Early fault diagnosis method and system for rotating machinery under dynamic working condition

Abstract

The invention, which relates to the technical field of intelligent fault diagnosis, discloses an early fault diagnosis method and system for a rotating machine under a dynamic working condition, comprising the following steps: S1, preparing data; 2, training a DA-WGAN model and carrying out data expansion, which comprises the following steps: S21, generating an adversarial network; s22, introducing a conditional version of the GAN model, namely a C-GAN model, and defining an objective function of the C-GAN to realize input of class labels; s23, using a Wasserstein distance as a measure of a distance between real data distribution and generated data distribution, and adopting a gradient penalty method (GP) to meet a Lipschitz continuity condition, and defining an objective function # imgabs0 # of GP; S24, realizing a domain adversarial loss function through a binary cross entropy loss function, and defining DA-WGAN loss based on the domain adversarial loss function and the improved objective function # imgabs1 #, so as to obtain a DA-WGAN model; and completing establishment of a domain recognizer in the generative model and setting of domain confrontation loss. The method has the effect of improving the problem of insufficient label samples for early fault diagnosis of the rotating machinery.

Description

Technical Field

[0001] The present application relates to the technical field of intelligent fault diagnosis, and in particular to a method and system for early fault diagnosis of rotating machinery under dynamic conditions. Background Art

[0002] Rotating machinery (such as cutting tools, bearings, gearboxes, etc.) is widely used in many industries such as machinery and petrochemicals, and has a significant impact on the normal operation of the equipment. Component failures may affect the processing quality and work efficiency of the product, or may cause major mechanical accidents and casualties. Therefore, it is very important to use reliable and fast online monitoring technology to perform early fault diagnosis on rotating machinery to avoid equipment deterioration, reduced efficiency and unexpected failures.

[0003] With the rapid development of deep learning theory and methods, its monitoring principle based on data-based pattern recognition has received widespread attention in the industry and has achieved many good results. However, deep learning models require a large number of labeled samples to obtain good classification results, which is very difficult for early faults in rotating machinery because it is difficult to classify data samples under each working condition under variable working conditions, especially early fault characteristics, which poses a challenge to early fault identification under data-driven dynamic working conditions. Summary of the invention

[0004] In order to improve the problem of insufficient label samples for early fault diagnosis of rotating machinery, the present application provides a method and system for early fault diagnosis of rotating machinery under dynamic conditions.

[0005] In a first aspect, the present application provides a method for early fault diagnosis of rotating machinery under dynamic conditions, which adopts the following technical solution:

[0006] A method for early fault diagnosis of rotating machinery under dynamic conditions, comprising:

[0007] S1. Data preparation, which includes:

[0008] S11, obtaining early fault data; wherein the early fault data includes labeled source domain data and unlabeled target domain data;

[0009] S12, dividing the early fault data into a training set and a validation set, wherein the training set is used for model training and the validation set is used for model evaluation;

[0010] S2: Training the DA-WGAN model and performing data expansion, which includes:

[0011] S21, Generative Adversarial Network;

[0012] S22. Introduce the conditional version of the GAN model: the C-GAN model, and define the objective function of the C-GAN to implement the input of class labels;

[0013] S23, use Wasserstein distance as the measure of the distance between the real data distribution and the generated data distribution, and use the gradient penalty method (GP) to meet the Lipschitz continuity condition, and define the objective function of GP

[0014] S24. Implement the domain adversarial loss function through the binary cross entropy loss function, and based on the domain adversarial loss function and the improved objective function Define DA-WGAN loss, complete the construction of domain identifier and the setting of domain adversarial loss in the generative model;

[0015] S3. Training a multi-scale transfer learning classification model based on the attention mechanism, which includes: using the labeled data similar to the target working condition generated by the DA-WGAN model as the source domain, using the unlabeled data of the target working condition as the target domain, establishing a multi-scale ResNet under the attention mechanism to extract domain-invariant features, and using the transfer learning method of MK-MMD for training;

[0016] S4, application, which includes: applying the model established based on S1-S3, and continuously obtaining data in the application to update the model.

[0017] Optionally, the generative adversarial network comprises:

[0018] The DA-WGAN model is defined to include a generator G, a discriminator D, and a domain adaptation module;

[0019] The generator G maps the noise source z to the input space, and the discriminator D distinguishes whether the sample originates from the real data distribution or the generated data distribution, and the process is expressed as:

[0020]

[0021] in, is the actual data distribution, The generated data distribution is implicitly defined by x=G(z), z~p(z), where z is sampled from a simple noise distribution p;

[0022] Define the loss function JS of the standard GAN,

[0023] Where KL is the Kullback-Leibler divergence and Pm is the mixture

[0024] Optionally, the objective function of the C-GAN includes:

[0025]

[0026] Among them, y represents the input of the hidden layer combined with the prior noise, which is the class label.

[0027] Optionally, the formula corresponding to the Wasserstein distance includes:

[0028]

[0029] in, Represents the actual data distribution and generate data distribution The entire set of feasible joint distributions γ(x,y) of To distribute Convert to Distribution The minimum cost of transmission quality.

