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

Abstract

The application discloses a kind of dynamic working condition under rotating machinery early fault diagnosis method and system, it is related to intelligent fault diagnosis technical field, it includes: S1, data preparation;2, training DA-WGAN model and carrying out data expansion, it includes: S21, generating adversarial network;S22, introduce the conditional version of GAN model: C-GAN model, and define the objective function of C-GAN, realize the input of class label;S23, use Wasserstein distance as the distance between the measurement of real data distribution and generated data distribution, and satisfy Lipschitz continuity condition using gradient penalty method (GP), define the objective function of GP S24, realize domain adversarial loss function by binary cross-entropy loss function, and define DA-WGAN loss based on domain adversarial loss function and improved objective function, complete the construction of domain recognizer in generation model and the setting of domain adversarial loss.The present application has the effect of improving the problem of insufficient rotating machinery early fault diagnosis label sample.

Description

TECHNICAL FIELD

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

[0002] Rotating machinery (such as cutting tools, bearings, gearboxes, etc.) is widely used in many industries such as machinery and petrochemical industry, and has an important influence on the normal operation of equipment. Component failure may affect product processing quality and work efficiency, or even cause major mechanical accidents and casualties. Therefore, it is very important to use reliable and fast online monitoring technology to diagnose early faults of rotating machinery to avoid equipment deterioration, efficiency reduction and unexpected failures.

[0003] With the rapid development of deep learning theory and method, its monitoring principle based on data pattern recognition has received widespread attention from the industry and has achieved many good results. However, deep learning models require a large number of labeled samples to achieve good classification results, which is very difficult for early fault diagnosis of rotating machinery. Under variable working conditions, it is difficult to label data samples under each working condition, especially early fault features, which poses a challenge to data-driven early fault recognition under dynamic working conditions. SUMMARY

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

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

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

[0007] S1, data preparation, comprising:

[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, and the training set is used for model training and the validation set is used for model evaluation;

[0010] S2, training a DA-WGAN model and expanding data, comprising:

[0011] S21, generating a generative adversarial network;

[0012] S22, introduce the conditional version of the GAN model: C-GAN model, and define the objective function of C-GAN, realize the input of class label;

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

[0014] S24, realize the domain adversarial loss function through binary cross entropy loss function, and define DA-WGAN loss based on domain adversarial loss function and improved objective function

[0015] S3, train the multi-scale transfer learning classification model based on attention mechanism, which includes: taking the label data similar to the target working condition generated by the DA-WGAN model as the source domain, taking 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, including: applying the model established based on S1-S3, and continuously updating the model in the application.

[0017] Optionally, the generative adversarial network comprises:

[0018] The DA-WGAN model comprises a generator G, a discriminator D and a domain adaptive module;

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

[0020]

[0021] wherein, is the real data distribution, is the generated data distribution defined implicitly by x=G(z), z~p(z), and z is sampled from a simple noise distribution p;

[0022] The loss function JS of the standard GAN is defined as

[0023] wherein, KL is Kullback-Leibler divergence, Pm is a mixture

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

[0025]

[0026] where y represents the input combined with the prior noise as the hidden layer, and y represents the class label.

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

[0028]

[0029] where, represents the real data distribution and the generated data distribution the entire feasible joint distribution γ(x, y) set, is the minimum cost of the transmission quality of converting the distribution to the distribution

[0030] Optionally, the objective function includes:

[0031] In the discriminator D, y is connected with y in the joint hidden representation; in the generator G, y is connected with P(z) in the same representation form;

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

[0033]

[0034] The Lipschitz continuity condition is satisfied by the gradient penalty method, and a new objective function is obtained:

[0035] where λ is the gradient penalty coefficient, is sampled along the straight line between the real data distribution and the generated data distribution

[0036] The domain adversarial loss function includes:

[0037]

[0038] where is the domain loss function, D domain represents the predicted probability of the domain identifier that the sample belongs to the target domain, and 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;

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

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

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

[0042]

