have appeared to improve the quality of data augmentation. In addition, in order to improve the quality of generated samples, data preprocessing techniques are usually applied before feeding the original vibration signal to GAN. Typical data preprocessing means include the use of two-dimensional (2D) representation and time-frequency analysis methods.

[0005] Based on the above description, the existing fault diagnosis method has the following shortcomings:

[0006] 1) The input of the generator in the GAN is mostly random noise or random noise with some labels, which cannot fully capture the diverse feature representation of the data, reducing the convergence speed of the model.

[0007] 2) The generalization ability of the GAN needs to be improved because the quality of the generated samples is low under the condition of extremely limited training data.

[0008] 3) The 2D image calculation time is long, which will reduce the efficiency of fault diagnosis.

[0009] 4) The time-frequency analysis method requires prior knowledge and professional knowledge of signal processing, which may reduce the adaptability of the intelligent diagnosis method. SUMMARY

[0010] The purpose of the present application is to provide a data imbalance-oriented mechanical intelligent fault diagnosis method, system, device, medium and product, which can efficiently generate high-quality samples from a small amount of original vibration signals under the condition of data imbalance and extremely limited training data, thereby improving the performance, efficiency and adaptability of the intelligent fault diagnosis method.

[0011] To achieve the above purpose, the present application provides the following solutions:

[0012] In a first aspect, the present application provides a data imbalance-oriented mechanical intelligent fault diagnosis method, comprising:

[0013] Obtaining vibration experimental data under different mechanical states, and generating a data set by taking the vibration experimental data as sample data; the mechanical states include a healthy state and a plurality of fault states;

[0014] Dividing the data set to obtain a training set and a test set; the number of vibration experimental data of the healthy state in the training set is greater than the number of vibration experimental data of the fault state; the number of vibration experimental data of the healthy state in the test set is balanced with the number of vibration experimental data of the fault state;

[0015] Constructing a generalization classification regularization generative adversarial network;

[0016] Training the generalization classification regularization generative adversarial network using the training set until a first set condition is met, to obtain an original sample data set, a generalization sample and a trained generalization classification regularization generative adversarial network;

[0017] Verifying whether the trained generalization classification regularization generative adversarial network meets a second set condition using the original sample data set and the generalization sample;

[0018] adjusting parameters of the generalized classification regularization generative adversarial network when the second set condition is not met, and returning to the step of training the generalized classification regularization generative adversarial network with the training set until the first set condition is met, obtaining the original sample data set, the generalized sample, and the trained generalized classification regularization generative adversarial network;

[0019] when the second set condition is met, using the trained generalized classification regularization generative adversarial network as a sample generation model;

[0020] inputting the vibration experimental data of the fault state in the training set into the sample generation model to obtain a generated sample, and constructing a balanced training set with the vibration experimental data of the healthy state in the training set;

[0021] constructing a fault diagnosis classifier;

[0022] training the fault diagnosis classifier with the balanced training set until the third set condition is met, obtaining the trained fault diagnosis classifier;

[0023] using the test set data or obtaining real-time vibration data of the mechanical equipment to be diagnosed as to-be-diagnosed data, and inputting the to-be-diagnosed data into the fault diagnosis classifier to obtain a mechanical fault diagnosis result.

[0024] Optionally, the generalized classification regularization generative adversarial network is constructed, and specifically includes:

[0025] constructing a generator module; the generator module is configured to sample and operate the vibration experimental data of the fault state in the training set to obtain an original sample data set, a candidate data set, and a generalized sample; the original sample data set includes a signal space and a label space corresponding to the signal space; the generalized sample is used to constitute an optimization target and a loss function of the generator module;

[0026] constructing a classifier module; the classifier module is configured to perform fault classification based on the generalized sample and the original sample data set to obtain an output loss, and apply the output loss to the generator module;

[0027] constructing a discriminator module; the discriminator module is configured to obtain a true sample score and a false sample score based on the generalized sample and the original sample data set; samples in the original sample data set are taken as true samples, and the generalized sample is taken as a false sample.

[0028] Optionally, the generator module is constructed, and specifically includes:

[0029] constructing a random sampling module; the random sampling module is configured to sequentially sample and randomly sample the vibration experiment data of the fault states in the training set to obtain an original sample data set and a candidate data set;

[0030] constructing an initial sample generation module; the initial sample generation module is configured to obtain an initial sample based on the original sample data set;

[0031] constructing a generalization module; the generalization module is configured to calculate a consistency correlation coefficient of the candidate data set and the initial sample, and generate a generalization sample.

[0032] Optionally, the initial sample generation module is specifically constructed and includes:

[0033] constructing a first signal / label feature learning module; the first signal / label feature learning module is configured to obtain original sample features based on the original sample data set;

[0034] constructing a generation module; the generation module is configured to obtain the initial sample based on the original sample features.

