A turnout switch device mechanical fault diagnosis method based on complementary feature enhancement and consistency learning

Abstract

The application discloses a kind of based on complementary feature enhancement and consistency learning's turnout switch equipment mechanical fault diagnosis method, it is related to equipment safety evaluation technical field, comprising the following steps: step 1, the vibration signal data of turnout switch equipment is collected;Step 2, construct double branch structure, and feature extraction is carried out through double branch structure;Step 3, complementary feature enhancement module is constructed, the feature extracted is fused through complementary feature enhancement module, and the feature of fusion is obtained, and extractor E1, E2 and main diagnostic ware C are updated through loss function minimization;Step 4, based on domain adversarial learning framework, construct domain invariant feature;Step 5, auxiliary diagnostic ware Ca is constructed, and main diagnostic ware C and auxiliary diagnostic ware Ca are jointly built as double diagnostic ware, consistency learning is carried out through double diagnostic ware, and E1 and E2 are optimized again;Step 6, fault diagnosis is carried out through main diagnostic ware C and the diagnosis result is obtained.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of equipment safety evaluation, in particular to a turnout switch machine mechanical fault diagnosis method based on complementary feature enhancement and consistency learning. BACKGROUND

[0002] The turnout switch machine is an indispensable part of train operation, and its running state is directly related to the safety of train travel. Mechanical failure of the turnout switch device often leads to unstable operation of the device, and even causes serious accidents. Therefore, it is of great significance to diagnose the mechanical failure of the turnout switch device efficiently and accurately. With the development of intelligent methods based on deep learning, the research on mechanical equipment fault diagnosis has made a lot of achievements.

[0003] The rapid development of deep learning in the field of computer vision promotes the progress of intelligent fault diagnosis. Traditional deep learning methods assume that training data and test data obey the same distribution, however, in actual scenarios, when the operating conditions of the equipment change, it is easy to cause changes in data distribution, leading to a decline in the diagnostic performance of the trained diagnostic model. In addition, it is not practical to collect a large amount of fault data and perform labeling, because it is relatively difficult for the equipment to fail in natural operation. In the face of fault diagnosis under varying conditions, unsupervised domain adaptation technology shows great potential. In the absence of labeled data in the target domain, unsupervised domain adaptation transfers the knowledge learned from the source domain to the target domain, which is one of the important methods to solve the domain transfer problem.

[0004] Most existing unsupervised domain adaptation fault diagnosis methods map the features of the source domain and the target domain to a high-dimensional space by maximizing the mean difference (MMD), and then calculate the distance between the two feature distributions. By minimizing the distance between the feature distributions, the transfer from the source domain to the target domain can be achieved. Another commonly used method is to introduce an adversarial learning method to achieve domain adaptation through adversarial training. The basic idea is based on the idea of generative adversarial network (GAN), introducing a domain discriminator into the model, and through the adversarial training of the domain discriminator and the feature generator, the feature distributions of the source domain and the target domain are made as similar as possible. In this way, the transfer from the source domain to the target domain can be achieved. However, most of these methods focus on aligning the feature distributions of the source domain and the target domain, ignoring the rich fault information obtained from the original signal. In addition, in order to meet the distribution alignment, some class-related information is easily sacrificed, thereby limiting the further improvement of the diagnostic precision. Therefore, there is an urgent need for a turnout switch device mechanical fault diagnosis method based on complementary feature enhancement and consistency learning to solve the problems existing in the prior art. SUMMARY

[0005] The application aims to provide a turnout switch device mechanical fault diagnosis method based on complementary feature enhancement and consistency learning, to solve the problems of limited precision, poor robustness and poor reliability of existing diagnosis models in the prior art.