[0043] wherein 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 not available, the data in the target domain is defined as follows:

[0044]

[0045] wherein, 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 the source domain data respectively obey the probability distribution P and Q, a deep neural network is established and can be classified by learning the transferable features of the unlabeled data in the target domain, which is represented as: wherein, β(·) represents a deep neural network, is the prediction result of the model;

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

[0047] The total loss of domain adaptation is defined as: wherein, λ is a weighting parameter; represents a partial loss, which reduces the feature difference between the source domain and the target domain data; is the maximum cross-entropy loss, which is used to measure the difference distance between the prediction of the source domain data and the real label, 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 comprises building a feature extractor, the building of the feature extractor comprising:

[0050] An AMResNet module in parallel with three residual modules is established, the AMResNet module using convolution layers with kernel sizes of 1x1, 1x3 and 1x9 to process input data in three branches in parallel to extract feature information of different scales, and at the end of the AMResNet module, the features extracted by each branch are taken as the scale input-attention mechanism to obtain the weights of different scales, and the weighted features of each branch are connected according to the dimension of the channel to realize the building of the feature extractor.

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

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

[0053] where λ MK-MMD is a trade-off parameter in the total loss, and represents a multi-kernel version of MMD.

[0054] Optionally, the training of the multi-scale transfer learning classification model based on the attention mechanism comprises building a class discriminator, the building of the class discriminator comprising: inputting the high-dimensional feature vector aligned by the domain adaptation into a class classifier for classification, and the class discriminator being built by two fully connected layers.

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

[0056] A rotating machinery early fault diagnosis system under dynamic working conditions, comprising a memory and a processor, the memory storing a computer program capable of being loaded and executed by the processor to perform the rotating machinery early fault diagnosis method under dynamic working conditions according to any one of the above.

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

[0058] 1), The application fundamentally solves the constraint that a certain amount of same-condition label data is required for the premise of realizing data expansion by the published data generation method (such as standard GANs and various GANs); the DA-WGAN proposed in the application can generate label data under multiple different conditions only by using label data under one condition, which is more in line with the actual situation that label data under a complex and variable processing environment is difficult to collect;

[0059] 2), In the fault diagnosis under variable conditions, the prior art proposes some methods based on transfer learning, and the premise that can be realized by these methods is that there are enough source domain label data to train the model; the application is aimed at early fault diagnosis under variable conditions without enough label data under any condition; the application considers not only the complex and variable condition environment, but also the coupling situation that early fault signals are weak and more difficult to collect;

[0060] 3), The application uses the gradient-penalized Wasserstein generative adversarial network (WGAN-GP) for data enhancement, and introduces the domain adaptation technology into the model; a small amount of label data is used to introduce inter-class features by using a discriminator, and a large amount of unlabelled multiple different condition data is used to introduce variable condition features by using domain adaptation, so as to guide the generation of a large amount of label data under different conditions; compared with the traditional generative adversarial network, the method is more suitable for processing complex data distribution under different conditions in the industry, the generated data has higher quality, and the application of the data enhancement technology in early fault diagnosis is expanded;

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

[0062] 5), The advantages of the generative adversarial network and the transfer learning are combined, and the dependence of the traditional fault diagnosis model on the label data is broken; by introducing unlabelled data and performing domain adaptation adjustment in the training process, the model can automatically learn the feature transfer relationship between different conditions, and realizes efficient utilization of the unlabelled data; this setting solves the problems of unbalanced fault data and difficult marking under different conditions, and significantly improves the adaptability and classification precision of the model under dynamic conditions, which shows good industrial application prospect. BRIEF DESCRIPTION OF DRAWINGS

[0063] Figure 1 is a main flow schematic diagram of the method of the application;

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

[0065] Figure 3 is a DA-WGAN overall structure diagram;

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

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

[0068] Figure 6 is an MK-MMD domain adaptation classification framework schematic diagram;

[0069] Figure 7 is a schematic diagram of a bearing early fault diagnosis experiment table;

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

[0071] Figure 9 is a confusion matrix schematic diagram of the method of the present application in four types of transfer tasks. DETAILED DESCRIPTION

[0072] The following will be described in detail below in combination with the accompanying drawings Figures 1-9 The present application will be further described in detail.