[0030] Optionally, the objective function It includes:

[0031] In the discriminator D, and Concatenate y with the joint hidden representation. In the generator G, concatenate y with P(z) in the same representation.

[0032] The objective function between the generator G and the discriminator D is a minimax game equation and is:

[0033]

[0034] The gradient penalty method is used to satisfy the Lipschitz continuity condition and obtain a new objective function:

[0035] Where λ is the gradient penalty coefficient, is along the real data distribution and generate data distribution Linear sampling between ;

[0036] The domain adversarial loss function includes:

[0037]

[0038] in is the domain loss function, D domain represents the predicted probability of the domain identifier that the sample belongs to the target domain, d labelsis the domain label, usually represented by 0 and 1, 0 represents the source domain, 1 represents the target domain, and N is the number of samples;

[0039] The DA-WGAN loss is defined as:

[0040] Optionally, the training of the multi-scale transfer learning classification model based on the attention mechanism includes establishing an unsupervised domain adaptation framework; the establishing of the unsupervised domain adaptation framework includes:

[0041] Unsupervised domain adaptation is selected as the supervised learning algorithm, and the source domain is defined as:

[0042]

[0043] in represents the source domain, represents the i-th sample in the source domain, X s is the set of all samples, is the i-th label of the i-th sample, Y s is the set of all labels, n s Represents the total number of source domain samples. Assuming that the labels in the target domain are unavailable, the data in the target domain is defined as follows:

[0044]

[0045] in, represents the target domain, is the i-th sample in the target domain, X t is the set of all samples, n t The target domain and source domain data respectively obey the probability distribution of P and Q, which represents the total number of target samples. A deep neural network is established to classify the unlabeled data in the target domain by learning the features that can be transferred, which is expressed as: Among them, β(·) represents a deep neural network, is the prediction result of the model;

[0046] The purpose of domain adaptation is to minimize the target domain risk ε using source domain data supervision t (β), and: ε t (β) = Pr (x,y)～Q [β(x)≠y];

[0047] The total loss for domain adaptation is defined as: Among them, λ is the trade-off parameter; Represents partial loss to reduce the feature difference between source domain and target domain data; is the maximum cross entropy loss, which is used to measure the difference between the prediction and the true label of the source domain data, and:

[0048] Where C is the number of all possible labels, is the indicator function.

[0049] Optionally, the training of the multi-scale transfer learning classification model based on the attention mechanism includes building a feature extractor, and the building of the feature extractor includes:

[0050] An AMResNet module consisting of three residual modules in parallel is established. The AMResNet module uses convolutional layers with kernel sizes of 1×1, 1×3, and 1×9 to process input data in parallel with three branches to extract feature information of different scales. At the end of the AMResNet module, the features extracted by each branch are used as the input of the scale-attention mechanism to obtain weights of different scales. The weighted features of each branch are connected according to the dimension of the channel to realize the construction of the feature extractor.

[0051] Optionally, the training of the multi-scale transfer learning classification model based on the attention mechanism includes:

[0052] Add MK-MMD to the loss function to design a transfer learning strategy, realize feature conversion between source domain and target data, and define the final loss function as:

[0053] Among them, λ MK-MMD is a trade-off parameter in this total loss, and Indicates the multi-core version of MMD.

[0054] Optionally, the training of the multi-scale transfer learning classification model based on the attention mechanism includes building a class discriminator, and the building of the class discriminator includes: inputting the high-dimensional feature vector that has undergone domain adaptive alignment into the class classifier for classification, and the class discriminator is built by two fully connected layers.

[0055] In a second aspect, the present application provides a system for early fault diagnosis of rotating machinery under dynamic conditions, which adopts the following technical solution:

[0056] A system for early fault diagnosis of rotating machinery under dynamic working conditions comprises a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and execute any of the above-mentioned methods for early fault diagnosis of rotating machinery under dynamic working conditions.

[0057] In summary, this application includes the following beneficial technical effects:

[0058] 1) The present invention fundamentally solves the constraint that a certain amount of labeled data of the same working condition is required for data expansion in published data generation methods (such as standard GANs and various GANs variants); the DA-WGAN proposed in this application only needs labeled data of one working condition to generate labeled data of multiple different working conditions, which is more in line with the actual situation that label data of a variable processing environment is difficult to collect;

[0059] 2) In the case of fault diagnosis under variable working conditions, the prior art proposes some methods based on transfer learning, and the premise for the implementation of these methods is: there is enough source domain label data to train the model; this application is aimed at early fault diagnosis under variable working conditions where there is not enough label data under any working condition; the application considers not only the complex and variable working conditions, but also the coupling situation that the early fault signal is weak and more difficult to collect;

