A composite weight-based selective migration fault diagnosis method for mechanical equipment

Abstract

A selective migration fault diagnosis method for mechanical equipment based on composite weights comprises the following steps: first, using an acquisition device to collect vibration signals under various operating conditions from the mechanical equipment to establish data sets for different operating conditions; then, constructing a selective migration network based on composite weights, including a feature extractor, a state classifier, a domain identifier, and a domain adaptation module based on Wasserstein distance; using a back-propagation algorithm to pre-train the feature extractor and the state identifier; loading the pre-trained weights into the selective migration network, and completing the training using a back-propagation algorithm; and finally, obtaining the selective migration fault diagnosis result of the mechanical equipment. The present invention starts from both the state classifier and the domain identifier, obtains class-level weights and sample-level weights respectively, and uses a composite weighted model to re-weight the objective functions of the state classifier and the domain adaptation module based on Wasserstein distance, thereby improving the reliability of selective migration and effectively realizing selective migration fault diagnosis of mechanical equipment across operating conditions.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent diagnosis of mechanical equipment, and in particular relates to a method for selective migration fault diagnosis of mechanical equipment based on composite weights. Background Art

[0002] With the rapid development of modern manufacturing, mechanical equipment has become an important pillar of modern industry. However, due to the complex working environment and external interference, mechanical equipment inevitably fails. Once a failure occurs, it may cause significant economic losses and casualties. Therefore, the fault diagnosis of mechanical equipment has attracted great attention in the industrial community.

[0003] With the development of deep learning, intelligent fault diagnosis based on deep models has made great progress. However, when using supervised deep models (application number CN201911155556.7, titled: Intelligent Fault Diagnosis Method for Mechanical Equipment Based on Partially Transferred Convolutional Networks) to diagnose mechanical equipment faults, there are three shortcomings: 1) In actual industrial environments, it is very difficult to obtain sufficient labeled data, so training supervised deep models is extremely difficult; 2) Due to factors such as changes in working conditions, external interference, and equipment loss, the original data and target test data in actual applications have different data distributions; 3) The label space of the original training dataset and the target test dataset is likely to be different. In actual fault diagnosis tasks, it is inevitable that the data label categories in the target domain are smaller than those in the source domain. These shortcomings not only lead to the possibility of misdiagnosis and missed diagnosis of faults, but also may cause significant losses to modern production. Summary of the Invention

[0004] In order to overcome the shortcomings of the prior art, the purpose of the present invention is to provide a mechanical equipment selective migration fault diagnosis method based on composite weights. The training method adopted is simple and can selectively migrate the diagnostic knowledge of the source domain to the target domain to realize intelligent cross-domain diagnosis of mechanical equipment.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is:

[0006] A method for selective migration fault diagnosis of mechanical equipment based on composite weights comprises the following steps:

[0007] Step 1: Collect vibration data of mechanical equipment under different working conditions, with each working condition corresponding to a different domain;

[0008] Step 2: Build a selective transfer network based on composite weights. The selective transfer network includes a feature extractor F and trainable parameters θ F , a state classifier C and trainable parameters θ C , a domain discriminator D and trainable parameters θD , a domain adaptation module based on Wasserstein distance; the feature extractor includes a source domain feature extractor and a target domain feature extractor;

[0009] For an intelligent fault diagnosis task of a mechanical equipment, the source domain dataset is in n s Represents source domain samples, source domain sample labels, and the number of source domain samples respectively; the target domain dataset is in n t Represent the target domain samples and the number of target domain samples respectively; train the feature extractor and state classifier under the labeled source domain data, and use the cross entropy loss function L ce Reduce the empirical risk loss on the source domain and obtain a feature space for class separation. The optimization goal of the above process is expressed as:

[0010]

[0011] Among them L c is the classification loss, F(x i ) is the feature extractor extracting samples x i The characteristics of C(F(x i )) is the softmax output of the classifier, y i is the corresponding source domain sample;