[0073] The embodiment of the present application discloses a rotating machine early fault diagnosis method under dynamic working conditions, which comprises the following steps:

[0074] It is used to solve the problem of classification performance decline caused by insufficient labeled data and unbalanced fault data distribution in early fault recognition under dynamic working conditions;

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

[0076] The method trains a gradient-penalized WassersteinGAN (WGAN) based on the labeled data of the source domain (known working conditions), and introduces the unlabeled data of other working conditions as the target domain in the training process, prompting the generative adversarial network to generate labeled data close to the target domain working conditions, solving the dependence of traditional methods on data labeling.

[0077] In addition, in order to further improve the practical application value of the generated data, considering that it is difficult for the generated data to completely simulate the actual data feature distribution of the target field, the method further proposes a multi-scale transfer learning model based on an attention mechanism, which is used for aligning the feature distribution of the generated data and the target domain data. By introducing the attention mechanism and the multi-scale feature alignment strategy, the feature differences between different data domains are effectively captured and corrected, and the representation ability of the generated data and the classification performance of the target domain data are improved.

[0078] With reference to Figure 1 and Figure 2 , the early fault diagnosis method of rotating machinery under dynamic working conditions comprises:

[0079] S1, data preparation, comprising:

[0080] S11, acquiring early fault data; wherein the early fault data comprises 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] The early fault stage is relatively weak, the signal strength may be low, and it is easy to be covered by environmental noise or other signals, which is the key challenge in early fault diagnosis research. The weakness and confusion of the features require higher sensitivity and noise resistance of the diagnosis technology, and stronger feature discrimination ability. Therefore, in the framework of the generative data model GANs, the method explores to expand the labeled data of different working conditions by generating representative early fault samples, and improves the recognition accuracy of early fault signals by combining the feature learning ability of the model.

[0082] The specific acquisition method of related data is described in another application embodiment of the present application, and therefore will not be described again.

[0083] S12, dividing 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 the training set, 20% are used as the test set, and it is ensured that there is no overlap between the training data and the test data to avoid data leakage.

[0085] S2, training a DA-WGAN (i.e., Domain-Adaptive Wasserstein GAN) model and expanding data, comprising:

[0086] S21, generating a generative adversarial network, specifically:

[0087] The DA-WGAN model comprises a generator G, a discriminator D and a domain adaptive moduleFigure 2 The domain adaptation module includes a domain discriminator under the middle discriminator. The GAN model training process includes defining a game between two adversarial networks:

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

[0089]

[0090] wherein, is the real data distribution, is the generated data distribution implicitly defined by x=G(z), z~p(z), z is sampled from a simple noise distribution p, such as 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, the Jensen-Shannon (JS) divergence is used as the loss function of the standard GAN as a measure of generated samples; it is defined as: wherein, KL is the Kullback-Leibler divergence, Pm is the mixture

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

[0093] In this embodiment, the training of GAN is carried out using unsupervised learning of randomly generated data, and bearing fault diagnosis often has multiple fault types, and when multiple classes need to be learned, the deployment is very complex. In order to overcome this problem, this paper introduces the 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] wherein, y is the class label, and can also be any other type of auxiliary information. y is combined with the prior noise as the input of the hidden layer.

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

[0098] Whether it is C-GAN or standard GAN, the same problem behavior will be shown: gradient collapse and unstable training (gradient vanishing) and other problems. Therefore, the method uses Wasserstein distance (also known as Earth-Mover) instead of JS divergence as the distance between real data distribution and generated data distribution, because the smoothness of Wasserstein distance is better than that of 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 formula corresponding to the Wasserstein distance includes:

[0101] Among them, P(x) represents the real data distribution and the generated data distribution The entire feasible joint distribution γ(x, y) set,

[0102] is the minimum cost of the transmission quality of converting distribution to distribution .

