Gear fault diagnosis method based on transfer learning fault diagnosis model

Abstract

The invention discloses a gear fault diagnosis method based on a transfer learning fault diagnosis model, and relates to the gear fault diagnosis method based on the transfer learning fault diagnosis model. The objective of the invention is to solve the problem of low gear fault diagnosis accuracy caused by few fault samples and unbalanced data in the actual gear fault diagnosis process. The method comprises the following steps: constructing a gear training set; constructing a transfer learning fault diagnosis model, and obtaining a trained transfer learning fault diagnosis model; the specific process is as follows: a transfer learning fault diagnosis model sequentially comprises a feature extractor and a classifier; obtaining a trained transfer learning fault diagnosis model; the method comprises the following steps: inputting a gear training set into a feature extractor of a model, and outputting features by the feature extractor; inputting the features into a classifier, and outputting a classification result by the classifier; until the total loss converges, obtaining a trained model; and inputting a test set of an actual gearbox into the trained model to obtain a diagnosis result. The method is used in the field of gear fault diagnosis.

Description

Technical Field

[0001] The invention relates to a gear fault diagnosis method based on a transfer learning fault diagnosis model. Background Art

[0002] Gear systems are widely used in various mechanical equipment such as helicopters, vehicles, wind turbines, and mining machinery due to their high transmission efficiency and strong load-bearing capacity. Gears inevitably produce various faults due to their long-term operation under complex and harsh working conditions such as high temperature, high speed, and heavy load. Therefore, intelligent fault diagnosis of gears is of great significance to ensure personnel safety and equipment operation.

[0003] In recent years, data-driven methods, especially fault diagnosis methods based on deep learning, have received extensive attention. However, most studies are conducted under the assumption that data is balanced. In fact, the problem of data imbalance is inevitable. In actual industrial production, the frequency of mechanical equipment failure is much lower than that in normal state, and the frequency of occurrence of different types of failures is also different. Therefore, the data collected in the fault state is very limited, while a large amount of data can be obtained in the normal state, which leads to the problem of data imbalance. When using an unbalanced data set to train a deep learning model, most normal samples will dominate the training of the model, resulting in a strong bias in the model and a loss of diagnostic ability for faulty samples. In recent years, a lot of research has been carried out to solve the problem of data imbalance, and great progress has been made. At present, class imbalance fault diagnosis methods can be roughly divided into two categories: class re-balancing and information augmentation.

[0004] Class rebalancing is the mainstream solution for unbalanced fault diagnosis, which aims to rebalance the negative impact of class imbalance on the number of training samples. It mainly includes three categories: resampling, class-sensitive learning, and Logit adjustment. Resampling rebalances the categories by adjusting the number of samples in each class in the batch. Chatterjee et al. used SMOTE and ENN to oversample and undersample the fault samples respectively, and adopted SMOTE-ENN sampling technology to achieve unbalanced fault diagnosis of wind turbine data. Class-sensitive learning rebalances the training effect of unbalanced data by adjusting the training loss value of each class. Focal loss uses a modulation factor containing the predicted probability to weight the categories, assigning higher weights to the minority class and lower weights to the majority class. Online Hard Example Mining sorts the cross entropy loss of each sample from large to small, determines the difficult sample according to the threshold, and calculates the average loss of the difficult sample as the loss value. SEESAW reduces the gradient of the majority class to other classes by introducing a mitigation factor and a compensation factor, thereby reducing the suppression of other classes. Specifically, the mitigation factor reduces the penalty for the minority class according to the ratio of the number of samples between the classes. The compensation factor increases the penalty for misclassifying a class. Logit adjustment solves the class imbalance problem by adjusting the prediction logic of the model. Menon et al. proposed LAL (logistic adjustment loss) by simply modifying the cross entropy loss, which adjusts the predicted probability of the model based on the label frequency, focusing on the minority class. The class rebalancing method is simple and effective. However, most of the methods in this method have a common disadvantage, that is, the diagnostic effect of the minority class is improved by sacrificing the diagnostic effect of the majority class. In addition, although the overall diagnostic effect is improved, the problem of small amount of data in some categories cannot be changed.

[0005] Information enhancement aims to improve model performance by introducing additional information during the training process, mainly including data enhancement and transfer learning. Among the data enhancement methods, virtual sample generation based on GAN is one of the methods that has been widely studied in recent years. WGAN uses Wasserstein distance instead of Jensen-Shannon divergence in traditional GAN ​​as the loss function of the discriminator, which theoretically solves the problem of gradient vanishing. VAE_GAN uses VAE as the generator of GAN to improve the quality of virtual sample generation. LSGAN uses the least squares loss function to replace the original binary cross entropy loss, transforming the GAN objective into a least squares optimization problem, alleviating the gradient vanishing problem and promoting more stable training dynamics. However, the generator aims to generate high-quality data by learning from real data sets, while the discriminator attempts to distinguish between real data and generated data. Due to their conflicting optimization goals, GAN training often suffers from instability and model collapse. Transfer learning can greatly improve the model training effect by transferring knowledge from the source domain to the target domain. Li et al. proposed a multi-domain adversarial transfer framework to improve the cross-domain fault diagnosis accuracy of rolling bearings under data imbalance conditions. Kuang et al. embedded class imbalance learning into the adversarial training process and performed two-level adversarial transfer learning including marginal and conditional distribution adaptation to achieve cross-domain fault diagnosis of imbalanced data. However, in general, there is also a problem of lack of fault data in the source domain, so transfer learning also needs to be assisted by other means to solve the class imbalance problem. Summary of the invention

[0006] The purpose of the present invention is to solve the problem of low gear fault diagnosis accuracy caused by small number of fault samples and unbalanced data in the actual gear fault diagnosis process, and to propose a gear fault diagnosis method based on a transfer learning fault diagnosis model.