[0060] 3) This application uses a gradient-penalized Wasserstein generative adversarial network (WGAN-GP) for data enhancement, and introduces domain adaptation technology into the model; a small amount of labeled data is used to introduce inter-class features using a discriminator, and a large amount of unlabeled data of various working conditions is used to introduce variable working condition features through domain adaptation, thereby guiding the generation of a large amount of labeled data under different working conditions; compared with traditional generative adversarial networks, this method is more suitable for processing complex data distribution under different working conditions in industry, generates higher quality data, and expands the application of data enhancement technology in early fault diagnosis;

[0061] 4) In the process of aligning generated data with target domain data, a multi-scale transfer learning model with attention mechanism is introduced; by focusing on key feature information through the attention mechanism, the model can align the generated data and target domain data at different scales, significantly improving the model's ability to recognize target domain data;

[0062] 5) It combines the advantages of generative adversarial networks and transfer learning, breaking through the dependence of traditional fault diagnosis models on labeled data; by introducing unlabeled data during the training process and performing domain adaptive adjustments, the model can automatically learn the feature migration relationship between different working conditions, and realize the efficient use of unlabeled data; this setting solves the problems of imbalance and labeling difficulty of fault data under different working conditions, and also significantly improves the adaptability and classification accuracy of the model under dynamic conditions, showing good prospects for industrial applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a schematic diagram of the main process of the method of the present application;

[0064] Figure 2 It is the overall structure diagram of DA-WGAN of this application;

[0065] Figure 3 This is the overall structure diagram of DA-WGAN;

[0066] Figure 4 is a schematic diagram of the domain adaptation rule of this application;

[0067] Figure 5 It is a schematic diagram of the shortcut connection of the residual block;

[0068] Figure 6 It is a schematic diagram of the MK-MMD domain adaptive classification framework;

[0069] Figure 7 It is a schematic diagram of the bearing early fault diagnosis test bench;

[0070] Figure 8 It is the spectrum analysis of the generated data and actual data of this application;

[0071] Fig. 9 Schematic diagram of the confusion matrix of the method of the present application in four types of migration tasks. DETAILED DESCRIPTION

[0072] The following is combined with Figure 1-9 This application is described in further detail.

[0073] The present application embodiment discloses a method for early fault diagnosis of rotating machinery under dynamic conditions, which:

[0074] It is used to solve the problem of reduced classification performance caused by insufficient label data and uneven distribution of fault data in early fault identification under dynamic working conditions;

[0075] In the condition monitoring and early fault diagnosis of industrial equipment, traditional supervised classification algorithms rely on a large amount of labeled fault data; however, when operating conditions change frequently, it is very difficult to obtain sufficient and balanced labeled data covering all operating conditions. In particular, the characteristics of early faults are weak, and it is difficult to accurately label fault data samples under different operating conditions, resulting in unbalanced data distribution and insufficient samples, affecting the accuracy and reliability of fault detection.

[0076] This method uses labeled data from the source domain (known working conditions) to train a gradient-penalized WassersteinGAN (WGAN), and introduces unlabeled data from other working conditions as the target domain during the training process, prompting the generative adversarial network to generate labeled data close to the target domain working conditions, thereby solving the problem of traditional methods' dependence on data labeling.

[0077] In addition, in order to further improve the practical application value of generated data, this method takes into account that it is difficult for generated data to fully simulate the actual data feature distribution of the target domain, and proposes a multi-scale transfer learning model based on the attention mechanism to align the feature distribution of generated data with that of the target domain data; this model effectively captures and corrects the feature differences between different data domains by introducing the attention mechanism and the multi-scale feature alignment strategy, thereby improving the representation ability of generated data and the classification performance of target domain data.

[0078] Reference Figure 1 and Figure 2 , the early fault diagnosis methods of rotating machinery under dynamic conditions include:

[0079] S1. Data preparation, which includes:

[0080] S11, obtaining early fault data; wherein the early fault data includes labeled source domain data (referred to as: A working condition) and unlabeled target domain data (referred to as: other working conditions, B working condition).

[0081] In the early stages of a fault, the signal strength may be low and it is easily masked by environmental noise or other signals. This is the key challenge in early fault diagnosis research. The weakness and confusion of features require diagnostic technology to have higher sensitivity and noise resistance, as well as stronger feature differentiation capabilities. To this end, this method explores the generation of representative early fault samples under the framework of the generative data model GANs to expand the period label data of different working conditions, and combines the model's feature learning ability to improve the recognition accuracy of early fault signals.

[0082] The specific method of obtaining relevant data is explained in another application embodiment of the present application, so it will not be repeated here.

[0083] S12. Divide the early fault data into a training set and a validation set, and the training set is used for model training, and the validation set is used for model evaluation.

[0084] In this embodiment, the sliding window length of the model is 1024, 80% of the total samples are used as training sets, 20% are used as test sets, and it is ensured that there is no overlap between the training data and the test data to avoid data leakage.