[0012] The domain discriminator is used to distinguish whether the sample is from the source domain or the target domain. The samples of the source domain and the target domain are marked with domain labels d i are 1 and 0 respectively. Through training, a domain discriminator that distinguishes between the source domain and the target domain is obtained. The optimization objective of the training process is defined as follows:

[0013]

[0014] Among them L bce represents the binary cross entropy loss function;

[0015] The Wasserstein distance-based domain adaptation module is used to obtain fine-grained class-level feature alignment for the source domain distribution P s and target domain distribution P t , Wasserstein distance L w The definition is as follows:

[0016]

[0017] where the distribution P belongs to the joint distribution set Π(P s ,P t ), h s and ht Indicates the location of the source domain distribution and the location of the target domain distribution; the domain adaptation module based on Wasserstein distance has no parameters that need to be updated during model training;

[0018] For a state classifier trained on the source domain, samples in the target domain will be more likely to be classified into categories shared by the target and source domains rather than categories unique to the source domain. Therefore, the prediction results are used as class-level weights to select diagnostic knowledge for transfer. When samples from the target domain are input into the state classifier trained on the source domain, the soft label y′ of the source domain sample can be expressed as:

[0019]

[0020] Therefore, the class-level weight α is calculated as follows:

[0021]

[0022] Similarly, for the trained domain discriminator, the pseudo domain labels d′ of all samples in the source domain are expressed as:

[0023]

[0024] Therefore, the sample-level weight β i Calculated by the following formula:

[0025] β i =1-d′ i (i=1,2…,n s )

[0026] Source domain sample x i The composite weight w i Calculated by the following formula:

[0027] w i =β i ×α(y i |x i )(i=1,2,…,n s )

[0028] Among them, α(y i |x i ) is the source domain sample x i The label is y i The weight, β i is the source domain sample x i The weight of

[0029] All weights are normalized using the maximum value among the weights as follows:

[0030] w i =w i / max(w)(i=1,2,…,ns )

[0031] Among them, the weight w is the weight of all samples in training, which is the weighted value of the composite weight module CCR;

[0032] Step 3: Use the back propagation algorithm to pre-train the feature extractor and state classifier, and save the parameters of the feature extractor and state classifier;

[0033] Step 4: Load the pre-trained weights into the selective transfer network constructed in step 2. The overall training goal is:

[0034] L=L c +L d +γL w

[0035] Among them, γ is the trade-off parameter of the domain adaptation module based on Wasserstein distance;

[0036] Using the composite weight module CCR to L c and L w The weighted objective function L is obtained, which is optimized by the stochastic gradient descent Adam algorithm. The specific parameter update rules are as follows:

[0037]

[0038]

[0039]

[0040] in, and are learning rate and partial derivative respectively;

[0041] Step 5: Input the samples of the target domain into the trained feature learner and state classifier to obtain the fault diagnosis results.

[0042] The advantages of the present invention are:

[0043] (1) The present invention adopts a composite weight of class-level weight and sample-level weight to weight the shared health state, reduce the impact of the singular health state, and promote the identification of the shared health state;

[0044] (2) The present invention adopts a weighted domain adaptation module based on Wasserstein distance, which can achieve the alignment of the weighted source domain distribution with the target domain distribution, and promote the transfer of shared category diagnostic knowledge from the source domain to the target domain;

[0045] (3) The present invention uses an end-to-end deep convolutional network as the basic architecture of the present invention, forming an end-to-end fault diagnosis framework that can adaptively learn a shared feature space that is class separable and domain inseparable, thereby improving the accuracy of selective migration fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is an experimental platform for the planetary gearbox according to the embodiment of the present invention.

[0047] Figure 2 This is a diagram of the selective migration network structure of the present invention.

