Migration diagnosis method based on integrated weighted subfield adaptive network

Abstract

The present application relates to a kind of based on integrated weighting subdomain self-adapting network migration diagnosis method, belong to mechanical equipment fault diagnosis technical field.The method includes: S1: the original vibration signal on mechanical equipment is collected by sensor;Then, the original vibration signal collected is expanded sample using sliding window sampling technique;S2: EWSAN is constructed;S3: training sample is input into the EWSAN constructed, using the classification loss function L c Of source domain labeled sample, the EJDA domain adaptation loss function L EJDA Between source domain and unlabeled target domain, the edge distribution alignment loss function L MDA Between source domain and unlabeled target domain and the information entropy loss function L IE Of unlabeled target domain is used to EWSAN iterative update training;After multiple iterations training, error curve tends to be stable, EWSAN training is completed, and is used for the closed set and partial set migration diagnosis of bearing across.This application can improve the accuracy of migration diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of mechanical equipment fault diagnosis technology, and relates to a migration diagnosis method based on an integrated weighted subdomain adaptive network. Background Technology

[0002] Rotating machinery is widely used in important engineering fields such as energy and power, rail transportation, military defense, and aerospace. However, some key mechanical components, such as bearings, gears, and shafts, often operate under harsh conditions of high speed, heavy load, and instability. Once a failure occurs, it can cause irreparable economic and property losses or even casualties. Due to the scarcity of fault samples and the lack of prior knowledge of labels, transfer fault diagnosis methods based on domain adaptation technology have been extensively studied by experts and scholars. However, most of them only focus on the overall distribution alignment of the source and target domains, while ignoring the more refined conditional distribution alignment at the class level. On the other hand, the few joint distribution alignment mechanisms that consider class-level alignment have approximate estimations. At the same time, most methods can only be applied to one of the following cases: closed-set transfer diagnosis (source and target domains have the same category) and partial transfer diagnosis (target domain category is a subset of source domain).

[0003] Therefore, there is an urgent need for a device fault diagnosis method that can be applied to both closed-set migration and partial-set migration. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a migration diagnosis method based on an ensembleweighting subdomain adaptation network (EWSAN), which can be applied to both closed-set migration and partial-set migration, and can improve the accuracy of migration diagnosis.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A migration diagnostic method based on an integrated weighted subdomain adaptive network specifically includes the following steps:

[0007] S1: Raw vibration signals from mechanical equipment are collected using sensors to facilitate the construction of subsequent migration diagnostic tasks. Then, the collected raw vibration signals are expanded using sliding window sampling technology, and the expanded samples are divided into training samples and test samples.

[0008] S2: Construct an integrated weighted subdomain adaptive network (EWSAN) based on a one-dimensional convolutional neural network.

[0009] S3: Input the training samples into the constructed EWSAN, and use the classification loss function L for labeled samples in the source domain. cEnhanced Joint Distribution Alignment (EJDA) domain adaptation loss function L between the source domain and the unlabeled target domain. EJDA The edge distribution alignment loss function L between the source domain and the unlabeled target domain MDA and the unlabeled target domain information entropy loss function L IE EWSAN is trained iteratively; after multiple iterations, the error curve stabilizes, and EWSAN training is complete.

[0010] S4: Input the test samples into the trained EWSAN for cross-bearing closed set and partial set transfer diagnosis.

[0011] Furthermore, in step S2, the constructed EWSAN uses a one-dimensional convolutional neural network as the backbone network to adaptively extract fault features; the EWSAN includes a feature extractor and a multi-scale top-level classifier; the feature extractor consists of three convolutional blocks; the multi-scale top-level classifier is an integrated structure composed of multiple classifiers with different network parameters and normalization layers, each single classifier consists of a convolutional block (same as the feature extractor structure) and two fully connected layers; wherein, each convolutional block contains a convolutional layer, a max pooling layer and a batch normalization layer.

[0012] Furthermore, in step S2, the total loss function L of EWSAN training is iteratively updated. W The expression is:

[0013] L W =L C +λ(L IE +L EJDA )+(1-λ)L MDA

[0014] in, ρ is the ratio of the current number of network iterations to the total number of iterations.

[0015] Furthermore, in step S2, the RMSprop optimizer and loss function L are used. EJDA The feature extractor G is optimized according to the following formula. FE (·) and multi-scale top-level classifier G MTC The parameter θ of (·) FE and θ MTC ;

[0016]

[0017]

[0018] Where ε is the learning rate.

[0019] Furthermore, in step S2, the EJDA domain adaptation loss function L between the source domain and the unlabeled target domain... EJDA The expression is:

[0020]

[0021] Where m represents the number of kernel functions, and K represents the number of classifiers in the multi-scale top-level classifier. It is the number of categories; and Let represent the number of samples in the source and target domains of class c, respectively. w S (c) and w T (c) represents the class weights of the source and target domains of class c, respectively; φ p (·) represents the p-th mapping function, k p (·) represents the p-th kernel function; and Let represent the output vectors of the i-th source domain sample and the target domain sample on the k-th classifier in the c-th class, respectively.