[0007] A gear fault diagnosis method based on transfer learning fault diagnosis model. The specific process is as follows:

[0008] Step 1: Construct a gear training set;

[0009] The gear training set includes actual unbalanced gear sample data and gear simulated fault sample data;

[0010] The actual unbalanced gear sample data is the collected gearbox vibration acceleration signal;

[0011] The sample data of gear simulation fault are the lateral and longitudinal acceleration signals of the center of mass of the driven wheel;

[0012] Step 2: Build a transfer learning fault diagnosis model to obtain a trained transfer learning fault diagnosis model; the specific process is as follows:

[0013] Step 21. The transfer learning fault diagnosis model includes a feature extractor and a classifier in sequence;

[0014] The feature extractor is a 1D-ResNet18 network model;

[0015] The 1D is 1 dimension;

[0016] The classifier comprises a linear layer, a ReLU activation function layer, and a linear layer in sequence;

[0017] Step 22: Obtain a trained transfer learning fault diagnosis model; the specific process is:

[0018] The gear training set is input into the feature extractor of the transfer learning fault diagnosis model, and the feature extractor outputs features;

[0019] The features are input into the classifier, and the classifier outputs the classification result;

[0020] Set the overall loss Loss of the transfer learning fault diagnosis model;

[0021] Until the overall loss converges, a trained transfer learning fault diagnosis model is obtained;

[0022] Step 3: Input the actual gearbox test set into the trained model to obtain the diagnosis result;

[0023] The diagnosis results include normal gear data and faulty gear data.

[0024] The beneficial effects of the present invention are:

[0025] (1) The present invention establishes an effective gearbox virtual model based on digital twin technology and generates a large amount of high-quality virtual fault data.

[0026] (2) The present invention adopts a new distribution difference measurement method (NDDM), combined with CORrelationAlignment (CORAL) and Local Maximum Mean Discrepancy (LMMD), to measure the distribution distance from the covariance and mean respectively, and pays attention to the marginal distribution alignment and subdomain conditional distribution alignment of the virtual domain and the actual domain, thereby improving the accuracy of gear fault diagnosis.

[0027] (3) The present invention adopts the LMF (large margin aware focal) loss function, and dynamically considers hard samples and category distribution at the same time by linearly combining focal loss and LDAM loss (label distribution aware margin loss), thereby further improving the generalization ability of the model and improving the accuracy of gear fault diagnosis.

[0028] (4) The present invention compares and verifies the proposed method with other methods on the laboratory self-test gear dataset and the Southeast University gear dataset. The results show that the proposed method is effective in diagnosing imbalance faults and improves the accuracy of gear fault diagnosis.

[0029] This paper proposes a digital twin-driven unbalanced gear fault diagnosis method based on a new distribution difference metric method and large margin-aware focal loss. First, an effective gearbox digital twin model is established to obtain a large amount of high-quality virtual fault data. Secondly, NDDM (new distribution difference metric) is used to measure the distribution distance between the virtual domain and the actual domain from the two aspects of covariance and mean, and the marginal distribution and subdomain conditional distribution of the two domains are aligned at the same time. Then, the characteristics of the two components FL (focal loss) and LDAM (label distribution-aware margin loss) of LMF (large margin-aware focal loss) are used respectively to consider the category distribution while considering hard samples, thereby improving the generalization ability of the model. Finally, the effectiveness of the proposed method is verified by the laboratory self-test gear dataset and the Southeast University gear dataset. By comparing with other methods, the advantages of this method in unbalanced fault diagnosis are proved, and the effectiveness of the joint application of digital twins and transfer learning is proved, providing a new solution to the fault diagnosis problem under unbalanced data, and solving the problem of low gear fault diagnosis accuracy caused by few fault samples and unbalanced data in the actual gear fault diagnosis process. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a flow chart of the present invention;

[0031] Figure 2 Figure 2 shows the virtual model of the gearbox: (a) gearbox, (b) three failure modes;

[0032] Figure 3 is the new distribution difference metric map;

[0033] Figure 4 Schematic diagram of large margin-aware focal loss. (a) Schematic diagram of large margin-aware focal loss. (b) Schematic diagram of label distribution-aware margin loss.

[0034] Figure 5 is the overall framework diagram, G η , G μ They are feature extractor and classifier respectively;

[0035] Figure 6 Diagram of the laboratory experiment setup, (a) experimental platform, (b) gears in four states;

[0036] Figure 7The average accuracy of each method is shown in Figure 2. (a) N1 dataset, (b) N2 dataset, (c) N3 dataset, and (d) the average of the experimental results of the three datasets (a), (b), and (c).

[0037] Figure 8 ROC curves of the methods are as follows: (a) the proposed method, (b) synthetic minority oversampling combined with edited nearest neighbor, (c) Seesaw, (d) logistic adjustment loss, (e) Wasserstein generative adversarial network, and (f) least squares generative adversarial network.

[0038] Fig. 9 The confusion matrix diagrams of each method are: (a) the proposed method, (b) synthetic minority oversampling combined with edited nearest neighbor, (c) Seesaw, (d) logistic adjustment loss, (e) Wasserstein generative adversarial network, and (f) least squares generative adversarial network.