[0035] Optionally, the discriminator module is specifically constructed and includes:

[0036] constructing a second signal / label feature learning module; the second signal / label feature learning module is configured to obtain sample features based on the generalization sample and the original sample data set;

[0037] constructing a discriminator module; the discriminator module is configured to obtain true sample scores and false sample scores based on the sample features.

[0038] In a second aspect, the present application provides a mechanical intelligent fault diagnosis system for data imbalance, which is used to implement the mechanical intelligent fault diagnosis method for data imbalance provided above; the system includes:

[0039] a data acquisition module, configured to acquire vibration experiment data under different mechanical states, and generate a data set by taking the vibration experiment data as sample data; the mechanical states include a healthy state and multiple fault states;

[0040] a data division module, configured to divide the data set to obtain a training set and a test set; the number of vibration experiment data of the healthy state in the training set is greater than the number of vibration experiment data of the fault states; the number of vibration experiment data of the healthy state in the test set is balanced with the number of vibration experiment data of the fault states;

[0041] a network training module, configured to construct a generalization classification regularization generative adversarial network, and train the generalization classification regularization generative adversarial network using the training set until a first preset condition is met, to obtain an original sample data set, a generalization sample, and a trained generalization classification regularization generative adversarial network;

[0042] a network verification determination module, configured to verify whether the trained generalization classification regularization generative adversarial network meets a second preset condition using the original sample data set and the generalization sample; when the second preset condition is not met, adjusting parameters of the generalization classification regularization generative adversarial network, and returning to execute steps in the network training module; when the second preset condition is met, taking the trained generalization classification regularization generative adversarial network as a sample generation model;

[0043] a training set balancing module, configured to input vibration experiment data of a fault state in the training set to the sample generation model to obtain a generated sample, and construct a balanced training set together with vibration experiment data of a healthy state in the training set;

[0044] a classifier construction training module, configured to construct a fault diagnosis classifier; and train the fault diagnosis classifier using the balanced training set until a third preset condition is met, to obtain a trained fault diagnosis classifier;

[0045] a fault diagnosis module, configured to obtain to-be-diagnosed data including the test set data and real-time vibration data of a to-be-diagnosed mechanical equipment, and input the to-be-diagnosed data into the trained fault diagnosis classifier to obtain a fault diagnosis result.

[0046] In a third aspect, a computer device is provided, including a memory, a processor, a computer program stored in the memory and executable on the processor, and the processor executes the computer program to implement steps of the mechanical intelligent fault diagnosis method for data imbalance.

[0047] In a fourth aspect, a computer readable storage medium is provided, and the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement steps of the mechanical intelligent fault diagnosis method for data imbalance.

[0048] In a fifth aspect, a computer program product is provided, including a computer program, and the computer program is executed by a processor to implement steps of the mechanical intelligent fault diagnosis method for data imbalance.

[0049] According to the embodiments provided in the present application, the following technical effects are achieved:

[0050] The application provides a mechanical intelligent fault diagnosis method and system for data imbalance, a device, a medium and a product. BRIEF DESCRIPTION OF DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0052] Figure 1 An application environment diagram of a mechanical intelligent fault diagnosis method for data imbalance in an embodiment of the present application;

[0053] Figure 2 A flowchart of a mechanical intelligent fault diagnosis method for data imbalance provided in an embodiment of the present application;

[0054] Figure 3 A generalization classification regularization generative adversarial network structure diagram provided in an embodiment of the present application;

[0055] Figure 4 An implementation flowchart of a mechanical intelligent fault diagnosis method for data imbalance provided in another embodiment of the present application;

[0056] Figure 5 A result diagram of fault diagnosis by using test set data to train and generate samples by using different data enhancement methods in the case of data imbalance and extremely limited data in an embodiment of the present application;

[0057] Figure 6 An ablation experiment result diagram provided in an embodiment of the present application;

[0058] Figure 7 A structure diagram of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0059] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0060] The above objects, features and advantages of the present application will become more apparent from the following detailed description considered in conjunction with the accompanying drawings and specific embodiments.

[0061] The mechanical intelligent fault diagnosis method for data imbalance provided in the embodiments of the present application can be applied in an application environment as shown in Figure 1 The terminal 102 communicates with the server 104 through a network. The data storage system can store data required to be processed by the server 104. The data storage system can be separately arranged, integrated on the server 104, placed on a cloud or other servers. The terminal 102 can send the vibration data of the mechanical equipment to be diagnosed to the server 104. After receiving the vibration data of the mechanical equipment to be diagnosed, the server 104 inputs the real-time acquired vibration data into the fault diagnosis classifier to obtain the mechanical fault diagnosis result. The server 104 can feed back the obtained mechanical fault diagnosis result to the terminal 102. In addition, in some embodiments, the mechanical intelligent fault diagnosis method for data imbalance can also be implemented by the server 104 or the terminal 102 alone, for example, the terminal 102 can directly perform mechanical intelligent fault diagnosis on the vibration data of the mechanical equipment to be diagnosed, or the server 104 can acquire the vibration data of the mechanical equipment to be diagnosed from the data storage system and perform mechanical intelligent fault diagnosis on the vibration data of the mechanical equipment to be diagnosed.