[0103] In the discriminator D, connect and with y in the joint hidden representation; in the generator G, connect y with P(z) in the same representation form;

[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 unstable training, but also provide a reliable indicator for the training progress of GAN; in addition, the indicator is highly related to the quality of generated samples, however, the fitting ability of deep neural network is not fully utilized, and the gradient will disappear or explode after weight sharing.

[0107] To solve the above problems, the method adopts a gradient penalty (GP) method to meet the Lipschitz continuity condition, and the GP solves the problem of vanishing or exploding gradients during training, and considers the GP to obtain a new objective function:

[0108]

[0109] where λ is a gradient penalty coefficient, is a straight line sampled between the real data distribution and the generated data distribution

[0110] The domain adversarial loss function comprises:

[0111]

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

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

[0114] S24, the domain adversarial loss function is realized by a binary cross-entropy loss function, and the DA-WGAN loss is defined based on the domain adversarial loss function and the improved objective function , the domain discriminator in the generation model is built and the domain adversarial loss is set, and specifically:

[0115] The method first introduces GAN into the domain adaptation strategy, and the domain discriminator is a key component of DA-WGAN, and its task is to distinguish data from two different working conditions (A and B), without caring whether the data is real or generated. The purpose of the (domain discriminator) network is to help the generator generate data closer to the B working condition.

[0116] The domain adversarial loss function is used to measure the prediction accuracy of the domain discriminator D domain , and the goal of the domain adversarial loss function is to enable the domain discriminator to accurately distinguish whether the data is from the source domain or the target domain; in the method, the domain adversarial loss function is realized by a binary cross-entropy loss function, and the domain adversarial loss function comprises:

[0117] wherein is a domain loss function, D domain represents the predicted probability of the domain discriminator 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;

[0118] Based on the domain adversarial loss function and the improved target function Define DA-WGAN loss, DA-WGAN loss is defined as:

[0119]

[0120] S3, training of the multi-scale transfer learning classification model based on attention mechanism, which includes: taking the label data similar to the target working condition generated by the DA-WGAN model as the source domain, taking 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, specifically:

[0121] 1), unsupervised domain adaptation framework:

[0122] When the feature distribution of the source domain and target domain data sets is different, the traditional supervised learning algorithm is often difficult to achieve effective classification, and experiments have shown that the unsupervised domain adaptation is more suitable for the scene of the method:

[0123] Reference Figure 4 The premise of domain adaptation only needs the data of the source domain data set to have labels, so the source domain can be defined as follows:

[0124] Wherein 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 not available, the data of the target domain is defined as follows:

[0125]

[0126] Wherein, 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 the source domain data respectively obey the probability distribution P and Q, and the deep neural network can be established and learned to classify the unlabeled data in the target domain by the transferable features, which is represented as: Where, β(·) represents a deep neural network, The prediction result of the model.

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

[0128] The total loss of domain adaptation is defined as: Where λ is a trade-off parameter. represents a partial loss, which reduces the feature difference between the source domain and the target domain data. is the maximum cross-entropy loss, which is used to measure the difference distance between the prediction of the source domain data and the real label, 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 is established by three residual modules in parallel, which uses convolution layers with kernel sizes of 1x1, 1x3 and 1x9 to process input data in three branches in parallel to extract feature information of different scales, and at the end of the AMResNet module, the features extracted by each branch are used as the scale input-attention mechanism, to obtain the weights of different scales, and the weighted features of each branch are connected according to the dimension of the channel, to realize the construction of the feature extractor.

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

[0133] Residual learning is a technique that can effectively avoid model degradation, which uses the characteristics that the learning effect of shallow network is better than that of deep network when the model degrades, and directly inputs the features learned by shallow network to deep network across layers, so as to effectively solve the problem of model degradation caused by too much training error and improve the overall performance of the network, as shown in Figure 5 .