[0048] Figure 3 Kappa coefficient evaluation results for the planetary gearbox selective migration tasks T4 and T6 in the embodiment. DETAILED DESCRIPTION

[0049] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0050] A method for selective migration fault diagnosis of mechanical equipment based on composite weights comprises the following steps:

[0051] Step 1: Use signal acquisition equipment to collect vibration data of mechanical equipment under different working conditions. Each working condition corresponds to a different domain. The planetary gearbox experimental platform in this embodiment is as follows: Figure 1 As shown in the figure, one end of the test planetary gearbox is connected to the drive motor through the sensor, and the other end of the test planetary gearbox is connected to the load motor through the companion planetary gearbox. The experimental platform has three operating conditions, namely load 50Nm, load 150Nm and load 250Nm; the experimental gear has four health states, including normal (Nor), sun gear fault (SG), planet gear fault (PG) and ring gear fault (RG). Each health state contains 200 samples, and each sample includes 1024 data points.

[0052] Step 2: If Figure 2 As shown, a selective transfer network based on composite weights is built. The selective transfer network includes a feature extractor F and trainable parameters θ F , a state classifier C and trainable parameters θ C , a domain discriminator D and trainable parameters θ D , a domain adaptation module based on Wasserstein distance; it should be noted that the feature extractor includes a source domain feature extractor and a target domain feature extractor;

[0053] For the selective migration fault diagnosis task of the planetary gearbox, the source domain dataset is in n sRepresents source domain samples, source domain sample labels, and the number of source domain samples respectively; the target domain dataset is in n t Represent the target domain samples and the number of target domain samples respectively; train the feature extractor and state classifier under the labeled source domain data, and use the cross entropy loss function L ce To reduce the empirical risk loss in the source domain and obtain a feature space for class separation, the optimization goal of the above process can be expressed as:

[0054]

[0055] Among them L c is the classification loss, F(x i ) is the feature extractor extracting samples x i The characteristics of C(F(x i )) is the softmax output of the classifier, y i is the corresponding source domain sample;

[0056] The domain discriminator is used to distinguish whether the sample is from the source domain or the target domain. The samples of the source domain and the target domain are marked with domain labels d i are 1 and 0 respectively. Through training, a domain discriminator that can distinguish between the source domain and the target domain can be obtained. The optimization objective of the training process is defined as follows:

[0057]

[0058] Among them L bce represents the binary cross entropy loss function;

[0059] The domain adaptation module based on Wasserstein distance is used to obtain fine-grained class-level feature alignment. Wasserstein distance is widely used to measure the difference between two distributions and is very suitable for domain adaptation intelligent fault diagnosis tasks. s and target domain distribution P t , Wasserstein distance L w The definition is as follows:

[0060]

[0061] where the distribution P belongs to the joint distribution set Π(P s ,P t ), h s and h t Indicates the position of the source domain distribution and the position of the target domain distribution. It should be noted that the domain adaptation module based on Wasserstein distance has no parameters that need to be updated during model training.

[0062] For a state classifier trained on the source domain, samples in the target domain will be more likely to be classified into categories shared by the target and source domains rather than categories unique to the source domain. Therefore, the prediction results can be used as a class-level weight to select diagnostic knowledge for transfer. When samples from the target domain are input into the state classifier trained on the source domain, the soft label y′ of the source domain sample can be expressed as:

[0063]

[0064] Therefore, the class-level weight α can be calculated as follows:

[0065]

[0066] Similarly, for the trained domain discriminator, source domain samples of shared categories are difficult to distinguish, while source domain samples of non-shared categories are easily identified as coming from the source domain. Based on this, the pseudo domain label d′ of all samples in the source domain can be expressed as:

[0067]

[0068] Therefore, the sample-level weight β i It can be calculated by the following formula:

[0069] β i =1-d′ i (i=1,2…,n s )

[0070] Source domain sample x i The composite weight w i It can be calculated by the following formula:

[0071] w i =β i ×α(y i |x i )(i=1,2,…,n s )

[0072] Among them, α(y i |x i ) is the source domain sample x i The label is y i The weight, β i is the source domain sample x i The weight of