[0062] The terminal 102 can be, but is not limited to, various desktop computers, notebook computers, smart phones, tablet computers, Internet of Things devices and portable wearable devices. The Internet of Things device can be a smart speaker, a smart television, a smart air conditioner, a smart vehicle device, etc. The portable wearable device can be a smart watch, a smart bracelet, a head-mounted device, etc. The server 104 can be implemented by an independent server or a server cluster composed of multiple servers, and can also be a cloud server.

[0063] In an exemplary embodiment, as shown in Figure 2 A mechanical intelligent fault diagnosis method for data imbalance is provided. The method is executed by a computer device, which can be executed by a terminal or a server alone or by a terminal and a server together. In the embodiments of the present application, the method is applied to the server 104 in Figure 1 The method includes the following steps 200 to 210. Wherein:

[0064] Step 200: Acquire vibration experimental data under different mechanical states, and generate a data set by taking the vibration experimental data as sample data. The mechanical states include a healthy state and multiple fault states.

[0065] In practical application, the training set (i.e., the vibration data set for training) and the test set (i.e., the vibration data set for testing) in different fault states can be obtained through sensors installed on the rotating machinery. The test set is used to verify the performance of the trained generalization classification regularization generative adversarial network in subsequent verification. The data in the data set are mainly vibration signal data of the gear, and these data can reflect different health states of the gear. The different health states of the gear include normal gear, gear fracture, gear tooth loss, gear root crack, and gear surface wear.

[0066] Step 201: Divide the data set to obtain the training set and the test set. The number of vibration experimental data of the health state in the training set is much larger than that of the fault state, that is, there are a large number of health state samples and a small number of fault state samples in the training set. The number of vibration experimental data of the health state in the test set is balanced with that of the fault state.

[0067] Step 202: Construct the generalization classification regularization generative adversarial network. The generalization classification regularization generative adversarial network is essentially a conditional generative adversarial network with original signal feature representation as the network input, a generalization module and a classifier module for auxiliary training, and an L2 norm regularization term added to the discriminator objective function. The trained generalization classification regularization generative adversarial network takes vibration data as input and generates samples as output. The constructed generalization classification regularization generative adversarial network is as shown in Figure 3 .

[0068] Step 203: Train the generalization classification regularization generative adversarial network with the training set until the first set condition is met, and obtain the original sample data set, the generalization sample, and the trained generalization classification regularization generative adversarial network.

[0069] Step 204: Verify whether the trained generalization classification regularization generative adversarial network meets the second set condition using the original sample data set and the generalization sample.

[0070] Step 205: When the second set condition is not met, adjust the parameters of the generalization classification regularization generative adversarial network and return to step 203.

[0071] Step 206: When the second set condition is met, the trained generalization classification regularization generative adversarial network is used as a sample generation model.

[0072] Step 207: Input the vibration experimental data of the fault state in the training set into the sample generation model to obtain the generated sample, and construct a balanced training set with the vibration experimental data of the health state in the training set.

[0073] Step 208: Construct a fault diagnosis classifier. The fault diagnosis model takes the original sample dataset and generalized samples as network inputs and outputs the fault classification results of the samples.

[0074] Step 209: Train the fault diagnosis classifier using a balanced training set until the third set condition is met, and then obtain the trained fault diagnosis classifier.

[0075] Step 210: Obtain vibration data of the mechanical equipment to be diagnosed, and input the vibration data into the fault diagnosis classifier to obtain the mechanical fault diagnosis result.

[0076] By implementing steps 200 to 210 above and constructing a sample generation model, high-quality samples can be efficiently generated from a small amount of raw vibration signals in the vibration data of the mechanical equipment to be diagnosed, even under conditions of imbalanced data and extremely limited training data. This improves the efficiency of high-quality sample generation. When using deep neural network-based methods for fault diagnosis, employing these high-quality samples can improve the performance and efficiency of the fault diagnosis classifier, making neural network-based fault diagnosis methods more adaptable to scenarios with imbalanced data and extremely limited training data.

[0077] In another exemplary embodiment of this application, in order to improve the quality of generated samples and effectively enhance their generalization ability with extremely limited training data, step 202 essentially involves initializing the generalization classification regularized generative adversarial network to fully learn the data distribution and output a balanced training set, such as... Figure 4 As shown, the process of constructing a generalized classification regularized generative adversarial network includes:

[0078] (1) Constructing the generator module. The generator module is configured to sample and process vibration experimental data of fault states in the training set to obtain the original sample dataset, candidate dataset, and generalized samples. The original sample dataset includes the signal space and the corresponding label space. The generalized samples are used to construct the optimization objective and loss function of the generator module.

[0079] The process of building the generator module includes:

[0080] (1-1) Constructing a random sampling module. The random sampling module is configured to perform sequential and random sampling on the vibration experimental data of fault states in the training set to obtain the original sample dataset and the candidate dataset.