[0134] The residual network module usually adopts two or more layers of convolution operation, as shown in Figure 5 , which is mainly composed of an identity mapping part and a residual part, and the formula is as follows: y=F(x,W i)+x;

[0135] Where x represents the input of the current layer and y represents the output of the current layer, note the difference between this and the class label y in this method; F(x,W i ) is the residual function, representing the residual mapping to be learned. Figure 5 With two layers of convolution operations, the residual function can be expressed as: F = W2σ(W1x), where W represents the convolution operation and σ is generally a nonlinear activation function.

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

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

[0138] The above A i It is the attention feature after passing through a convolutional 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, this attention weight is used to weight the original features, resulting in a weighted feature vector:

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

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

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

[0143] We design a transfer learning strategy by adding MK-MMD to the loss function to achieve feature transformation between the source and target data. The final loss function is defined as follows:

[0144] Where, λ MK-MMD This is the trade-off parameter in the total loss, and This indicates a multi-core version of MMD. In this embodiment, only Gaussian kernels are 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. Defined in kernel Hilbert space (RKHS), MMD is the squared distance between the kernel embeddings of marginal distributions P(Xs) and Q(Xt). The MMD formula is as follows:

[0147]

[0148] in This indicates the use of the RKHS with kernel k (generally a Gaussian kernel is used as the kernel), and φ(·) represents the mapping to the RKHS. It represents the mathematical expectation.

[0149] In practical applications, the parameter selection for each kernel is crucial to the final performance of the mapping. To address this issue, this method utilizes MK-MMD, which maximizes the testing capability of two samples while simultaneously minimizing Type II error. For MK-MMD, this method provides an efficient estimate of the mapping using a convex combination of m kernels:

[0150]

[0151] Where, {β u} represents the weighted parameters for different kernels.

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

[0153] 4) Building a classifier, which includes: inputting the high-dimensional feature vectors that have undergone domain adaptive alignment into the classifier for classification. The classifier is built from two fully connected layers.

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

[0155] In another embodiment of this method, an application example is given: taking a bearing as an example in rotating machinery, an experiment on early bearing fault diagnosis under varying operating conditions is conducted:

[0156] (1) Build an early bearing fault diagnosis experimental platform to collect data. For example... Figure 7As shown, the bearing early fault diagnosis test bench includes a drive motor, a dynamic torque sensor, and an eddy current brake; a transmission shaft and a balance disc 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 and a complete fault simulation system.

[0157] (2) Fault Category Definition: This method considers six bearing health conditions, including one normal condition, three single faults, and two compound faults, as specifically defined in Table 1. Single faults include inner ring concavity, outer ring concavity, and misalignment; compound faults are the superposition of inner ring concavity and misalignment, and outer ring concavity and misalignment. To simulate early-stage faults, this method uses pitting etching to create concavity damage on the inner and outer rings, with the fault dimensions defined as a diameter of 0.2 mm and a depth of 0.1 mm. Misalignment faults are achieved by adding 0.2 mm shims to the bearing section.

[0158] Table 1. Six Bearing Health Conditions

[0159]

[0160] This experimental design aims to cover the most common failure modes and compound failure scenarios in rolling bearings, fully simulating actual operating conditions. Common bearing failures mainly focus on damage to the inner and outer rings, as well as issues such as misalignment during installation, which has been validated and adopted by numerous studies and classic datasets (such as the CWRU dataset). Building upon this, this method further introduces compound failures (such as the superposition of inner or outer ring indentation and misalignment), expanding the richness and complexity of the experiment to study the signal interaction effects under multiple failure modes. This design not only covers typical single failure types in bearings but also systematically explores the characteristics of compound failures.

[0161] (3) Operating Conditions: Three different speeds (2000 rpm, 3000 rpm, and 2000 rpm) and two radial loads (500 N and 1000 N) were considered in the experiment. Data for four operating conditions were collected for experimental verification, as shown in Table 2. For each fault, under a specific speed and torsional load, the sampling frequency was 25600 Hz, and a set of data was collected every 8 seconds, with 204,800 data points collected each time, for a total of 5 sets. For each fault, 1,024,000 data points were collected under each operating condition.