[0085] S2. Training the DA-WGAN (Domain-Adaptive Wasserstein GAN) model and performing data expansion, which includes:

[0086] S21. Generate adversarial networks, specifically:

[0087] The DA-WGAN model consists of a generator G, a discriminator D, and a domain adaptation module ( Figure 2 The domain adaptation module includes the domain identifier. The GAN model training process involves defining a game between two adversarial networks:

[0088] Reference Figure 3 , the generator (network) G maps the noise source to the input space, and the discriminator (network) D distinguishes whether the sample comes from the real data distribution or the generated data distribution; the training process is a minimax two-player game, and the optimization process is expressed as:

[0089]

[0090] in, is the actual data distribution, The generated data distribution is implicitly defined by x=G(z), z~p(z), where z is sampled from a simple noise distribution p, such as a uniform, normal, or Gaussian distribution. The discriminator D is optimized to maximize the probability that the training sample, the sample from the generator G, is correctly classified, and the generator G is optimized to minimize or

[0091] In the related art, Jensen-Shannon (JS) divergence is used as the loss function of the standard GAN as a measure of the generated samples; it is defined as: Where KL is the Kullback-Leibler divergence and Pm is the mixture

[0092] S22. Introduce the conditional version of the GAN model: the C-GAN model, and define the objective function of C-GAN to realize the input of class labels.

[0093] In this embodiment, the training of GAN is performed using unsupervised learning with randomly generated data. Bearing fault diagnosis often has multiple fault types, and deployment is very complicated when there are multiple classes to learn. To overcome this problem, this paper introduces a conditional version of GAN, namely C-GAN, which uses additional label information to generate supervised data to simplify the training procedure; therefore, as long as the corresponding label is input, the required class can be generated from the trained generator.

[0094] The objective function of C-GAN includes:

[0095]

[0096] Where y is the class label or any other type of auxiliary information. y is combined with the prior noise as the input to the hidden layer.

[0097] S23, use Wasserstein distance as the measure of the distance between the real data distribution and the generated data distribution, and use the gradient penalty method (GP) to meet the Lipschitz continuity condition, and define the objective function of GP Specifically:

[0098] Both C-GAN and standard GAN exhibit the same problematic behaviors: gradient collapse and unstable training (gradient disappearance) and other problems. Therefore, this method uses Wasserstein distance (also known as Earth-Mover) instead of JS divergence as a measure of the distance between the real data distribution and the generated data distribution, because the smoothness of Wasserstein distance is better than JS divergence.

[0099] On the one hand, Wasserstein distance can theoretically solve the problem of gradient vanishing in GAN and C-GAN; on the other hand, experiments show that using Wasserstein distance instead of JS divergence can effectively solve the mode collapse problem.

[0100] The formulas corresponding to the Wasserstein distance include:

[0101] in, Represents the actual data distribution and generate data distribution The entire set of feasible joint distributions γ(x,y) of

[0102] To distribute Convert to Distribution The minimum cost of transmission quality.

[0103] In the discriminator D, and Concatenate y with the joint hidden representation. In the generator G, concatenate y with P(z) in the same representation.

[0104] Formally, the objective function between the generator G and the discriminator D is a minimax game equation and is:

[0105]

[0106] In theory, Wasserstein distance can not only solve the problem of training instability, but also provide a reliable indicator for the training progress of GAN; in addition, this indicator is highly correlated with the quality of generated samples. However, the fitting ability of deep neural networks has not been fully utilized, and the gradient will disappear or explode after weight sharing.

[0107] In order to solve the above problems, this method adopts the gradient penalty (GP) method to meet the Lipschitz continuity condition. GP solves the problem of gradient disappearance or gradient explosion in training. Considering GP, a new objective function is obtained:

[0108]

[0109] Where λ is the gradient penalty coefficient, is along the real data distribution and generate data distribution Linear sampling between ;

[0110] The domain adversarial loss function includes:

[0111]

[0112] ;in is the domain loss function, D domain represents the predicted probability of the domain identifier that the sample belongs to the target domain, d labels is the domain label, usually represented by 0 and 1, 0 represents the source domain, 1 represents the target domain, and N is the number of samples;

[0113] The DA-WGAN loss is defined as:

[0114] S24. Implement the domain adversarial loss function through the binary cross entropy loss function, and based on the domain adversarial loss function and the improved objective function Define the DA-WGAN loss and complete the construction of the domain identifier and the setting of the domain adversarial loss in the generative model. Specifically:

[0115] This method, for the first time, introduces the domain adaptation strategy into GAN. The domain identifier is a key component of DA-WGAN. Its task is to distinguish data from two different conditions (A and B), regardless of whether the data is real or generated. The purpose of the (domain identifier) ​​network is to help the generator generate data that is closer to condition B.