[0073] All weights are normalized using the maximum value among the weights as follows:

[0074] w i =w i / max(w)(i=1,2,…,n s )

[0075] Among them, the weight w is the weight of all samples in training, which is the weighted value of the composite weight module CCR;

[0076] Step 3: Use the back propagation algorithm to pre-train the feature extractor and state classifier, and save the parameters of the feature extractor and state classifier;

[0077] Step 4: Load the pre-trained weights into the selective transfer network constructed in step 2. The overall training goal is:

[0078] L=L c +L d +γL w

[0079] Among them, the trade-off parameter γ of the domain adaptation module based on Wasserstein distance is set to 10;

[0080] Using the composite weight module CCR to L c and L w The weighted objective function L is obtained, which is optimized by the stochastic gradient descent Adam algorithm. The specific parameter update rules are as follows:

[0081]

[0082]

[0083]

[0084] in, and are learning rate and partial derivative respectively;

[0085] Step 5: Input the samples of the target domain into the trained feature learner and state classifier to obtain the fault diagnosis results.

[0086] To verify the effectiveness of the present invention, other methods were selected for comparison: convolutional neural network (CNN), maximum mean discrepancy (MMD), domain adversarial neural network (DANN), deep convolutional transfer learning network (DCTLN), importance sampling adversarial network (IWAN), class weighted adversarial network (CWDA), weighted subdomain adaptation network (WSAN), and weighted adversarial transfer network (WATN). All selective transfer tasks for the planetary gearbox are set as shown in Table 1, totaling 6 tasks.

[0087] Table 1

[0088] source domain Target domain Source domain label Target domain label Task Name 50Nm 250Nm Nor,SG,PG,RG Nor,SG,PG T1 150Nm 50Nm Nor,SG,PG,RG Nor,SG,RG T2 150Nm 250Nm Nor,SG,PG,RG Nor,PG,RG T3 250Nm 150Nm Nor,SG,PG,RG Nor,SG,RG T4 250Nm 150Nm Nor,SG,PG,RG Nor,SG,PG T5 250Nm 150Nm Nor,SG,PG,RG Nor,RG T6

[0089] In order to eliminate the randomness of network training, each method was repeated ten times in each task to obtain the average accuracy and variance. The cross-condition fault diagnosis results of all six planetary gearbox selective migration tasks are shown in Table 2;

[0090] Table 2

[0091] Task T1 T2 T3 T4 T5 T6 CNN 71.19±2.08 91.92±1.79 86.33±1.82 75.63±2.47 92.80±1.87 72.35±3.35 MMD 80.83±2.70 69.56±2.89 80.77±4.31 70.35±3.38 70.67±4.06 68.55±2.62 DANN 90.70±1.60 93.68±2.15 88.83±2.35 80.50±3.67 84.93±2.68 79.10±1.34 DCTLN 95.93±1.31 68.06±3.87 83.93±2.82 67.85±3.17 70.27±3.65 79.40±2.21 IWAN 98.30±1.17 93.32±2.61 89.75±1.69 95.53±1.37 96.60±1.43 95.43±1.81 CWDA 96.82±1.86 92.92±1.99 90.85±1.59 95.72±2.05 95.55±0.65 95.98±2.46 WSAN 96.50±2.34 92.33±2.56 91.00±1.74 97.15±1.43 96.00±1.41 91.73±2.42 WATN 97.88±1.77 90.51±1.67 91.50±1.84 93.34±1.70 96.23±1.50 93.65±2.37 The present invention 99.58±0.51 94.55±2.61 92.48±1.93 99.72±0.31 99.15±0.96 98.93±0.72

[0092] By comparison, the fault diagnosis accuracy of the method of the present invention for various planetary gearbox selective migration tasks is significantly improved, and the accuracy of all tasks demonstrates the effectiveness of the method of the present invention. Figure 3 As shown in FIG, for the selective transfer tasks T4 and T6, the Kappa coefficient is used to further measure the classification accuracy of each method. The method of the present invention is significantly higher than the comparison method, which further illustrates the superiority of the method of the present invention.