[0162] Table 2. Experimental parameters

[0163]

[0164] (4) Transfer Task Division: To verify the effectiveness of the proposed method, four transfer tasks were designed: task (D1→D2), task (D2→D1), task (D3→D4), and task (D4→D3). The left and right sides of the "→" sign represent the source domain dataset and the target domain dataset, respectively. For example, in the transfer task (D1→D2), the labeled data of the source domain D1 working condition is first expanded using a data generation model. Then, a large amount of labeled data is generated and used as the source domain data. The classification model is trained using the MK-MMD classification framework. The target domain D2 working condition data is unlabeled, but the target domain data is also used in the training.

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

[0166] In this embodiment, two upsampling layers are introduced in the generator G to gradually increase the dimensionality of the feature space, thereby enabling the generated data to have sufficient resolution and detail. Furthermore, the last layer of generator G employs the Tanh activation function, limiting the output range to [-1, 1] to maintain consistency with the range of the normalized data.

[0167] Discriminator D consists of multiple convolutional layers, progressively shrinking the feature map and finally outputting the true / false discrimination result through a linear layer. Discriminator D also uses Leaky ReLU activation to alleviate the vanishing gradient problem. More convolutional layers (e.g., six layers in total) are added to Discriminator D to capture more subtle distributional differences between the source and target domains. The last layer uses Sigmoid activation to output the probability of domain discrimination within the range [0,1]. Batch normalization (BN) is used in both the discriminator and the domain discriminator to stabilize the training process and ensure consistent distributional features across different input batches.

[0168] (6) Generating Fault Data: After generating noisy signals, it is crucial to evaluate and filter their quality. First, feature extraction and filtering 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, DA-WGAN is trained using the labeled data of condition D1 and the unlabeled data of condition D2 to generate labeled data for condition D2 with features similar to those 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 several aspects: in the time domain, the waveform of the generated signal is similar to that of the original signal, exhibiting similar amplitude and frequency characteristics; in the frequency domain, the spectral structure of the generated signal closely matches the dominant frequency component of the original signal; and in the time-frequency domain, the time-frequency distribution of the generated signal remains consistent with the original data.

[0170] (7) Result Comparison:

[0171] This method is compared with four other typical GAN ​​methods (GAN, WGAN, WGNGP, and SNGAN): the generative model is trained using 512,000 data points from the source domain, and the generated data is used to classify and identify the target working conditions. 512,000 samples (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 for different data augmentation methods.

[0173] Table 3. Diagnostic results (%) of transfer task D1→D2

[0174]

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

[0176]

[0177]

[0178] Table 5. Diagnostic results (%) of transfer task D3→D4

[0179]

[0180] Table 6. Diagnostic results (%) of transfer task D4→D3

[0181]

[0182] The table above shows that the fault classification accuracy of the five generative adversarial network variants (GAN, WGAN, SNGAN, WGNGP, DA-WGAN) varies significantly with different amounts of generated sample data under different transfer tasks (D1→D2, D2→D1, D3→D4, D4→D3). Overall, the diagnostic accuracy of all methods significantly improves with the increase of the amount of generated training data, especially when the amount of generated data increases from 512,000 to 512,000 × 3, resulting in a substantial improvement in the diagnostic accuracy of each method. This indicates that increasing the amount of data has a positive effect on the transfer learning effect of the model.

[0183] Reference Figure 9The result is the confusion matrix. Our method demonstrates significant fault diagnosis capabilities in the classification performance across four transfer tasks (D1→D2, D2→D1, D3→D4, D4→D3). In all transfer tasks, samples for most fault categories are concentrated on the diagonal of the confusion matrix, indicating that the model has high accuracy in classifying healthy states (category 0) and combined fault states (e.g., inner ring fault + misalignment fault, outer ring fault + misalignment fault).