[0006] To achieve the above-mentioned purpose, the application provides a turnout switch device mechanical fault diagnosis method based on complementary feature enhancement and consistency learning, comprising the following steps:

[0007] Step 1, collecting vibration signal data of the turnout switch device; and labeling by hand, and dividing the labeled data into a source domain D s and a target domain D t ;

[0008] Step 2, constructing a double-branch structure, and extracting features through the double-branch structure, wherein the first branch focuses on capturing time sequence features through an extractor E1 and the second branch focuses on capturing frequency domain features through an extractor E2

[0009] Step 3, constructing a complementary feature enhancement module, fusing the extracted time sequence features and frequency domain features through the complementary feature enhancement module to obtain fused features f i s , and updating the extractors E1 and E2 and a main diagnostic device C through loss function minimization to capture more comprehensive fault feature information;

[0010] Step 4, constructing domain-invariant features based on a domain adversarial learning framework;

[0011] Step 5, constructing an auxiliary diagnostic device Ca, and jointly constructing a double diagnostic device with the main diagnostic device C and the auxiliary diagnostic device Ca, performing consistency learning through the double diagnostic device to optimize E1 and E2 again; through the mutual cooperation of the two diagnostic devices, the robustness of the features is enhanced, and the generalization ability of the model to fault modes under different working conditions is improved

[0012] Step 6, performing fault diagnosis through the main diagnostic device C to obtain a diagnosis result.

[0013] Preferably, the source domain D s with labels and the target domain D t without labels in step 1 are represented as follows:

[0014]

[0015] wherein, represents a source domain sample, represents a health status category label, and n srepresents the number of source domain samples, M represents the number of health status categories; represents the target domain sample, n t Represents the number of target domain samples. The source and target domains have the same label space, expressed as |Y s |=|Y t |=M.

[0016] Preferably, in step 3, the extracted time series features are enhanced by a complementary feature enhancement module. and frequency domain characteristics Fusion is performed to obtain the fused feature f i s The process is as follows:

[0017] S31. Calculate time series characteristics and frequency domain characteristics Intimacy A i , the calculation expression is as follows:

[0018]

[0019] S32. Calculate the features that E1 and E2 both focus on. The calculation expression is as follows:

[0020]

[0021] S33. Calculate the complementary features of E1 and E2 The calculation expression is as follows:

[0022]

[0023] S34. Obtain fused feature f i s , the expression is as follows:

[0024]

[0025] Where σ(·) represents Softmax, (·) T represents the matrix transpose, ψ(·), g(·), θ(·) both represent linear mapping functions, which are implemented using 1×1 convolution. FC represents a fully connected layer to integrate information. i It reflects the correlation between the features of the two branches, that is, the information that is of common concern, so 1-A i It indicates the difference between the features of the two branches, that is, the information that they do not pay attention to each other.

[0026] Preferably, through the loss function L ecThe expression for minimizing the updates to extractors E1, E2 and the main diagnostics C is as follows:

[0027]

[0028] Where CE(·) represents the cross entropy loss function.

[0029] Preferably, in step 4, based on the domain adversarial learning framework, the expression for constructing the domain-invariant feature is as follows:

[0030] L D (D)=CE(D(f i m ),d i );

[0031] L adv (E1,E2)=CE(D(f i m ),d i );

[0032] Where CE(·) represents the cross entropy loss function, f i m Represents the characteristics of the source domain and the target domain, m = {s, t}, d i Represents the domain label, when x i When coming from the source domain, d i =1, when x i When coming from the target domain i = 0, D represents the domain discriminator, L D represents the domain classification loss, L adv Denotes domain adversarial loss, and updates L alternately D and L adv , conduct adversarial training.

[0033] Preferably, in step 5, consistency learning is performed through a dual diagnostic device, and the expressions of E1 and E2 are optimized again as follows:

[0034]

[0035] In the formula, ||·|| 2 represents the MSE loss function, L con represents the consistency loss function, n b Indicates the batch size.

[0036] Preferably, the expression of the diagnosis result obtained by the main diagnostic device C during the fault diagnosis in step 6 is as follows:

[0037] Accuracy=C(f i t );

[0038] In the formula, f i t Indicates The target domain data features extracted by E1 and E2 and obtained by the complementary feature enhancement module.