[0039] Fig.10 Diagram of the Southeast University experiment setup, (a) experimental platform, (b) gears in four states;

[0040] Fig.11 This is a quantitative evaluation diagram of the characteristics of each method in the Southeast University gear imbalance dataset;

[0041] Fig.12 The F1 score graphs of each method, (a) S1 dataset, (b) S2 dataset, (c) S3 dataset, (d) the experimental results of the three datasets (a), (b), and (c) are averaged together;

[0042] Fig.13 t-SNE graph of each method, (a) the proposed method, (b) synthetic minority oversampling combined with edited nearest neighbors, (c) Seesaw, (d) logistic adjustment loss, (e) Wasserstein generative adversarial network, and (f) least squares generative adversarial network. DETAILED DESCRIPTION

[0043] Specific implementation method 1: The specific process of a gear fault diagnosis method based on a transfer learning fault diagnosis model in this implementation method is as follows:

[0044] Step 1: Construct a gear training set;

[0045] The gear training set includes actual unbalanced gear sample data and gear simulated fault sample data;

[0046] The actual unbalanced gear sample data is the vibration acceleration signal collected on the upper side of the gearbox and the outer side of the gearbox end cover;

[0047] The sample data of gear simulation fault are the lateral and longitudinal acceleration signals of the center of mass of the driven wheel;

[0048] Step 2: Build a transfer learning fault diagnosis model to obtain a trained transfer learning fault diagnosis model; the specific process is as follows:

[0049] Step 21. The transfer learning fault diagnosis model includes a feature extractor and a classifier in sequence;

[0050] The feature extractor is a 1D-ResNet18 network model;

[0051] The 1D is 1 dimension;

[0052] The classifier comprises a linear layer, a ReLU activation function layer, and a linear layer in sequence;

[0053] Step 22: Obtain a trained transfer learning fault diagnosis model; the specific process is:

[0054] The gear training set is input into the feature extractor of the transfer learning fault diagnosis model, and the feature extractor outputs features;

[0055] The features are input into the classifier, and the classifier outputs the classification result;

[0056] Set the overall loss Loss of the transfer learning fault diagnosis model;

[0057] Until the overall loss converges, a trained transfer learning fault diagnosis model is obtained;

[0058] Step 3: Input the actual gearbox test set into the trained model to obtain the diagnosis result;

[0059] The diagnosis results include normal gear data and faulty gear data.

[0060] Specific implementation method 2: This implementation method is different from the specific implementation method 1 in that a gear training set is constructed in step 1, and the gear training set includes actual unbalanced gear sample data and gear simulated fault sample data; the specific process is:

[0061] Step 11: Collect actual unbalanced gear data, where the actual unbalanced gear sample data includes normal gear sample data and faulty gear sample data;

[0062] Normal gear data is more, while faulty gear data is less, which is called unbalanced gear data;

[0063] The actual unbalanced gear sample data is the vibration acceleration signal collected on the upper side of the gearbox and the outer side of the gearbox end cover;

[0064] The gear fault sample data includes broken gears, cracked gears and peeled gears;

[0065] Step 1 and 2: Obtain gear simulation fault sample data; the specific process is as follows:

[0066] The faulty gear 3D model (3D diagram, such as the gear missing tooth 3D diagram, Figure 2 (a) is input into ADAMS software, and ADAMS software outputs gear simulation fault sample data;

[0067] The sample data of gear simulation fault are the lateral and longitudinal acceleration signals of the center of mass of the driven wheel;

[0068] In ADAMS software, define the material of the 3D model, set constraints such as the revolute pair, define the contact function between the gears, add drive and load to the driving wheel and the driven wheel respectively. ADAMS software will automatically generate the dynamic equation, set the end time and step length, and click simulation. ADAMS software will automatically output the simulated fault gear data;

[0069] It can accurately simulate the vibration response of gears under different fault types and generate a large amount of high-quality virtual fault data. Figure 2 As shown in (b), three fault modes of tooth breakage, cracking and spalling are pre-simulated in the virtual space, the sampling frequency is set to 10kHz, and the lateral and longitudinal acceleration signals of the center of mass of the driven wheel are collected as virtual fault data.

[0070] The other steps and parameters are the same as those in the first embodiment.

[0071] Specific implementation method three: This implementation method is different from specific implementation method one or two in that the process of obtaining the overall loss Loss of the transfer learning fault diagnosis model in step 22 is:

[0072] Step 1) Set a new distribution difference metric loss L NDDM ;

[0073] Step 2) Construct large margin-aware focal loss L LMF ;

[0074] Step 3), based on the new distribution difference metric L NDDM and constructed large margin-aware focal loss L LMF , obtain the overall loss Loss of the transfer learning fault diagnosis model.

[0075] The other steps and parameters are the same as those in the first or second embodiment.

[0076] Specific implementation method 4: This implementation method is different from any one of the specific implementation methods 1 to 3 in that a new distribution difference metric loss L is set in step 1). NDDM ; The specific process is:

[0077] In order to reduce the distribution space difference between virtual data and real data, this paper adopts a new distribution difference measurement method that combines CORrelationAlignment (CORAL) (correlation alignment) and LocalMaximumMeanDiscrepancy (LMMD) (local maximum mean difference). The distribution distance is measured from the covariance and mean, and the marginal distribution alignment and subdomain conditional distribution alignment of the virtual domain and the real domain are paid attention to, such as Figure 3 shown.

[0078] Step 11), construct the relevant alignment loss L CORAL (P,Q);

[0079] Step 12), construct the local maximum mean difference loss L LMMD (P (c) ,Q (c) );

[0080] Step 13), based on the loss L constructed in step 11) CORAL (P,Q) and the loss L constructed in step 12) LMMD (P (c) ,Q (c) ), and obtain the new distribution difference metric loss L NDDM .

[0081] The other steps and parameters are the same as those in Specific Embodiments 1 to 3.