[0081] In practical applications, original sample datasets and candidate datasets are obtained with a very limited number of original vibration signals. The extremely limited vibration dataset under c fault states can be represented as L. D =[L1,L2,…,L c ], where Li denotes the vibration data set in the i-th state, and the data length is L in each state. Sequential sampling is performed in each state to generate m samples, and the length of each sample is l. The sampling is performed according to the index R denotes the i-th sample after sampling in the c-th state, and [·] denotes the index according to the index R D Sampling is performed, and x i denotes the vibration signal data point in each state, and finally a signal space is generated, denoted as The i-th sample has a label Corresponding label space is generated The two constitute the original sample data set.

[0082] At the same time, L D is randomly sampled to construct two different candidate data sets with corresponding labels, that is, and This process can be described as follows:

[0083]

[0084]

[0085] In the formula, and The generation principle of the data set is similar to , but the index R i and R j are randomly generated natural numbers. and denote the two candidate data sets.

[0086] (1-2) Construct a generalization module. The generalization module is configured to calculate the consistency correlation coefficient of the candidate data set and the initial sample, and generate a generalization sample. The output of the random sampling module is the input of the generalization module, and the consistency correlation coefficient of the initial generated sample and the original sample is calculated in combination with the two candidate data sets to generate a generalization sample. Take α i and β i as the consistency correlation coefficient of the i-th initial generated sample and the original sample in and , and the formula is as follows:

[0087]

[0088] In the formula, ρ i and ρ i ' are the Pearson correlation coefficients between the i-th initial generated sample and the samples in the two candidate data sets, and are the data sets and the standard deviation of the signal space X and respectively represent the mean of the signal space X r and .

[0089] For the final generated generalization sample can be expressed as:

[0090]

[0091] The generalization sample constitutes the optimization objective and loss function of the proposed generator module, which is expressed as:

[0092]

[0093] In the formula, L G is the loss of the generator module, represents the minimization of the loss function by adjusting the parameters of the generator module G, represents the expectation when the input data follows the sample distribution of the X r data set, g G (·) is the output layer that converts the results calculated by the discriminator D into sample classification scores, c C (·) is the output layer that converts the results calculated by the classifier C into sample classification scores, Y g is the label of the fake sample, θ c is the training parameter of the classifier.

[0094] (2) Construct an initial sample generation module. The initial sample generation module is configured to obtain initial samples based on the original sample data set.

[0095] In actual application process, the construction process of the initial sample generation module includes:

[0096] (2-1) Construct a first signal / label feature learning module. The first signal / label feature learning module is configured to obtain original sample features based on the original sample data set.

[0097] The original sample is input into the generator to fully capture different feature representations and generate an initial sample. To avoid the label features being covered by the signal features, two independent modules are used to realize the feature learning of signals and labels. Signals and labels are converted into high-level transferable features through two fully connected (FC) layers. To avoid gradient explosion and disappearance, normalization and linear rectification function (ReLu) activation functions are used in the network. Considering the model performance and computational efficiency, batch normalization (BN) and instance normalization (IN) are used alternately. After initial feature learning, the signal feature f x and the label feature f y are concatenated by the Concatenated(·) function and then input into the generation module. The calculation method of this process is as follows:

[0098] f x =F x (X r ,θ g ), f y =F y (Y r ,θ g ).

[0099] f X =Concatenated(f x ,f y ).

[0100] In the formula, F x (·) and F y (·) represent the signal and label feature learners respectively, θ g represents the training parameters of the generator, f x and f y represent the signal features and label features, and f X represents the original sample features.

[0101] The generation module is constructed as shown in (2-2). The generation module is configured to obtain an initial sample based on the original sample features.

[0102] In the subsequent processing process, the output of the random sampling module is the input of the generation module, that is, the original sample features f X will be input into the generation module to generate an initial sample The generation module is composed of a convolution layer (Conv) and two fully connected layers (FC). Assuming that g(·) represents the generation module, the initial generated sample can be described as:

[0103]

[0104] (3) Constructing a classifier module. The classifier module is configured to perform fault classification based on the generalized samples and the original sample dataset, obtain an output loss, and apply the output loss to the generator module.

[0105] The purpose of constructing the classifier module is to enable the generator module to pay more attention to class differences, thereby improving the stability of the generated samples. Finally, the classifier module outputs the loss and to the generator module. Wherein, the optimization objective and the loss function formula are as follows:

[0106]

[0107] In the formula, represents the minimization of the loss function value by adjusting the parameters of the classifier module C. represents the expectation when the data follows the sample distribution of X r .

[0108] (4) Constructing a discriminator module. The discriminator module is configured to obtain true sample scores and false sample scores based on the generalized samples and the original sample dataset. The samples in the original sample dataset are taken as true samples, and the generalized samples are taken as false samples. That is, input the original sample (true sample) or the generalized sample (false sample), and classify the sample.