[0039] Therefore, the application adopts the above-mentioned turnout switch equipment mechanical fault diagnosis method based on complementary feature enhancement and consistency learning, which has the following beneficial effects:

[0040] (1) The application uses double-branch structure for feature extraction through complementary feature learning, and fuses the features of the two branches through a complementary feature enhancement module, effectively capturing more comprehensive fault feature information and improving the accuracy of fault recognition;

[0041] (2) The application integrates double-diagnostic devices for consistency learning based on a domain adversarial learning framework, which alleviates the distribution difference between the source domain and the target domain, realizes domain-invariant feature representation, and improves the robustness of the features and the generalization ability of the model to fault patterns under different working conditions, thereby enhancing the reliability of fault diagnosis;

[0042] (3) The method of the application can be widely applied to different types of turnout switch equipment, such as gears, bearings, etc., and has good universality and adaptability; the implementation of this method can provide strong technical support for the health management of turnout switch equipment machinery, and is helpful to realize intelligent and automated maintenance of turnout switch equipment mechanical systems.

[0043] The technical solutions of the application will be further described in detail below with reference to the accompanying drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0044] Fig. 1 The application is a kind of mechanical fault diagnosis method for turnout switch equipment based on complementary feature enhancement and consistency learning, and its overall flowchart is shown in the figure.

[0045] Fig. 2 The complementary feature enhancement module of the embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0046] The detailed description of the embodiments of the application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but only represents selected embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the application.

[0047] Please refer to Figs. 1-2 A kind of mechanical fault diagnosis method for turnout switch equipment based on complementary feature enhancement and consistency learning, comprising the following steps:

[0048] Step 1, collect the vibration signal data of the turnout switching equipment; and manually label and divide the labeled data into source domain D s and target domain D t ; wherein the labeled source domain D s and the unlabeled target domain D t are represented as follows:

[0049]

[0050]

[0051] wherein, represents the source domain sample, represents the health status category label, n s represents the number of source domain samples, and M represents the number of health status categories; represents the target domain sample, n t represents the number of target domain samples, and the source domain and the target domain have the same label space, represented as |Y s |=|Y t |=M.

[0052] Step 2, construct a double-branch structure and extract features through the double-branch structure, wherein the first branch focuses on capturing time series features through extractor E1 and the second branch focuses on capturing frequency domain features through extractor E2

[0053] Step 3, construct a complementary feature enhancement module, fuse the extracted time series features and frequency domain features through the complementary feature enhancement module to obtain fused features f i s , and update the extractors E1, E2 and the main diagnostic C through loss function minimization to capture more comprehensive fault feature information; wherein the extracted time series features and frequency domain features are fused through the complementary feature enhancement module to obtain fused features f i s The process is as follows:

[0054] S31, calculate the affinity A i of the time series features and the frequency domain features , and the calculation expression is as follows:

[0055]

[0056] S32, calculate the features of common interest of E1 and E2 The calculation expression is as follows:

[0057]

[0058] S33, calculate the complementary features of E1 and E2 The calculation expression is as follows:

[0059]

[0060] S34, obtain the fused feature f i s The expression is as follows:

[0061]

[0062] In the formula, σ(·) represents Softmax, (·) T represents matrix transposition, ψ(·), g(·), θ(·) all represent linear mapping functions, realized by 1x1 convolution, FC represents a fully connected layer, and information is integrated, wherein, A i reflects the correlation between the features of the two branches, that is, the information of common interest, and 1-A i represents the difference between the features of the two branches, that is, the information of mutual interest.

[0063] By minimizing the loss function L ec The expression for updating the extractor E1, E2 and the main diagnostic C is as follows:

[0064]

[0065] In the formula, CE(·) represents a cross-entropy loss function.

[0066] Step 4, based on the domain adversarial learning framework, construct domain-invariant features; the specific expression is as follows:

[0067] L D (D)=CE(D(f i m ),d i );

[0068] L adv (E1,E2)=CE(D(f i m ),d i );

[0069] In the formula, CE(·) represents a cross-entropy loss function, f i mRepresents the characteristics of the source domain and the target domain, m = {s, t}, d i Represents the domain label, when x i When coming from the source domain, d i =1, when x i When coming from the target domain i = 0, D represents the domain discriminator, L D represents the domain classification loss, L adv Denotes domain adversarial loss, and updates L alternately D and L adv , conduct adversarial training.