[0082] Specific implementation method 5: This implementation method is different from any one of the specific implementation methods 1 to 4 in that the related alignment loss L is constructed in step 11). CORAL (P,Q); the specific process is:

[0083]

[0084] Among them, L CORAL (P,Q) represents the correlation alignment loss (CORAL);

[0085] P and Q represent the distribution of virtual domain and real domain, respectively;

[0086] The virtual domain is the gear simulation fault sample data;

[0087] The actual domain is the actual unbalanced gear sample data;

[0088] C S and C T They represent the feature covariance matrix of virtual domain samples and the feature covariance matrix of actual domain samples respectively;

[0089] Virtual domain sample features are features output after the gear simulation fault sample data is input into the feature extractor;

[0090] The actual domain sample features are the features output after the actual unbalanced gear sample data is input into the feature extractor;

[0091] ||·|| F represents the Frobenius norm;

[0092] d represents the feature dimension of each sample, which is 512 here;

[0093] I is a column vector whose elements are all equal to 1;

[0094] n S and n T They represent the sample data volume of the virtual domain and the sample data volume of the actual domain respectively;

[0095] represents the virtual domain sample feature matrix, It represents the features output after the first sample in the virtual domain is input into the feature extractor. It represents the features output after the second sample in the virtual domain is input into the feature extractor. Indicates the nth S The features output after samples are input into the feature extractor;

[0096] represents the actual domain sample feature matrix, It represents the features output after the first sample in the actual domain is input into the feature extractor. It represents the features output after the second sample in the actual domain is input into the feature extractor. Indicates the nth T The features output after samples are input into the feature extractor;

[0097] The subscripts S and T denote the virtual domain and the real domain, respectively;

[0098] The superscript ' indicates transposition.

[0099] The other steps and parameters are the same as those in Specific Embodiments 1 to 4.

[0100] Specific implementation method 6: This implementation method is different from any one of the specific implementation methods 1 to 5 in that the local maximum mean difference loss L is constructed in step 12). LMMD (P (c) ,Q (c) ); the specific process is:

[0101] LMMD considers the distribution of related subdomains within the same category in both virtual and real domains.

[0102] D S and D T Divide into subdomains based on category and c is the class label;

[0103] L LMMD (P (c) ,Q (c) ) is expressed as:

[0104]

[0105] Among them, L LMMD (P (c) ,Q (c) ) represents the local maximum mean difference loss (i.e., LMMD loss);

[0106] x s Denotes the virtual domain sample feature matrix D S Features in;

[0107] x t Represents the actual domain sample feature matrix D T Features in;

[0108] P (c) and Q (c) They are and Distribution of

[0109] ||·|| H represents the reproducing kernel Hilbert space (RKHS);

[0110] represents the mapping function in the reproducing kernel Hilbert space (RKHS);

[0111] It means to find the expectation of features in the virtual domain;

[0112] It means to find the expectation of features in the actual domain;

[0113] E c It means to find the expectation for category c;

[0114] If a sample is based on the weight w c Belongs to a certain class, then the above formula is rewritten as:

[0115]

[0116] in, and They are and The weights of class c, and

[0117] is the total number of categories;

[0118] is the i-th sample in the virtual domain, is the jth sample in the actual domain;

[0119] The i-th sample in the virtual domain The real label The weight of each sample in the virtual domain is represented by a one-hot vector It is expressed as:

[0120]

[0121] Each sample in the actual domain The real label The weight of each sample in the actual domain is represented by a one-hot vector It is expressed as:

[0122]

[0123] The other steps and parameters are the same as those in Specific Implementation Methods 1 to 5.

[0124] Specific implementation method 7: This implementation method is different from any one of the specific implementation methods 1 to 6 in that the loss L constructed based on step 11) in step 13) CORAL (P,Q) and the loss L constructed in step 12) LMMD (P (c) ,Q (c) ), and obtain the new distribution difference metric loss L NDDM (i.e. NDDM loss); the specific process is:

[0125] Distribution difference metric loss L NDDM It is expressed as:

[0126]

[0127] The other steps and parameters are the same as those in Specific Embodiments 1 to 6.

[0128] Specific implementation eight: This implementation differs from any one of specific implementations one to seven in that the large margin perceived focus loss L is constructed in step 2). LMF ; The specific process is:

[0129] In the class imbalance problem, the shortcoming of the cross-entropy loss function is that it gives equal weight to all classes. In order to further solve the class imbalance problem and improve the generalization ability of the model, the present invention adopts the LMF (large margin perceptual focus) loss function, such as Figure 4 As shown in (a), the function dynamically considers both hard samples and category distribution by linearly combining Focalloss (focus loss) and LDAMloss (label distribution aware margin loss).

[0130] Step 21), construct the focal loss L FL (p t )(Focal loss);

[0131] FocalLoss reduces the loss contribution by reducing the weight of the majority class;

[0132] Focus loss L FL (p t ) is specifically expressed as:

[0133] L FL (p t )=-(1-p t ) γ log(p t )

[0134] Among them, p t is the sample prediction probability of the model output after the transfer learning fault diagnosis model, and γ is the adjustment factor;

[0135] When γ = 0, the loss function Focal loss becomes cross entropy loss; the larger γ is, the lower the weight of most samples is, and the higher the attention paid to hard samples is;

[0136] Step 22), construct label distribution aware margin loss L LDAM (x,y)(LDAM loss); the specific process is:

[0137] like Figure 4 As shown in (b), the LDAM loss regularizes the minority class by maximizing the margin of the minority class (the minimum distance from the data point to the decision boundary);

[0138] For a sample x with label y, the label distribution aware margin loss L LDAM (x,y) is represented as:

[0139]

[0140] Among them, y is the true label of sample x, z cThe output of the model after the transfer learning fault diagnosis model is input for sample x. Sample x belongs to class c, Δ c is the margin of class c, c represents the category, n c is the number of samples in category c, and N is a fixed constant; obviously, the smaller the number of samples, the larger the margin.

[0141] Obviously, both Focal loss and LDAM loss are optimizations of the cross entropy loss function. The former makes the model pay more attention to a few samples, while the latter considers the category distribution of the data set to determine the weight.

[0142] Step 23), based on the focus loss L FL (p t ) and label distribution aware margin loss L LDAM (x,y), obtain the large margin-aware focal loss L LMF .

[0143] The other steps and parameters are the same as those in Specific Embodiments 1 to 7.