[0116] The domain adversarial loss function is used to measure the domain recognition D domain The prediction accuracy of the domain adversarial loss function is to enable the domain identifier to accurately distinguish whether the data is from the source domain or the target domain. In this method, the domain adversarial loss function is implemented by the binary cross entropy loss function. The domain adversarial loss function includes:

[0117] in is the domain loss function, D domain represents the predicted probability of the domain identifier that the sample belongs to the target domain, dlabels is the domain label, usually represented by 0 and 1, 0 represents the source domain, 1 represents the target domain, and N is the number of samples;

[0118] Domain-based adversarial loss function and improved objective function Define DA-WGAN loss,DA-WGAN loss is defined as:

[0119]

[0120] S3. Training a multi-scale transfer learning classification model based on the attention mechanism, which includes: using the labeled data similar to the target working condition generated by the DA-WGAN model as the source domain, using the unlabeled data of the target working condition as the target domain, establishing a multi-scale ResNet under the attention mechanism to extract domain-invariant features, and using the MK-MMD transfer learning method for training, specifically:

[0121] 1) Unsupervised Domain Adaptation Framework:

[0122] When the feature distributions of the source domain and target domain datasets are different, traditional supervised learning algorithms often have difficulty achieving effective classification. It has been proven that unsupervised domain adaptation is more suitable for the following scenarios:

[0123] Reference Figure 4 The premise of domain adaptation only requires that the data in the source domain dataset are all labeled, so the source domain can be defined as follows:

[0124] in represents the source domain, represents the i-th sample in the source domain, X s is the set of all samples, is the i-th label of the i-th sample, Y s is the set of all labels, n s Represents the total number of source domain samples. Assuming that the labels in the target domain are unavailable, the data in the target domain is defined as follows:

[0125]

[0126] in, represents the target domain, is the i-th sample in the target domain, X t is the set of all samples, n t Represents the total number of target samples. The target domain and source domain data obey the probability distribution of P and Q respectively. A deep neural network is established to classify the unlabeled data in the target domain by learning the features that can be transferred, which is expressed as: Among them, β(·) represents a deep neural network, is the prediction result of the model.

[0127] The purpose of domain adaptation is to minimize the target domain risk ε using source domain data supervision t (β), and: ε t (β) = Pr (x,y)～Q [β(x)≠y];

[0128] The total loss for domain adaptation is defined as: Among them, λ is the trade-off parameter; Represents partial loss to reduce the feature difference between source domain and target domain data; is the maximum cross entropy loss, which is used to measure the difference between the prediction and the true label of the source domain data, and:

[0129] Where C is the number of all possible labels and 1 is the indicator function.

[0130] 2) Build a feature extractor, which includes:

[0131] An AMResNet module consisting of three residual modules in parallel is established. The AMResNet module uses convolutional layers with kernel sizes of 1×1, 1×3, and 1×9 to process input data in parallel with three branches to extract feature information of different scales. At the end of the AMResNet module, the features extracted by each branch are used as the input of the scale-attention mechanism to obtain weights of different scales. The weighted features of each branch are connected according to the dimension of the channel to realize the construction of the feature extractor.

[0132] As an important part of the proposed domain-invariant feature extraction method, the Attention-Enhanced Multi-Scale ResNet Module (AMResNet) is a multi-scale ResNet module with an attention mechanism, which is also a basic component of the feature generator Gf. At the same time, its main structural features and the application of the attention mechanism are described in detail as follows:

[0133] Residual learning is a technology that can effectively avoid model degradation. It uses the fact that the learning effect of shallow networks is better than that of deep networks when the model is degraded. It directly inputs the features learned by shallow networks into deep networks across layers, thereby effectively solving the problem of reducing model degradation caused by excessive training errors and improving the overall performance of the network. Figure 5 shown.

[0134] The residual network module usually uses two or more layers of convolution operations, such as Figure 5 As shown, it mainly consists of the identity mapping part and the residual part, and the formula is as follows: y = F(x,W i)+x;

[0135] Among them, x represents the input of the current layer, and y represents the output of the current layer. Please note that this is different from the class label y in this method; F(x,W i ) is the residual function, which represents the residual mapping to be learned. Figure 5 There are two layers of convolution operations, and the residual function can be expressed as: F = W 2 σ(W 1 x), where W represents the convolution operation and σ is generally a nonlinear activation function.

[0136] Regarding weighting the importance of features of different scales extracted by each residual block, specifically:

[0137] For each residual block output feature F i , where i represents different scales (e.g. 1x1, 1x3, 1x7); A i =σ(W Ai *F i +b Ai );

[0138] A i is the attention feature after a convolution layer and bias processing, b Ai It is obtained through an activation function σ (usually Sigmoid); the convolutional layer learns how to extract useful information from the input feature vector to calculate the importance of each element.

[0139] Then, use this attention weight to weight the original features to get the weighted feature vector:

[0140] Here, ⊙ represents element-wise multiplication.

[0141] Finally, the weighted features of all scales are fused to obtain the final feature representation F:

[0142] 3) Multi-Kernel Maximum Mean Discrepancy (MK-MMD), which includes:

[0143] Add MK-MMD to the loss function to design a transfer learning strategy, realize feature conversion between source domain and target data, and define the final loss function as:

[0144] Among them, λ MK-MMD is a trade-off parameter in this total loss, and represents the multi-core version of MMD. In this embodiment, only Gaussian kernel is used, and the number of kernels is equal to 5.