[0070] Step 5: Construct the auxiliary diagnostic device Ca. Combine the main diagnostic device C and the auxiliary diagnostic device Ca into a dual diagnostic device. Perform consistency learning through the dual diagnostic device to further optimize E1 and E2. Through the mutual cooperation of the two diagnostic devices, the robustness of the features is enhanced, and the generalization ability of the model for fault modes under different working conditions is improved. Among them, the expressions of E1 and E2 optimized again through consistency learning through the dual diagnostic device are as follows:

[0071]

[0072] In the formula, ||·|| 2 represents the MSE loss function, L con represents the consistency loss function, n b Indicates the batch size, that is, Batchsize.

[0073] Step 6: Perform fault diagnosis through the main diagnostic device C to obtain the diagnosis result; the specific calculation expression is as follows:

[0074] Accuracy=C(f i t );

[0075] Where, f i t express The target domain data features are obtained by extracting features through E1 and E2 and enhancing them through the complementary feature enhancement module.

[0076] Example

[0077] To verify the effectiveness of the proposed method, experiments were conducted on the CWRU bearing dataset. The effectiveness of the proposed method was verified on multiple diagnostic tasks to improve the fault identification accuracy and diagnostic efficiency of turnout equipment.

[0078] Rolling bearings are key mechanical components in switch machines, and the CWRU dataset is provided by the University of West Reserve, and the experimental platform is operated at four different loads and speeds. The data of the experiment in this paper is obtained at a sampling frequency of 48KHZ, containing 10 kinds of health states: healthy (NC), three different degrees of inner ring fault (IF0, IF1, IF2), three different degrees of outer ring fault (OF0, OF1, OF2) and three different degrees of rolling body fault (BF0, BF1, BF2), and the detailed information of the dataset is shown in Table 1.

[0079] Table 1 Statistical table of data for experiment

[0080]

[0081]

[0082] The one-dimensional convolutional neural network CNN (Baseline), the deep domain adversarial network DANN and the correlation alignment method CORAL are compared, all the methods adopt the same network structure and are compared in the same environment. According to the method of the existing research, each experiment is carried out five times, and the average diagnostic accuracy and standard deviation are taken as the final result. Table 2 shows the accuracy (C 01 indicates the transfer task of working condition C0 as the source domain and C1 as the target domain), it can be seen that the method proposed in this paper has obvious advantages. In the most difficult task (namely C 30 ), the worst model accuracy reaches 73.34%±2.31%, while the method proposed in this application achieves an accuracy of 89.24%±2.24% in this task. The results show that the method proposed in this application is effective in the variable working condition fault diagnosis task.

[0083] Table 2 Comparison of diagnostic accuracy (%) of other methods on CWRU dataset

[0084]

[0085] Therefore, the switch machine mechanical fault diagnosis method based on complementary feature enhancement and consistency learning is adopted, first, the double branch structure is used for feature extraction, and the features of the two branches are fused through the complementary feature enhancement module, and the more comprehensive fault feature information is effectively captured, then the complementary feature enhancement module is constructed, the extracted features are fused through the complementary feature enhancement module, and the domain invariant features are constructed based on the domain adversarial learning framework; then the consistency learning is carried out through the double diagnostic device; finally, the fault diagnosis is carried out through the main diagnostic device to obtain the diagnosis result.