[0144] Specific implementation method 9: This implementation method is different from any one of the specific implementation methods 1 to 8 in that the step 23) is based on the focus loss L FL (p t ) and label distribution aware margin loss L LDAM (x,y), obtain the large margin-aware focal loss L LMF ; The specific process is:

[0145] L LMF =αL LDAM +βL FL

[0146] Among them, α and β are adjustable constants, and when α=β, a good effect will be produced.

[0147] Large margin-aware focus loss L LMF is the focal loss L FL (p t ) and label distribution aware margin loss L LDAM A linear combination of (x,y).

[0148] The other steps and parameters are the same as those in Specific Embodiments 1 to 8.

[0149] Specific implementation method 10: This implementation method is different from any one of specific implementation methods 1 to 9 in that the new distribution difference metric L is set in step 3). NDDM and constructed large margin-aware focal loss L LMF , obtain the overall loss Loss of the transfer learning fault diagnosis model;

[0150] The overall loss Loss expression is:

[0151] Loss(θ η ,θ μ )=λ t L NDDM +L LMF

[0152] Among them, θ η and θ μ are the learning parameters of the feature extractor and classifier, respectively;

[0153] is the time-varying weight coefficient, t and t m are the current number of iterations and the maximum number of iterations respectively.

[0154] Combining the above three parts, the present invention proposes a digital twin drive imbalance fault diagnosis method based on a new distribution difference measurement method and large margin perception focus loss, in order to achieve a breakthrough in the field of imbalance fault diagnosis. Figure 5 shown.

[0155] The other steps and parameters are the same as those in Specific Embodiments 1 to 9.

[0156] Overall process:

[0157] Data balance: Establish an effective gearbox virtual model based on digital twin technology and generate a large amount of high-quality virtual fault data.

[0158] Model training: First, the simulated fault data and the actual imbalanced data are simultaneously input into the feature extractor, where the 1D-ResNet18 network is selected as the feature extractor. Then, the features and labels of the virtual domain and the actual domain are input into the new distribution difference metric method (NDDM) to obtain the new distribution difference metric loss. Subsequently, the features of all data are input into the classifier to obtain the predicted label, and two linear layers are selected as the classifier. The output logical value and label of the classifier are input into the large margin perceptual focal loss (LMF) to obtain the large margin perceptual focal loss. The new distribution difference metric loss plus the large margin perceptual focal loss is the overall loss.

[0159] During the training iteration, the overall loss is minimized using the Adam optimizer with weight decay regularization, and the model with the lowest loss value of the training set is selected as the well-trained model. The test set of the actual gearbox is input into the trained model to obtain the diagnosis results, and multi-dimensional visualization analysis is performed on this basis.

[0160] Class rebalancing and information enhancement are the mainstream solutions for unbalanced fault diagnosis, but the former often sacrifices the effect of the majority class, and the latter often faces the problems of unstable training and lack of fault data. To avoid these problems, a digital twin-driven unbalanced fault diagnosis method based on a new distribution difference metric method and large margin-aware focal loss is proposed. Firstly, based on the analysis of the nonlinear dynamic characteristics of the gear pair, an effective gearbox virtual model is established based on digital twin technology, and a large amount of high-quality virtual fault data is generated. Secondly, a new distribution difference metric method (NDDM) is adopted to focus on the marginal distribution alignment and subdomain conditional distribution alignment of the virtual domain and the actual domain. Finally, the LMF loss function is used to dynamically consider hard samples and class distribution at the same time to further improve the generalization ability of the model. Through experimental verification on two datasets, the proposed unbalanced fault diagnosis method is significantly better than other methods, which proves the effectiveness of the proposed method in the field of unbalanced fault diagnosis. The effectiveness of NDDM and LMF is verified by ablation experiments.

[0161] Related Work

[0162] Digital twin

[0163] Digital twin technology is the key to realizing intelligent manufacturing and industrial digital transformation, and has received extensive attention and research from industry and academia. Driven by data and models, digital twins can perform operations such as monitoring, simulation, prediction, and optimization. In the field of fault diagnosis, because machines are often repaired before failure, the lack of fault data is one of the key challenges. Digital twins can provide real-time representation of physical equipment, generate data that can be used by fault diagnosis algorithms, and combine data-driven methods to achieve predictive maintenance of equipment.

[0164] In 2021, Xia et al. established a digital twin model of a three-cylinder pump and developed a new network structure based on a new sparse denoising autoencoder, which realized intelligent fault diagnosis using simulation data and limited measurement data. In 2023, Zhang et al. proposed a local domain adaptive network driven by digital twins, which generated simulation data by establishing a rolling bearing dynamics model for rolling bearing fault diagnosis when training data is insufficient. Wang et al. proposed an intelligent fault diagnosis model for a three-cylinder pump assisted by digital twins, and trained a transfer learning model based on domain adversarial strategy and Wasserstein distance using simulation data. Wang et al. proposed an application method combining a numerical simulation model with machine learning classification, which obtained normal and faulty simulation data by modifying relevant fault influencing factors and introducing the required fault types, and used a machine learning model to predict the probability of each fault, guiding the adjustment of actual parameters and the determination of maintenance plans. The present invention further considers the impact of measurement data imbalance, that is, the fact that fault data is far less than normal data, on the fault diagnosis effect.