[0145] Specifically:

[0146] MMD was first proposed by Professor Long Mingsheng of Tsinghua University in 2015 and has since been used by many other researchers for transfer learning. MMD defined in kernel Hilbert space (RKHS) is the squared distance between the kernel embeddings of the marginal distributions P(Xs) and Q(Xt). The MMD formula is as follows:

[0147]

[0148] in represents the RKHS using kernel k (generally a Gaussian kernel is used as the kernel), φ(·) represents the mapping to the RKHS, represents mathematical expectation.

[0149] For practical applications, the parameter selection of each kernel is crucial to the final performance of the mapping. To address this problem, this method uses MK-MMD, which can maximize the testing power of two samples and minimize the type II error at the same time. For MK-MMD, this method uses a convex combination of m kernels to provide an effective estimate of the mapping:

[0150]

[0151] Among them, {β u} are weight parameters for different kernels.

[0152] A transfer learning strategy is designed by adding MK-MMD to the loss function to achieve Figure 6 Feature transformation between source and target domains as shown in .

[0153] 4) Building a classifier, which includes: inputting the high-dimensional feature vector that has undergone domain adaptive alignment into the class classifier for classification, and the classifier is built with two fully connected layers.

[0154] S4, application, which includes: applying the model established based on S1-S3 to the status monitoring and fault diagnosis scenarios of industrial equipment, and continuously obtaining data in the application to update the model.

[0155] In another embodiment of the method, an application example is given: taking a bearing as an example of a rotating machinery, an early fault diagnosis experiment of a bearing under variable working conditions is performed:

[0156] (1) Build a bearing early fault diagnosis test bench to collect data. Figure 7As shown, the bearing early fault diagnosis test bench includes a drive motor, a dynamic torque sensor, and an eddy current brake; the transmission shaft and the balance disk on the transmission shaft are coaxially fixed between the output shaft of the drive motor and the eddy current brake, and a bearing working position is selected and arranged, thereby forming an experimental load to form a complete fault simulation system.

[0157] (2) Definition of fault categories: This method considers six types of bearing health conditions, including one normal condition, three single faults and two compound faults. The specific definitions are shown in Table 1. Single faults include inner ring depression, outer ring depression and misalignment; compound faults are the superposition of inner ring depression and misalignment, and outer ring depression and misalignment. In order to simulate early faults, this method uses pitting to create depression damage on the inner and outer rings. The fault size is defined as 0.2 mm in diameter and 0.1 mm in depth. Misalignment faults are achieved by adding 0.2 mm gaskets to the bearing segments.

[0158] Table 1. Six types of bearing health status

[0159]

[0160] This experimental design aims to cover the most common failure modes and compound failure conditions in rolling bearings, fully simulating the operating conditions under actual working conditions. Common bearing failures are mainly concentrated on damage to the inner and outer rings, as well as misaligned installation conditions, which have been verified and adopted by a large number of studies and classic data sets (such as the CWRU data set). On this basis, this method further introduces compound failures (such as inner or outer ring depressions and misalignment superposition), expanding the richness and complexity of the experiment to study the signal interaction effects under multiple fault modes. This design not only covers typical single fault types in bearings, but also systematically explores the characteristics of compound failures.

[0161] (3) Working condition setting: Three different speeds (2000rpm, 3000rpm and 2000rpm) and two radial loads (500N and 1000N) were considered in the experiment. A total of four working condition data were collected for experimental verification, as shown in Table 2. For each fault at a specific speed and torsional load, the sampling frequency was 25600Hz, and a set of data was collected every 8s. 204800 data points were collected each time, and a total of 5 sets were collected. 1024000 data points were collected for each fault under each working condition.

[0162] Table 2. Experimental parameters

[0163]

[0164] (4) Migration task division: In order to test the effectiveness of the method proposed in this invention, four migration tasks are designed, namely task (D1→D2), task (D2→D1), task (D3→D4), and task (D4→D3). The left and right sides of “→” represent the source domain dataset and the target domain dataset, respectively. For example, in the migration task (D1→D2), the data generation model is first used to expand the labeled data of the source domain D1 working condition, and then a large amount of labeled data is generated as the source domain data. The classification model is trained using the MK-MMD classification framework. The target domain D2 working condition data is not labeled, but the target domain data is also involved in the training.

[0165] (5) In order to reproduce the proposed method, this method attempts to tune the layer configuration and parameter settings of DA-WGAN.

[0166] This embodiment introduces two upsampling layers in the generator G, which gradually increase the dimension of the feature space so that the generated data has sufficient resolution and details. In addition, the last layer of the generator G uses the Tanh activation function to limit the output range to [-1, 1] to keep consistent with the range after data normalization.

[0167] The discriminator D consists of a multi-layer convolutional structure, which shrinks the feature map layer by layer and finally outputs the true and false discrimination results through the linear layer. The discriminator D also uses Leaky ReLU activation to alleviate the gradient vanishing problem. More convolutional layers (e.g., six layers in total) are added to the discriminator D to enable it to capture more detailed distribution differences between the source domain and the target domain data. The last layer uses Sigmoid activation to output the probability of domain discrimination in the range of [0,1]. Batch normalization (BN) operations are used in the discriminator and domain discriminator to stabilize the training process and ensure that the network has consistent distribution characteristics in different input batches.