[0086] It should be pointed out finally that the above examples are only used to illustrate the technical solutions of the present application but not to limit it, and although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can still be modified or replaced equivalently, and these modifications or equivalent replacements should not make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. A mechanical fault diagnosis method for turnout equipment based on complementary feature enhancement and consistency learning, characterized in that: The following steps are involved: Step 1: collecting vibration signal data of the turnout equipment; The labeled data is divided into source domain D s and target domain D t ; Step 2: Construct a dual-branch structure and perform feature extraction through the dual-branch structure, where the first branch focuses on capturing time series features through extractor E1 The second branch focuses on capturing frequency domain features through extractor E2 Step 3: Construct a complementary feature enhancement module to enhance the extracted time series features. and frequency domain characteristics Fusion is performed to obtain the fused feature f i s , and update the extractors E1, E2 and the main diagnostic device C by minimizing the loss function; Step 4: Construct domain-invariant features based on the domain adversarial learning framework; Step 5: Build the auxiliary diagnostic device Ca. Combine the main diagnostic device C and the auxiliary diagnostic device Ca into a dual diagnostic device. Perform consistency learning through the dual diagnostic device to optimize E1 and E2 again. Step 6: Perform fault diagnosis through the main diagnostic device C to obtain the diagnosis result.

2. A method for diagnosing mechanical faults of turnout equipment based on complementary feature enhancement and consistency learning according to claim 1, characterized in that: The labeled source domain D in step 1 s and the unlabeled target domain D t The specific representation is as follows: in, represents the source domain sample, Indicates the health status category label, n s represents the number of source domain samples, M represents the number of health status categories; represents the target domain sample, n t Represents the number of target domain samples. The source and target domains have the same label space, expressed as |Y s |=|Y t |=M.

3. The method for mechanical fault diagnosis of turnout equipment based on complementary feature enhancement and consistency learning according to claim 2 is characterized in that: In step 3, the extracted time series features are enhanced by the complementary feature enhancement module and frequency domain characteristics Fusion is performed to obtain the fused feature f i s The process is as follows: S31. Calculate time series characteristics and frequency domain characteristics Intimacy A i , the calculation expression is as follows: S32. Calculate the features that E1 and E2 focus on together The calculation expression is as follows: S33. Calculate the complementary features of E1 and E2 The calculation expression is as follows: S34. Obtain fused feature f i s , the expression is as follows: Where σ(·) represents Softmax, (·) T represents the matrix transpose, ψ(·), g(·) and θ(·) both represent linear mapping functions, which are implemented using 1×1 convolution. FC represents a fully connected layer to integrate information.

4. A method for diagnosing mechanical faults of turnout equipment based on complementary feature enhancement and consistency learning according to claim 3, characterized in that: Through the loss function L ec The expression for minimizing the updates to extractors E1, E2 and the main diagnostics C is as follows: L ec (E1,E2,C)=CE(C(f i s ),y i s ); Where CE(·) represents the cross entropy loss function.

5. The method for mechanical fault diagnosis of turnout equipment based on complementary feature enhancement and consistency learning according to claim 4 is characterized in that: In step 4, based on the domain adversarial learning framework, the expression for constructing domain-invariant features is as follows: L D (D)=CE(D(f i m ),d i ); L adv (E1,E2)=CE(D(f i m ),d i ); Where CE(·) represents the cross entropy loss function, f i m Represents the characteristics of the source domain and the target domain, m = {s, t}, d i Represents the domain label, when x i When coming from the source domain, d i =1, when x i When coming from the target domain i = 0, D represents the domain discriminator, L D represents the domain classification loss, L adv Denotes domain adversarial loss, and updates L alternately D and L adv , conduct adversarial training.

6. The method for mechanical fault diagnosis of turnout equipment based on complementary feature enhancement and consistency learning according to claim 5, characterized in that: In step 5, consistency learning is performed through the dual diagnostics, and the expressions of E1 and E2 are optimized again as follows: In the formula, ||·|| 2 represents the MSE loss function, L con represents the consistency loss function, n b Indicates the batch size.

7. A method for diagnosing mechanical faults of turnout equipment based on complementary feature enhancement and consistency learning according to claim 6, characterized in that: In step 6, the diagnostic result obtained by the main diagnostic device C is expressed as follows: Accuracy=C(f i t ); Where, f i t express The target domain data features are obtained by extracting features through E1 and E2 and enhancing them through the complementary feature enhancement module.

Images