[0109] In order to improve the generalization ability of the sample and the stability of the model, a discriminator module containing a new regularization loss function is developed. Also, in order to increase the weight of the label feature in the entire sample, a second signal / label feature learning module is proposed. Then, the signal and the label feature will be concatenated in the concatenation layer and input into the discriminator module. Finally, the discriminator module will output the scores of discriminating true samples and false samples. Its description is as follows:

[0110]

[0111] In the formula, S F and S R represent the false sample score and the true sample score, respectively. represents the expectation when the data follows the sample distribution of X or . f D and f G are output layers that convert the results calculated by the discriminator module D into classification scores of true samples and false samples, respectively. θ d represents the training parameters of the discriminator. ​

[0112] In order to improve the stability of the trained model, a new regularization term R D (D) is introduced in the discriminator module loss L LC (D), and the optimization objective and loss function formula are expressed as:

[0113]

[0114] In the formula, L D is the discriminator module loss, indicates that the function value is minimized by adjusting the parameters of the discriminator module D, λ represents the coefficient of the regularization term, and α F and α R are two variables called anchor points, which are used to track the predictions of the discriminator for true samples and false samples. ||·|| 2 represents the quadratic norm. R LC (D) is an intermediate quantity and plays a role in formula simplification.

[0115] Based on this, the process of constructing the discriminator module can be described as follows:

[0116] (4-1) Constructing a second signal / label feature learning module. The second signal / label feature learning module is configured to obtain sample features based on the generalization sample and the original sample dataset.

[0117] (4-2) Constructing a discrimination module. The discrimination module is configured to obtain true sample scores and false sample scores based on the sample features.

[0118] In another exemplary embodiment of the present application, based on the specific construction process of the generalization classification regularization generative adversarial network given above, the training method can be:

[0119] The samples in the original sample dataset are sequentially input into the constructed generalization classification regularization generative adversarial network, and feature extraction, initial sample generation, generalization sample generation, discriminator parameter update, classifier parameter update, and generator parameter update are sequentially performed for each input sample. Loop until the generalization classification regularization generative adversarial network reaches Nash equilibrium (i.e., satisfies the first setting condition), and a trained generalization classification regularization generative adversarial network is obtained.

[0120] In another exemplary embodiment of the present application, the construction process of the balanced training set in step 207 can be described as:

[0121] (1) input the vibration experimental data of the fault state in the training set into the sample generation model to obtain generated samples under different fault states. In this embodiment, the essence is to input extremely limited fault data into the trained generalization classification regularization generative adversarial network to obtain generated samples under different fault states.

[0122] (2) combine the generated samples with the vibration experimental data of the healthy state in the training set to construct a balanced training set.

[0123] For example, in the actual application process, the samples in the original sample data set X r are input into the trained generalization classification regularization generative adversarial network, and the generated generalization samples of a single sample are Xg. Until a specified number of samples of all fault types are generated, they are combined with the healthy samples to obtain a balanced training set.

[0124] Further, in the actual application process, in order to further ensure the integrity of the generated balanced training set and improve the fault diagnosis accuracy, the Fourier transform visualization technology can be used to evaluate the quality of the generated samples. If necessary, the generalization classification regularization generative adversarial network is retrained. The generated samples and the original samples are subjected to Fourier transform, and the quality of the generated samples is analyzed according to the differences between the two kinds of samples in the time domain and the frequency domain. If the quality of the generated samples cannot meet the similarity requirement of the original samples, the generalization classification regularization generative adversarial network is retrained until the generated samples are highly similar to the original samples in the frequency domain (i.e., the second set condition is met). Finally, the trained generalization classification regularization generative adversarial network is saved, the unbalanced training set is expanded using the trained generalization classification regularization generative adversarial network, and the balanced training set is generated.

[0125] (3) train the fault diagnosis classifier using the balanced training set until the classification accuracy of the fault diagnosis classifier converges (i.e., the third set condition is met), and then use the trained fault diagnosis classifier as the fault diagnosis model. In the application process, the balanced training set can be divided into a training set and a validation set to complete the training of the fault diagnosis classifier.

[0126] Based on the above description, the application first introduces the original sample data set and the candidate data set, and does sequence sampling and random sampling on the extremely limited fault samples. Secondly, the feature learning and initial sample generation are performed on the original data by using the generation module in the generalization classification regularization generative adversarial network, and the generalization sample is generated in combination with the candidate data set. Then, the generation performance is improved with the assistance of the discriminator module and the classifier module, the L2 norm regularization is integrated into the loss function of the discriminator module, after training, the loss function of the generalization classification regularization generative adversarial network reaches Nash equilibrium, and the generalization classification regularization generative adversarial network will output samples. Subsequently, the trained generator module is extracted, and the vibration experimental data of the fault state in the training set are used to generate samples, which are combined with the vibration experimental data of the healthy state in the training set to form a balanced training set. Finally, the fault diagnosis classifier is trained and used to classify the real-time acquired vibration data. The application fully considers the influence of the imbalance of the data in the industrial scene and the extremely limited training data on the diagnosis, and provides a new method for the state monitoring and fault diagnosis of the rotating machinery, and provides important technical support for the safe and reliable operation of the rotating machinery equipment.