[0165] Maximum mean discrepancy

[0166] The maximum mean difference is the most frequently used distribution difference measurement method in transfer learning. It aims to map variables to a high-dimensional space through a mapping function and calculate the mean difference between two distributions in this space. It is specifically expressed as:

[0167]

[0168] Where P and Q represent the distribution of virtual domain and real domain respectively, x s and x t denote samples from the virtual domain and the real domain respectively, n S and n T Indicates the amount of data, and Indicates n S and n T A single sample, E[·] represents the expected operation, ||·|| H represents the Reproducing Kernel Hilbert Space (RKHS), Represents the mapping function in RKHS. The mapping function is generally replaced by a kernel function, i.e., kernel technique. Usually, a Gaussian kernel function is chosen, specifically:

[0169]

[0170] Where σ is the bandwidth of the Gaussian kernel. After using the kernel technique, the formula L MMD (P,Q) can be rewritten as:

[0171]

[0172] in, and Indicates that it is different from and , and k(·,·) represents the kernel function used to measure the similarity between two samples in the feature space.

[0173] Experimental verification

[0174] This experiment uses the laboratory self-test gear dataset and the Southeast University gear dataset to fully verify the advantages of the proposed imbalance fault diagnosis method. In addition, the present invention also conducts ablation experiments and tests different modules to evaluate their impact on system performance.

[0175] Laboratory self-tested gear data set

[0176] The laboratory self-test gear data set includes one normal state and five fault states. The experimental platform used for data acquisition is as follows: Figure 6As shown in (a), the sampling frequency is set to 10kHz. This experiment selects four health states: normal, missing teeth, gear crack, and gear peeling. Figure 6 As shown in (b), the vibration acceleration signals on the upper side of the gearbox and the outer side of the gearbox end cover were tested when the load was 0.5A and the motor speed was 1200rpm.

[0177] For the laboratory self-tested gear dataset, three unbalanced datasets N1, N2 and N3 with imbalance rates of 29 times, 61 times and 125 times respectively are constructed, as shown in Table 1. Taking dataset N1 as an example, it includes 464 normal samples, 16 missing teeth samples, 16 gear crack samples and 16 gear peeling samples. In addition, for the proposed method, in the N1, N2 and N3 datasets, virtual fault data of 512 samples of each of the three fault states are also included.

[0178] Table 1 Unbalanced data set of gears tested in the laboratory

[0179]

[0180] Experimental verification of laboratory self-tested gear dataset

[0181] This experiment aims to compare the proposed imbalance fault diagnosis method with class rebalancing and information augmentation methods, including: resampling method SME, class sensitive learning methods FL, OHEM and SEESAW, Logit adjustment method LAL, data augmentation methods WGAN, VAE_GAN and LSGAN, a total of 8 methods. In order to increase the reliability of the experimental results, five independent tests were conducted on each method, and the experimental conditions were kept consistent. This section uses accuracy and feature quantitative evaluation index J to quantitatively evaluate the performance of various methods, and further verifies the effectiveness of the proposed method through ROC curve and confusion matrix.

[0182] Table 2 Diagnostic accuracy of each method in the laboratory self-test gear imbalance dataset

[0183]

[0184] As shown in Table 2, in the N1, N2 and N3 datasets, the average, maximum and minimum values ​​of the accuracy of the proposed method are the highest. Figure 7As shown in the figure, the annular bar graphs of the average accuracy of all methods in the N1, N2, and N3 datasets and the averaged annular bar graphs are plotted respectively. It is obvious that the proposed method has the highest accuracy. As the imbalance increases, the advantages of the proposed method become more obvious, but the accuracy of the data enhancement method decreases significantly. This is because the limited fault information easily leads to the model being unable to fully learn the characteristics of the fault. The feature quantitative evaluation index J quantifies the clustering effect of different methods through the intra-class covariance and inter-class covariance. For details, see the reference. As can be seen from Table 3, in the N1, N2, and N3 datasets, the average, maximum, and minimum values ​​of the J value of the proposed method are higher than those of other methods, which further proves the effectiveness of the proposed method.

[0185] Table 3 Quantitative evaluation of the characteristics of each method in the laboratory self-test gear imbalance dataset

[0186]

[0187] In order to further compare the performance of various methods under each health state, SME, SEESAW, LAL, WGAN, and LSGAN with better diagnostic effects are selected for comparison with the proposed method. Figure 8 The ROC curve of the average diagnosis result of five experiments on the N1 data set is: Fig. 9 The confusion matrix of the average diagnosis results of five experiments on the N2 dataset. Figure 8 (ROC curve) shows that the proposed method has good feature extraction performance in all gear health conditions, but other methods have poor performance in some health conditions. The area enclosed by the ROC curve and the coordinate axis is defined as the area under the curve (AUC). The AUC can intuitively evaluate the prediction performance. The higher the AUC value, the higher the prediction accuracy. It is obvious that the proposed method is better than other methods. Fig. 9 (Confusion Matrix) It can be seen that all methods have high diagnostic accuracy for normal samples. Although the diagnostic accuracy of the proposed method for missing teeth samples is 3% and 1% lower than that of SME and SEESAW, respectively, the diagnostic accuracy of the proposed method for gearpeeling samples is 7% and 20% higher than that of SME and SEESAW, respectively. Moreover, the proposed method has the highest diagnostic accuracy for gear crack and gearpeeling samples.

[0188] Southeast University Gear Dataset

[0189] The Southeast University gear data set includes one normal state and four fault states. The experimental platform used for data collection is as follows: Fig.10As shown in (a), the sampling frequency is set to 5120 Hz. This experiment selects four health states: normal, missing teeth, gear crack, and gear peeling. Fig.10 As shown in (b), the vibration signal of the motor in the Z direction and the vibration signal of the planetary gear in the X direction were tested when the load was 2V and the motor speed was 1800rpm.

[0190] For the Southeast University gear dataset, three unbalanced datasets S1, S2, and S3 with imbalance ratios of 61, 125, and 253 times, respectively, were constructed, and the detailed data are shown in Table 4. In addition, for the proposed method, virtual fault data of 512 samples each of the three fault states are also included in the S1, S2, and S3 datasets.