[0184] This application also discloses an early fault diagnosis system for rotating machinery under dynamic operating conditions.

[0185] The early fault diagnosis system for rotating machinery under dynamic operating conditions has a memory that stores a computer program that can be loaded by a processor and executed as described above for the early fault diagnosis method for rotating machinery under dynamic operating conditions.

[0186] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

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

2. The method for early fault diagnosis of rotating machinery under dynamic operating conditions according to claim 1, characterized in that, The generative adversarial network includes: The DA-WGAN model is defined as consisting of 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. The process is represented as follows: ； in, It is the actual data distribution. It is by Implicitly defined generated data distribution, From simple noise distribution Sampling from the middle; Define the loss function of a standard GAN. , ; in, It is the Kullback-Leibler divergence, and Pm is the mixture ( + ) / 2.

3. The method for early fault diagnosis of rotating machinery under dynamic operating conditions according to claim 1, characterized in that, The objective function of the C-GAN includes: ； in, It is a class tag. It is combined with prior noise as input to the hidden layer.

4. The method for early fault diagnosis of rotating machinery under dynamic operating conditions according to claim 1, characterized in that, The formula corresponding to the Wasserstein distance includes: ； in, Represents the true data distribution and the distribution of generated data The entire feasible joint distribution Collection, W To distribute Transform into distribution The lowest cost for achieving high-quality transmission.

5. The method for early fault diagnosis of rotating machinery under dynamic operating conditions according to claim 1, characterized in that, The objective function It includes: In discriminator D, and It is concatenated with y in the joint hidden representation; in the generator G, y is concatenated with y in the same representation. Connect them; The objective function between the generator G and the discriminator D is a minimax game equation, and is: ； By using the gradient penalty method to satisfy the Lipschitz continuity condition, a new objective function is obtained: ； in It is the gradient penalty coefficient. It is along the distribution of real data and the distribution of generated data Linear sampling between; The domain adversarial loss function includes: ;in It is a domain loss function. The domain identifier predicts the probability that a sample belongs to the target domain. These are domain labels, represented by 0 and 1, where 0 represents the source domain and 1 represents the destination domain. It is the sample size; The DA-WGAN loss is defined as: 。 6. The method for early fault diagnosis of rotating machinery under dynamic operating conditions according to claim 1, characterized in that: The training of the attention-based multi-scale transfer learning classification model includes establishing an unsupervised adaptive framework. The establishment of the unsupervised domain adaptive framework includes: Unsupervised domain adaptation is chosen as the supervised learning algorithm, and the source domain is defined as: ； in Represents the source domain. Represents the first in the source domain One sample, It is the set of all samples. It is the first The first sample One tag, It is a collection of all tags. This represents the total number of samples in the source domain. Assuming that labels are unavailable in the target domain, the data definition for the target domain is as follows: ； in, Indicates the target domain. It is the first in the target domain One sample, It is the set of all samples. Indicates the total number of target samples; The target domain and source domain data respectively follow and Based on the probability distribution, a deep neural network is built and unlabeled data in the target domain is classified by learning transferable features, as follows: ,in, This represents a deep neural network. The model's prediction results; The purpose of domain adaptation is to minimize target domain risk by using source domain data supervision. ,and: ; The total loss for domain adaptation is defined as: ,in, It is a trade-off parameter; This represents partial loss, used to reduce feature differences between source and target domain data; It is the maximum cross-entropy loss, used to measure the distance between the predicted and true labels of the source domain data, and: ,in, It is the total number of all possible tags. It is an index function.

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

8. The method for early fault diagnosis of rotating machinery under dynamic operating conditions according to claim 1, characterized in that, The training of the attention-based multi-scale transfer learning classification model includes: We design a transfer learning strategy by adding MK-MMD to the loss function to achieve feature transformation between the source and target data. The final loss function is defined as follows: ,in, This is the trade-off parameter in the total loss, and This indicates the multi-core version of MMD.

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

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

Images