[0127] In another exemplary embodiment of the application, as shown in Figure 5 The generalization classification regularization generative adversarial network (GCRGAN) is the method proposed in the application. CGAN represents the conditional generative adversarial network, WGAN-GP represents the generative adversarial network containing Wasserstein distance and gradient penalty term, ACGAN represents the generative adversarial network containing auxiliary classifier, ACWGAN-GP represents the ACGAN containing Wasserstein distance and gradient penalty term, Overlapping segmentation represents the sample expansion realized by using original data oversampling, and Direct diagnosis represents the fault diagnosis method using unbalanced data set and one-dimensional convolutional neural network. In this embodiment, the proposed GCRGAN method and the above method are used to identify the health state of the planetary gearbox test data. The results show that the proposed GCRGAN method can accurately and effectively identify the fault type of 2500 test samples (fault type unknown). In five repeated running tests, the highest result of the diagnosis accuracy of this method is higher than that of the existing advanced method, and the average diagnosis accuracy reaches 99.28%, and the standard deviation is 0.19%, which proves the high accuracy and strong stability of the proposed GCRGAN method. This reflects the superiority of the provided method in data enhancement under the condition of data imbalance and extremely limited data.

[0128] In another exemplary embodiment of the application, as shown inFigure 6 As shown, ablation experiments were performed to verify the role of each component in the method provided in the application. It should be noted that the GCRGAN (random noise) method means that the input of the generator in the GCRGAN is changed to random noise, the GCGAN method means that the GCRGAN method lacks an L2 regularization loss function item in the discriminator loss, the GRGAN means that the GCRGAN method lacks a classifier module in the generator module, and the CRGAN means that the GCRGAN method lacks a generalization module in the generator module. Figure 6 The convergence curves and running times of the five methods are shown in FIG. 1, wherein Figure 6 Part (a) of FIG. 1 is a schematic diagram of the ablation experiment results of the method provided in the application, Figure 6 Part (b) of FIG. 1 is a schematic diagram of the ablation experiment results of the GCRGAN (random noise) method, Figure 6 Part (c) of FIG. 1 is a schematic diagram of the ablation experiment results of the GCGAN method, Figure 6 Part (d) of FIG. 1 is a schematic diagram of the ablation experiment results of the GRGAN method, Figure 6 Part (e) of FIG. 1 is a schematic diagram of the ablation experiment results of the CRGAN method, Figure 6 Part (f) of FIG. 1 is a comparison diagram of the experiment results of the method provided in the application and the GCRGAN (random noise). Compared with the GCRGAN (random noise) method, the GCRGAN method provided in the application has lower generator loss and faster convergence speed. This indicates that inputting the original vibration signal can improve the performance of the GCRGAN. Compared with the GCGAN method, the generator loss of the GCRGAN method provided in the application is more stable, which indicates that it can generate more high-quality samples. Compared with the GRGAN method, the discriminator loss stability of the GCRGAN method provided in the application is stronger, which can better improve the generator. It is worth noting that the CRGAN method is over-fitted, which leads to its inability to generate high-quality samples.

[0129] In summary, compared with the prior art, the scheme provided in the application has the following improvements:

[0130] 1. In order to improve the quality of the generated samples, the application designs a new generator composed of a random sampling module, a generation module and a generalization module. The generator can effectively enhance the generalization ability of the generated samples under extremely limited training data.

[0131] 2. In order to fully capture different feature representations and accelerate model convergence, the application uses the original vibration signal as the input of the new generator. This method can improve the efficiency of high-quality sample generation under extremely limited training data.

[0132] 3. To further improve the generalization ability of the GAN model trained based on extremely limited training data, the application introduces a new regularization loss term in the loss function of the discriminator module to effectively stabilize the learning process of the GAN under extremely limited training data.

[0133] 4. To enable the generator to pay more attention to class differences and generate more high-quality samples, the application develops a classifier module in the proposed GAN model to distinguish between generated samples and real samples.

[0134] Based on the same inventive concept, the embodiments of the application also provide a data imbalance oriented mechanical intelligent fault diagnosis system for implementing the data imbalance oriented mechanical intelligent fault diagnosis method described above. The implementation scheme for solving problems provided by the system is similar to the implementation scheme described in the above method, so the specific limitations in one or more data imbalance oriented mechanical intelligent fault diagnosis system embodiments provided below can refer to the limitations of the data imbalance oriented mechanical intelligent fault diagnosis method described above, which will not be repeated here.

[0135] In an exemplary embodiment, the application provides a data imbalance oriented mechanical intelligent fault diagnosis system, which comprises:

[0136] A data acquisition module is configured to acquire vibration experimental data under different mechanical states and generate a data set by taking the vibration experimental data as sample data. The mechanical states include a healthy state and a plurality of fault states.