[0191] Table 4 Unbalanced data set of gears of Southeast University

[0192]

[0193] Experimental Verification of Southeast University Gear Dataset

[0194] Similar to the experimental verification of the laboratory self-tested gear dataset, 8 methods were selected to compare with the proposed method, and five independent experiments were conducted on each method. The accuracy, feature quantitative evaluation index J and F1 score were used to quantitatively evaluate the performance of various methods, and the t-SNE graph was used to further verify the effectiveness of the proposed method.

[0195] Table 5 Diagnostic accuracy of each method in the Southeast University gear imbalance dataset

[0196]

[0197] As shown in Table 5, in the S1, S2 and S3 datasets, the average, maximum and minimum values ​​of the accuracy of the proposed method are the highest. In the S1 dataset, the resampling method, class sensitive learning method and Logit adjustment method in the class rebalancing method have achieved good results. However, as the imbalance increases, the diagnostic effect of the class rebalancing method decreases significantly due to the constraint of limited fault information. Especially when the data is extremely unbalanced, that is, the S3 dataset, the resampling method SME almost loses its effect. The accuracy of the data enhancement method still shows a significant decrease. Compared with other methods, the diagnostic performance of the proposed method is obvious.

[0198] The J values ​​of each method in the S1, S2 and S3 datasets are shown in Fig.11As shown, the results of the five experiments are represented by black dots. As the imbalance increases, the J value of each method decreases to varying degrees, but the J value of the proposed method is larger. In the S1 dataset, only one J value of VAE_GAN is close to the proposed method. In the S2 dataset, the five J values ​​of the proposed method are higher than those of other methods. In the S3 dataset, the minimum J value of the proposed method is close to the maximum J value of other methods. This further proves the clustering effect of the proposed method. The F1 score and average F1 score of each method in the S1, S2, and S3 datasets are shown in Figure 2. Fig.12 The results are the average results of five experiments. As the imbalance increases, the advantages of the proposed method are obvious. Moreover, the proposed method has the best diagnostic effect on the health status of all gears in all cases.

[0199] In order to further compare the performance of various methods under each health state, SME, SEESAW, LAL, WGAN, and LSGAN with better diagnostic effects are still selected for comparison with the proposed method. Fig.13 The t-SNE graph of the experiment with the experimental results closest to the average value in the S1 dataset. It can be seen from the figure that the clusters formed by the proposed method in the same category are the tightest, and the boundaries between different categories are also the clearest. The boundaries between missing teeth and gear cracks in the LSGAN and LAL methods are somewhat blurred. SME clusters some gear crack samples into the missing teeth area. WGAN and SEESAW have difficulty in defining the boundaries between missing teeth and gear cracks. Normal samples with enough samples have good clustering effects in all.

[0200] Ablation experiment

[0201] In order to verify the effectiveness of LMF and NDDM, ablation experiments were conducted on the above six imbalanced data sets. The average accuracy of five independent experiments is shown in Table 6. None means that LMF and NDDM are not used, and only virtual fault data is used for training. LMF and NDDM respectively indicate that only LMF and NDDM are used, and Proposed indicates that LMF and NDDM are used at the same time. The results of the ablation experiment show that LMF improves the accuracy of imbalanced fault diagnosis, and NDDM plays a more important role in improving the accuracy. The combined use of LMF and NDDM has great potential in the field of imbalanced fault diagnosis.

[0202] Table 6 Comparison results of ablation experiments on imbalanced datasets

[0203]

[0204] The present invention may also have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A gear fault diagnosis method based on a transfer learning fault diagnosis model, characterized in that: The specific process of the method is: Step 1: Construct a gear training set; The gear training set includes actual unbalanced gear sample data and gear simulated fault sample data; The actual unbalanced gear sample data is the collected gearbox vibration acceleration signal; The sample data of gear simulation fault are the lateral and longitudinal acceleration signals of the center of mass of the driven wheel; Step 2: Build a transfer learning fault diagnosis model to obtain a trained transfer learning fault diagnosis model; the specific process is as follows: Step 21. The transfer learning fault diagnosis model includes a feature extractor and a classifier in sequence; The feature extractor is a 1D-ResNet18 network model; The 1D is 1 dimension; The classifier comprises a linear layer, a ReLU activation function layer, and a linear layer in sequence; Step 22: Obtain a trained transfer learning fault diagnosis model; the specific process is: The gear training set is input into the feature extractor of the transfer learning fault diagnosis model, and the feature extractor outputs features; The features are input into the classifier, and the classifier outputs the classification result; Set the overall loss Loss of the transfer learning fault diagnosis model; Until the overall loss converges, a trained transfer learning fault diagnosis model is obtained; Step 3: Input the actual gearbox test set into the trained model to obtain the diagnosis result; The diagnosis results include normal gear data and faulty gear data.

2. A gear fault diagnosis method based on transfer learning fault diagnosis model according to claim 1, characterized in that: In the step 1, a gear training set is constructed, and the gear training set includes actual unbalanced gear sample data and gear simulated fault sample data; the specific process is: Step 11: Collect actual unbalanced gear data, where the actual unbalanced gear sample data includes normal gear sample data and faulty gear sample data; The actual unbalanced gear sample data is the collected gearbox vibration acceleration signal; The gear fault sample data includes broken gears, cracked gears and peeled gears; Step 1 and 2: Obtain gear simulation fault sample data; the specific process is as follows: The three-dimensional model of the faulty gear constructed according to the actual gear shape and size is input into the ADAMS software, and the ADAMS software outputs the gear simulation fault sample data; The sample data of gear simulation fault are the lateral and longitudinal acceleration signals of the center of mass of the driven wheel.

3. A gear fault diagnosis method based on transfer learning fault diagnosis model according to claim 2, characterized in that: The process of obtaining the overall loss Loss of the transfer learning fault diagnosis model in step 22 is: Step 1) Set a new distribution difference metric loss L NDDM ; Step 2) Construct large margin-aware focal loss L LMF ; Step 3), based on the new distribution difference metric L NDDM and constructed large margin-aware focal loss L LMF , obtain the overall loss Loss of the transfer learning fault diagnosis model.