[0168] (6) Generate fault data: After generating the noise signal, it is crucial to evaluate and filter its quality. First, feature extraction and screening are performed on each segment of the signal through time domain, frequency domain, and time-frequency domain analysis to ensure the quality of the generated samples. Taking the D1→D2 transfer task as an example, the labeled data of condition D1 and the unlabeled data of condition D2 are used to train DA-WGAN to generate labeled data for condition D2 that is similar to the features of condition D2.

[0169] The time domain, frequency domain, and time-frequency domain representations generated by DA-WGAN are as follows Figure 8 As shown in the figure, the generated data is very similar to the original D2 data in many aspects: in the time domain, the waveform of the generated signal is similar to that of the original signal, showing similar amplitude and frequency characteristics; in the frequency domain, the spectral structure of the generated signal is very consistent with the main frequency component of the original signal; in the time-frequency domain, the time-frequency distribution of the generated signal is consistent with the original data.

[0170] (7) Results comparison:

[0171] This method is compared with four other typical GAN ​​methods (GAN, WGAN, WGANGP and SNGAN): 512,000 data points from the source domain are used to train the generative model, and the generated data is used to classify and identify the target working conditions. 512,000 (500 samples per class), 512,000×2 (1,000 samples per class), and 512,000×3 (1,500 samples per class) are generated respectively.

[0172] Tables 3 to 6 show the test results of different data enhancement methods.

[0173] Table 3. Diagnosis results of migration task D1→D2 (%)

[0174]

[0175] Table 4. Diagnosis results of migration task D2→D1 (%)

[0176]

[0177]

[0178] Table 5. Diagnosis results of migration task D3→D4 (%)

[0179]

[0180] Table 6. Diagnosis results of migration task D4→D3 (%)

[0181]

[0182] From the table above, we can see that under different migration tasks (D1→D2, D2→D1, D3→D4, D4→D3), the fault classification accuracy of the five generative adversarial network variants (GAN, WGAN, SNGAN, WGANGP, DA-WGAN) shows obvious differences when the number of generated sample data is different. In general, the diagnostic accuracy of all methods increases significantly with the increase of the amount of generated training data, especially when the amount of generated data increases from 512000 to 512000×3, the diagnostic accuracy of each method is greatly improved, which shows that increasing the amount of data has a positive effect on the transfer learning effect of the model.

[0183] Reference Fig. 9, which is the confusion matrix result, the classification performance of this method in four migration tasks (D1→D2, D2→D1, D3→D4, D4→D3) shows significant fault diagnosis ability. In all migration tasks, most samples of fault categories are concentrated on the diagonal of the confusion matrix, indicating that the model has a high accuracy in the classification of healthy state (category 0) and combined fault state (such as inner race fault + misalignment fault, outer race fault + misalignment fault).

[0184] The embodiment of the present application also discloses a system for early fault diagnosis of rotating machinery under dynamic conditions.

[0185] A system for diagnosing early faults of rotating machinery under dynamic conditions, wherein a memory stores a computer program that can be loaded by a processor and execute any of the above-mentioned methods for diagnosing early faults of rotating machinery under dynamic conditions.

[0186] The above are all preferred embodiments of the present application, and the protection scope of the present application is not limited thereto. Therefore, any equivalent changes made according to the structure, shape, and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for early fault diagnosis of rotating machinery under dynamic conditions, characterized in that: include: S1. Data preparation, which includes: S11, obtaining early fault data; wherein the early fault data includes labeled source domain data and unlabeled target domain data; S12, dividing the early fault data into a training set and a validation set, wherein the training set is used for model training and the validation set is used for model evaluation; S2: Training the DA-WGAN model and performing data expansion, which includes: S21, Generative Adversarial Network; S22. Introduce the conditional version of the GAN model: the C-GAN model, and define the objective function of the C-GAN to implement the input of class labels; S23, use Wasserstein distance as the measure of the distance between the real data distribution and the generated data distribution, and use the gradient penalty method (GP) to meet the Lipschitz continuity condition, and define the objective function of GP S24. Implement the domain adversarial loss function through the binary cross entropy loss function, and based on the domain adversarial loss function and the improved objective function Define DA-WGAN loss, complete the construction of domain identifier and the setting of domain adversarial loss in the generative model; S3. Training a multi-scale transfer learning classification model based on the attention mechanism, which includes: using the labeled data similar to the target working condition generated by the DA-WGAN model as the source domain, using the unlabeled data of the target working condition as the target domain, establishing a multi-scale ResNet under the attention mechanism to extract domain-invariant features, and using the transfer learning method of MK-MMD for training; S4, application, which includes: applying the model established based on S1-S3, and continuously obtaining data in the application to update the model.