Claims

1. A method for selective migration fault diagnosis of mechanical equipment based on composite weight, characterized in that: The following steps are involved: Step 1: Collect vibration data of mechanical equipment under different working conditions, with each working condition corresponding to a different domain; Step 2: Build a selective transfer network based on composite weights. The selective transfer network includes a feature extractor F and trainable parameters θ F , a state classifier C and trainable parameters θ C , a domain discriminator D and trainable parameters θ D , a domain adaptation module based on Wasserstein distance; the feature extractor includes a source domain feature extractor and a target domain feature extractor; For an intelligent fault diagnosis task of a mechanical equipment, the source domain dataset is in n s Represents source domain samples, source domain sample labels, and the number of source domain samples respectively; the target domain dataset is in n t Represent the target domain samples and the number of target domain samples respectively; train the feature extractor and state classifier under the labeled source domain data, and use the cross entropy loss function L ce Reduce the empirical risk loss on the source domain and obtain a feature space for class separation. The optimization goal of the above process is expressed as: Among them L c is the classification loss, F(x i ) is the feature extractor extracting samples x i The characteristics of C(F(x i )) is the softmax output of the classifier, y i is the corresponding source domain sample label; The domain discriminator is used to distinguish whether the sample is from the source domain or the target domain. The samples of the source domain and the target domain are marked with domain labels d i are 1 and 0 respectively. Through training, a domain discriminator that distinguishes between the source domain and the target domain is obtained. The optimization objective of the training process is defined as follows: Among them L bce represents the binary cross entropy loss function; The Wasserstein distance-based domain adaptation module is used to obtain fine-grained class-level feature alignment for the source domain distribution P s and target domain distribution P t , Wasserstein distance L w The definition is as follows: where the distribution P belongs to the joint distribution set Π(P s ,P t ), h s and h t Indicates the location of the source domain distribution and the location of the target domain distribution; The domain adaptation module based on Wasserstein distance has no parameters that need to be updated during model training; For a state classifier trained on the source domain, samples in the target domain will be more likely to be classified into categories shared by the target and source domains rather than categories unique to the source domain, so the prediction results are used as a class-level weight to select diagnostic knowledge for transfer; the samples in the target domain are input into the state classifier trained on the source domain, and the soft labels y of the source domain samples are ' Expressed as: Therefore, the class-level weight α is calculated as follows: Similarly, for the trained domain discriminator, the pseudo domain labels d of all samples in the source domain are ' Expressed as: Therefore, the sample-level weight β i Calculated by the following formula: β i =1-d' i ,i=1,2…,n s ; Source domain sample x i The composite weight w i Calculated by the following formula: w i =β i ×α(y i |x i ),i=1,2,…,n s ; Among them, α(y i |x i ) is the source domain sample x i The label is y i The weight, β i is the source domain sample x i The weight of All weights are normalized using the maximum value among the weights as follows: w i =w i / max(w),i=1,2,…,n s ; Among them, the weight w is the weight of all samples in training, which is the weighted value of the composite weight module CCR; Step 3: Use the back propagation algorithm to pre-train the feature extractor and state classifier, and save the parameters of the feature extractor and state classifier; Step 4: Load the pre-trained weights into the selective transfer network constructed in step 2. The overall training goal is: L=L c +L d +γL w Among them, γ is the trade-off parameter of the domain adaptation module based on Wasserstein distance; Using the composite weight module CCR to L c and L w The weighted objective function L is obtained, which is optimized by the stochastic gradient descent Adam algorithm. The specific parameter update rules are as follows: in, and are learning rate and partial derivative respectively; Step 5: Input the samples of the target domain into the trained feature learner and state classifier to obtain the fault diagnosis results.

Images