[0137] A data division module is configured to divide the data set to obtain a training set and a test set. The number of vibration experimental data of the healthy state in the training set is greater than the number of vibration experimental data of the fault state. The number of vibration experimental data of the healthy state in the test set is balanced with the number of vibration experimental data of the fault state.

[0138] A network construction and training module is configured to construct a generalization classification regularization generative adversarial network and train the generalization classification regularization generative adversarial network using the training set until a first set condition is met, to obtain an original sample data set, a generalization sample, and a trained generalization classification regularization generative adversarial network.

[0139] A network verification and determination module is configured to verify whether the trained generalization classification regularization generative adversarial network meets a second set condition using the original sample data set and the generalization sample. When the second set condition is not met, the parameters of the generalization classification regularization generative adversarial network are adjusted, and the step in the network training module is returned to be executed. When the second set condition is met, the trained generalization classification regularization generative adversarial network is taken as a sample generation model.

[0140] The training set balancing module is configured to input the vibration experimental data of the fault state in the training set into the sample generation model to obtain generated samples, and construct a balanced training set with the vibration experimental data of the healthy state in the training set.

[0141] The classifier construction training module is configured to construct a fault diagnosis classifier, train the fault diagnosis classifier with the balanced training set, and obtain the trained fault diagnosis classifier when a third set condition is met.

[0142] The fault diagnosis module is configured to obtain to-be-diagnosed data including test set data and real-time vibration data of the to-be-diagnosed mechanical equipment, and input the to-be-diagnosed data into the trained fault diagnosis classifier to obtain a fault diagnosis result.

[0143] Based on the above description, the present application has the following advantages relative to the prior art:

[0144] 1. The present application designs a feature learning module of original samples, takes the feature expression of samples and their categories as the input of the generator module, and accelerates the convergence speed of the model.

[0145] 2. The present application designs a conditional generative adversarial network with a loss function of L2 norm regularization, introduces the regularization term into the loss of the discriminator module, and makes the learning process of GAN more stable under extremely limited training data.

[0146] 3. The present application designs a classifier module and its loss function, and introduces them into the conditional generative adversarial network, so that the generator module can generate samples of higher quality.

[0147] 4. The present application proposes a sample generation method based on the consistency correlation coefficient. The consistency correlation coefficient of the initial generated sample and two candidate data sets is calculated to obtain a generalization sample by weighting the two kinds of samples, thereby improving the quality of the generated sample.

[0148] 5. The method proposed in the present application does not use a preprocessing method such as 2D image generation, the fault diagnosis classifier model is simple, real-time calculation is fast, and the fault diagnosis efficiency is high.

[0149] 6. The present application can perform state monitoring and fault diagnosis on rotating machinery under the condition that some category experimental data is extremely limited or the data set is unbalanced, can learn the distribution of extremely limited samples and complete the data, has no time-frequency analysis process, and widens the adaptability of the intelligent diagnosis method.

[0150] In an exemplary embodiment, a computer device, which can be a server or a terminal, is provided, and an internal structure diagram of the computer device can be as shown in Figure 7As shown in the figure. The computer device includes a processor, a memory, an input / output interface (I / O for short) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through the system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capability. 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, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store mechanical intelligent fault diagnosis data. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals through network connection. The computer program is executed by the processor to realize a mechanical intelligent fault diagnosis method for data imbalance.

[0151] Those skilled in the art can understand that, Figure 7 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement. In one exemplary embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to realize the steps in each of the above method embodiments.

[0152] In one exemplary embodiment, a computer readable storage medium is provided, storing a computer program, which is executed by a processor to realize the steps in each of the above method embodiments.

[0153] In one exemplary embodiment, a computer program product is provided, including a computer program, which is executed by a processor to realize the steps in each of the above method embodiments.

[0154] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant regulations.

[0155] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this 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), magnetic 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. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0156] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0157] The technical features of the above embodiments can be combined in any way. For the sake of brevity, 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.