2. The method for early fault diagnosis of rotating machinery under dynamic conditions according to claim 1, characterized in that: The generative adversarial network comprises: The DA-WGAN model is defined to include a generator G, a discriminator D, and a domain adaptation module; The generator G maps the noise source z to the input space, and the discriminator D distinguishes whether the sample originates from the real data distribution or the generated data distribution, and the process is expressed as: in, is the actual data distribution, The generated data distribution is implicitly defined by x=G(z), z~p(z), where z is sampled from a simple noise distribution p; Define the loss function JS of the standard GAN, Where KL is the Kullback-Leibler divergence and Pm is the mixture 3. The method for early fault diagnosis of rotating machinery under dynamic conditions according to claim 1, characterized in that: The objective function of the C-GAN includes: Here, y is the class label or any other type of auxiliary information, and y is combined with the prior noise as the input to the hidden layer.

4. The method for early fault diagnosis of rotating machinery under dynamic conditions according to claim 1 is characterized in that: The formula corresponding to the Wasserstein distance includes: in, Represents the actual data distribution and generate data distribution The entire set of feasible joint distributions γ(x,y) of To distribute Convert to Distribution The minimum cost of transmission quality.

5. The method for early fault diagnosis of rotating machinery under dynamic conditions according to claim 1, characterized in that: The objective function It includes: In the discriminator D, and Concatenate y with the joint hidden representation. In the generator G, concatenate y with P(z) in the same representation. The objective function between the generator G and the discriminator D is a minimax game equation and is: The gradient penalty method is used to satisfy the Lipschitz continuity condition and obtain a new objective function: Where λ is the gradient penalty coefficient, is along the real data distribution and generate data distribution Linear sampling between ; The domain adversarial loss function includes: in is the domain loss function, D domain represents the predicted probability of the domain identifier that the sample belongs to the target domain, d labels is the domain label, usually represented by 0 and 1, 0 represents the source domain, 1 represents the target domain, and N is the number of samples; The DA-WGAN loss is defined as:

6. The method for early fault diagnosis of rotating machinery under dynamic conditions according to claim 1, characterized in that: The training of the multi-scale transfer learning classification model based on the attention mechanism includes establishing an unsupervised domain adaptation framework; The unsupervised domain adaptation framework is established, which includes: Unsupervised domain adaptation is selected as the supervised learning algorithm, and the source domain is defined as: in represents the source domain, represents the i-th sample in the source domain, X s is the set of all samples, is the i-th label of the i-th sample, Y s is the set of all labels, n s Represents the total number of source domain samples. Assuming that the labels in the target domain are unavailable, the data in the target domain is defined as follows: in, represents the target domain, is the i-th sample in the target domain, X t is the set of all samples, n t Indicates the total number of target samples; The target domain and source domain data obey the probability distribution of P and Q respectively. A deep neural network is established to classify the unlabeled data in the target domain by learning the features that can be transferred, which is expressed as: Among them, β(·) represents a deep neural network, is the prediction result of the model; The purpose of domain adaptation is to minimize the target domain risk ε using source domain data supervision t (β), and: ε t (β) = Pr (x,y)～Q [β(x)≠y]; The total loss for domain adaptation is defined as: Among them, λ is the trade-off parameter; Represents partial loss to reduce the feature difference between source domain and target domain data; is the maximum cross entropy loss, which is used to measure the difference between the prediction and the true label of the source domain data, and: Where C is the number of all possible labels, is the indicator function.

7. The method for early fault diagnosis of rotating machinery under dynamic conditions according to claim 1, characterized in that: The training of the multi-scale transfer learning classification model based on the attention mechanism includes building a feature extractor, and the building of the feature extractor includes: An AMResNet module consisting of three residual modules in parallel is established. The AMResNet module uses convolutional layers with kernel sizes of 1×1, 1×3, and 1×9 to process input data in parallel with three branches to extract feature information of different scales. At the end of the AMResNet module, the features extracted by each branch are used as the input of the scale-attention mechanism to obtain weights of different scales. The weighted features of each branch are connected according to the dimension of the channel to realize the construction of the feature extractor.

8. The method for early fault diagnosis of rotating machinery under dynamic conditions according to claim 1, characterized in that: The training of the multi-scale transfer learning classification model based on the attention mechanism includes: Add MK-MMD to the loss function to design a transfer learning strategy, realize feature conversion between source domain and target data, and define the final loss function as: Among them, λ MK-MMD is a trade-off parameter in this total loss, and Indicates the multi-core version of MMD.

9. The method for early fault diagnosis of rotating machinery under dynamic conditions according to claim 1, characterized in that: The training of the multi-scale transfer learning classification model based on the attention mechanism includes building a class discriminator, and the building of the class discriminator includes: inputting the high-dimensional feature vector that has undergone domain adaptive alignment into the class classifier for classification, and the class discriminator is built by two fully connected layers.

10. A system for early fault diagnosis of rotating machinery under dynamic conditions, comprising a memory and a processor, characterized in that: The memory stores a computer program that can be loaded by the processor and execute the method for early fault diagnosis of rotating machinery under dynamic conditions as described in any one of claims 1-9.

Images