4. A gear fault diagnosis method based on transfer learning fault diagnosis model according to claim 3, characterized in that: In step 1), a new distribution difference metric loss L is set NDDM ; The specific process is: Step 11), construct the relevant alignment loss L CORAL (P,Q); Step 12), construct the local maximum mean difference loss L LMMD (P (c) ,Q (c) ); Step 13), based on the loss L constructed in step 11) CORAL (P,Q) and the loss L constructed in step 12) LMMD (P (c) ,Q (c) ), and obtain the new distribution difference metric loss L NDDM .

5. A gear fault diagnosis method based on transfer learning fault diagnosis model according to claim 4, characterized in that: In step 11), the relevant alignment loss L is constructed CORAL (P,Q); the specific process is: Among them, L CORAL (P,Q) represents the correlation alignment loss; P and Q represent the distribution of virtual domain and real domain, respectively; The virtual domain is the gear simulation fault sample data; The actual domain is the actual unbalanced gear sample data; C S and C T They represent the feature covariance matrix of virtual domain samples and the feature covariance matrix of actual domain samples respectively; Virtual domain sample features are features output after the gear simulation fault sample data is input into the feature extractor; The actual domain sample features are the features output after the actual unbalanced gear sample data is input into the feature extractor; ||·|| F represents the Frobenius norm; d represents the feature dimension of each sample; I is a column vector whose elements are all equal to 1; n S and n T They represent the sample data volume of the virtual domain and the sample data volume of the actual domain respectively; represents the virtual domain sample feature matrix, It represents the features output after the first sample in the virtual domain is input into the feature extractor. It represents the features output after the second sample in the virtual domain is input into the feature extractor. Indicates the nth S The features output after samples are input into the feature extractor; represents the actual domain sample feature matrix, It represents the features output after the first sample in the actual domain is input into the feature extractor. It represents the features output after the second sample in the actual domain is input into the feature extractor. Indicates the nth T The features output after samples are input into the feature extractor; The subscripts S and T denote the virtual domain and the real domain, respectively; The superscript ' indicates transposition.

6. A gear fault diagnosis method based on transfer learning fault diagnosis model according to claim 5, characterized in that: In step 12), the local maximum mean difference loss L is constructed. LMMD (P (c) ,Q (c) ); the specific process is: D S and D T Divide into subdomains based on category and c is the class label; L LMMD (P (c) ,Q (c) ) is expressed as: Among them, L LMMD (P (c ),Q (c) ) represents the local maximum mean difference loss; x s Denotes the virtual domain sample feature matrix D S Features in; x t Represents the actual domain sample feature matrix D T Features in; P (c) and Q (c) They are and Distribution of ||·|| H represents the reproducing kernel Hilbert space; represents the mapping function in the reproducing kernel Hilbert space; It means to find the expectation of features in the virtual domain; It means to find the expectation of features in the actual domain; E c It means to find the expectation for category c; If a sample is based on the weight w c Belongs to a certain class, then the above formula is rewritten as: in, and They are and The weights of class c, and is the total number of categories; is the i-th sample in the virtual domain, is the jth sample in the actual domain; The i-th sample in the virtual domain The real label The weight of each sample in the virtual domain is represented by a one-hot vector It is expressed as: Each sample in the actual domain The real label The weight of each sample in the actual domain is represented by a one-hot vector It is expressed as:

7. A gear fault diagnosis method based on transfer learning fault diagnosis model according to claim 6, characterized in that: The loss L constructed in step 13) is based on step 11) CORAL (P,Q) and the loss L constructed in step 12) LMMD (P (c) ,Q (c) ), and obtain the new distribution difference metric loss L NDDM ; The specific process is: Distribution difference metric loss L NDDM It is expressed as:

8. A gear fault diagnosis method based on transfer learning fault diagnosis model according to claim 7, characterized in that: In step 2), a large margin-aware focus loss L is constructed. LMF ; The specific process is: Step 21), construct the focal loss L FL (p t ); Focus loss L FL (p t ) is specifically expressed as: L FL (p t )=-(1-p t ) γ log(p t ) Among them, p t is the sample prediction probability of the model output after the transfer learning fault diagnosis model, and γ is the adjustment factor; Step 22), construct label distribution aware margin loss L LDAM (x,y); the specific process is: For a sample x with label y, the label distribution aware margin loss L LDAM (x,y) is represented as: Among them, y is the true label of sample x, z c The output of the model after the transfer learning fault diagnosis model is input for sample x. Sample x belongs to class c, Δ c is the margin of class c, c represents the category, n c is the number of samples in category c, and N is a fixed constant; Step 23), based on the focus loss L FL (p t ) and label distribution aware margin loss L LDAM (x,y), obtain the large margin-aware focal loss L LMF .

9. A gear fault diagnosis method based on transfer learning fault diagnosis model according to claim 8, characterized in that: In step 23), based on the focus loss L FL (p t ) and label distribution aware margin loss L LDAM (x,y), obtain the large margin-aware focal loss L LMF ; The specific process is: L LMF =αL LDAM +βL FL Among them, α and β are constants.

10. A gear fault diagnosis method based on transfer learning fault diagnosis model according to claim 9, characterized in that: In step 3), the new distribution difference metric L is set based on NDDM and constructed large margin-aware focal loss L LMF , obtain the overall loss Loss of the transfer learning fault diagnosis model; The overall loss Loss expression is: Loss(θ η ,i μ )=λ t L NDDM +L LMF Among them, θ η and θ μ are the learning parameters of the feature extractor and classifier, respectively; is the time-varying weight coefficient, t and t m are the current number of iterations and the maximum number of iterations respectively.

Images