[0158] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A mechanical intelligent fault diagnosis method for data imbalance, characterized in that, The mechanical intelligent fault diagnosis method facing data imbalance comprises: obtaining vibration experimental data under different mechanical states, and generating a data set by taking the vibration experimental data as sample data; the mechanical states include a healthy state and multiple fault states; dividing the data set to obtain a training set and a test set; the number of vibration experimental data of the healthy state in the training set is greater than the number of vibration experimental data of the fault state; the number of vibration experimental data of the healthy state in the test set is balanced with the number of vibration experimental data of the fault state; constructing a generalization classification regularization generative adversarial network; training the generalization classification regularization generative adversarial network by using the training set until a first set condition is met, to obtain an original sample data set, a generalization sample, and a trained generalization classification regularization generative adversarial network; verifying whether the trained generalization classification regularization generative adversarial network meets a second set condition by using the original sample data set and the generalization sample; when the second set condition is not met, adjusting parameters of the generalization classification regularization generative adversarial network, and returning to the step of training the generalization classification regularization generative adversarial network by using the training set until the first set condition is met, to obtain the original sample data set, the generalization sample, and the trained generalization classification regularization generative adversarial network; when the second set condition is met, taking the trained generalization classification regularization generative adversarial network as a sample generation model; inputting the vibration experimental data of the fault state in the training set into the sample generation model to obtain generated samples, and constructing a balanced training set with the vibration experimental data of the healthy state in the training set; constructing a fault diagnosis classifier; training the fault diagnosis classifier by using the balanced training set until a third set condition is met, to obtain a trained fault diagnosis classifier; taking the test set data or real-time vibration data of a mechanical equipment to be diagnosed as to-be-diagnosed data, and inputting the to-be-diagnosed data into the fault diagnosis classifier to obtain a mechanical fault diagnosis result; constructing a generalization classification regularization generative adversarial network comprises: constructing a generator module; the generator module is configured to sample and operate the vibration experimental data of the fault state in the training set to obtain an original sample data set, a candidate data set, and a generalization sample; the original sample data set comprises a signal space and a label space corresponding to the signal space; the generalization sample is used to constitute an optimization target and a loss function of the generator module; constructing a generator module comprises: constructing a random sampling module; the random sampling module is configured to sequentially sample and randomly sample the vibration experimental data of the fault state in the training set to obtain the original sample data set and the candidate data set; constructing an initial sample generation module; the initial sample generation module is configured to obtain an initial sample based on the original sample data set; constructing a generalization module; the generalization module is configured to calculate a consistency correlation coefficient of the candidate data set and the initial sample, and generate a generalization sample.

2. The method of claim 1, wherein, constructing a generalization classification regularization generative adversarial network further comprises constructing a classifier module; The classifier module is configured to perform fault classification based on the generalized sample and the original sample dataset, obtain an output loss, and apply the output loss to the generator module; An initial sample generation module is constructed, specifically including: A first signal / label feature learning module is constructed; the first signal / label feature learning module is configured to obtain original sample features based on the original sample dataset; A second signal / label feature learning module is constructed; the second signal / label feature learning module is configured to obtain sample features based on the generalized sample and the original sample dataset; 3. The method of claim 2, wherein, A discriminator module is constructed; the discriminator module is configured to obtain true sample scores and false sample scores based on the sample features. The data imbalance-oriented mechanical intelligent fault diagnosis system is used to implement the data imbalance-oriented mechanical intelligent fault diagnosis method according to any one of claims 1-4; the data imbalance-oriented mechanical intelligent fault diagnosis system includes: A data acquisition module is configured to acquire vibration experimental data under different mechanical states and generate a dataset by taking the vibration experimental data as sample data; the mechanical states include a healthy state and multiple fault states; 4. The method of claim 2, wherein, A data division module is configured to divide the dataset to obtain a training set and a test set; the number of vibration experimental data in the healthy state in the training set is greater than the number of vibration experimental data in the fault state; the number of vibration experimental data in the healthy state in the test set is balanced with the number of vibration experimental data in the fault state; A network construction and training module is configured to construct a generalized classification regularization generative adversarial network, train the generalized classification regularization generative adversarial network using the training set, and obtain an original sample dataset, a generalized sample, and a trained generalized classification regularization generative adversarial network when a first set condition is met; A network verification and determination module is configured to verify whether the trained generalized classification regularization generative adversarial network meets a second set condition using the original sample dataset and the generalized sample; when the second set condition is not met, the parameters of the generalized classification regularization generative adversarial network are adjusted, and the step in the network construction and training module is executed again; when the second set condition is met, the trained generalized classification regularization generative adversarial network is taken as a sample generation model; 5. A mechanical intelligent fault diagnosis system oriented to data imbalance, characterized in that, A training set balancing module is configured to input the vibration experimental data in the fault state in the training set to the sample generation model to obtain generated samples, and construct a balanced training set with the vibration experimental data in the healthy state in the training set; A classifier construction and training module is configured to construct a fault diagnosis classifier; the fault diagnosis classifier is trained using the balanced training set, and a trained fault diagnosis classifier is obtained when a third set condition is met. ​ ​ ​ ​ ​ A fault diagnosis module is configured to acquire diagnosis data including the test set data and real-time vibration data of the mechanical equipment to be diagnosed, and input the diagnosis data into the trained fault diagnosis classifier to obtain a fault diagnosis result.

6. A computer device comprising: A computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the mechanical intelligent fault diagnosis method for data imbalance according to any one of claims 1-4.

7. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the mechanical intelligent fault diagnosis method for data imbalance according to any one of claims 1-4.

8. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the mechanical intelligent fault diagnosis method for data imbalance according to any one of claims 1-4. The computer program is executed by the processor to implement the mechanical intelligent fault diagnosis method for data imbalance according to any one of claims 1-4.

Images