[0022] Furthermore, in step S2, the class weights w of the source and target domains of class c are... S (c) and w T The expression for (c) is:

[0023]

[0024] in, and Let represent the probability vectors of the i-th sample output by the Softmax classifier of the k-th classifier from the source and target domains, respectively; j represents the index position of the element in the vector. and Represents the i-th source domain sample and target domain samples The element corresponding to the tag index position;

[0025]

[0026] For source and target domain sample pairs and probability vector pairs The expression is:

[0027]

[0028] Where, θ FE and θ MTC These represent the feature extractor G. FE (·) and multi-scale top-level classifier G MTC The network parameters are in parentheses; to avoid class weights wS (c) and w T The values ​​of (c) are too small, so they are normalized to:

[0029]

[0030] Furthermore, in step S2, the classification loss function L for labeled samples in the source domain... c The expression is:

[0031]

[0032] Where I(·) is an indicator function, when When I = 1, otherwise, when When I = 0;

[0033] The edge distribution alignment loss function L between the source domain and the unlabeled target domain MDA The expression is:

[0034]

[0035] in, Denotes the Frobenius norm in the RKHS space;

[0036] Unlabeled target domain information entropy loss function L IE The expression is:

[0037] The beneficial effects of this invention are that the EWSAN constructed by this invention can be applied to both closed-set migration and partial-set migration. In EWSAN, an enhanced joint distributed alignment mechanism (EJDA) and an integrated weighted maximum mean difference index (EWMMD) are used to improve the accuracy of migration diagnosis.

[0038] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0039] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0040] Figure 1 This is a diagram of the Integrated Weighted Subdomain Adaptive Network (EWSAN) structure constructed in this invention;

[0041] Figure 2 This is a schematic diagram of the CWRU test bench structure;

[0042] Figure 3 This is a schematic diagram of the RDS test bench structure;

[0043] Figure 4 This is a schematic diagram of the IMS test bench structure. Detailed Implementation

[0044] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0045] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0046] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0047] Please see Figures 1-4 This invention provides a cross-bearing migration diagnostic method based on an Integrated Weighted Subdomain Adaptive Network (EWSAN), the specific steps of which are as follows:

[0048] Step 1: Acquire raw vibration signals of the bearings on the mechanical equipment using an accelerometer to facilitate the construction of subsequent migration diagnostic tasks. Then, expand the sample using sliding window sampling technology.

[0049] Step 2: Construct the Ensemble Weighted Subdomain Adaptive Network (EWSAN) based on a one-dimensional convolutional neural network, such as... Figure 1 As shown.

[0050] like Figure 1 As shown, EWSAN uses a one-dimensional convolutional neural network as its backbone to adaptively extract fault features. This network structure consists of a feature extractor and a multi-scale top-level classifier. The feature extractor comprises three convolutional blocks, each containing a convolutional layer, a max-pooling layer, and a batch normalization layer. The multi-scale top-level classifier is an ensemble structure composed of multiple classifiers with different network parameters and normalization layers. Each classifier consists of a convolutional block (similar to the feature extractor structure) and two fully connected layers. The following sections will detail the EJDA mechanism in the ensemble weighted subdomain adaptive network, the EWMMD distribution difference index, and the optimization process of EWSAN.

[0051] 1) Enhanced Joint Distributed Alignment (EJDA) Mechanism

[0052] First, the general Joint Distribution Alignment (JDA) mechanism is shown in the following equation:

[0053]

[0054] in, It represents the number of categories, T represents a feature representation, and X... S and Y S X represents the source domain sample and its corresponding label, respectively. T and Y T Let P(·) represent the target domain sample and its corresponding label, respectively, and let P(·) represent the corresponding distribution. Furthermore, from formula (1), we see that the JDA mechanism contains two approximate estimates:

[0055] ① This joint distribution P(X,Y) is approximately equal to the sum of the conditional distribution P(Y|X) and the marginal distribution P(X):

[0056] P(X,Y)=P(X)+P(Y|X) (2)

[0057] It is clear that this is only a necessary but not sufficient condition for a joint distribution.

[0058] ② The conditional distribution P(Y|X) is approximately equal to the class-conditional distribution P(Y|X):

[0059] P(Y|X)=P(X|Y) (3)

[0060] Based on Bayes' theorem P(X,Y)=P(X|Y)P(Y), a theoretically sound EJDA mechanism is proposed:

[0061]

[0062] 2) Ensemble Weighting Maximum Mean Discrepancy (EWMMD)

[0063] To enhance EWSAN's domain obfuscation capability, a distribution difference index, EWMMD, which includes sample weight information, was constructed:

[0064]

[0065] in, and Let w represent the number of samples in the source and target domains of class c, respectively; S (c) and w T (c) represents the class weights of the source and target domains of class c, respectively; φ p (·) represents the p-th kernel function. Assuming this multi-scale top-level classifier contains K classifiers, some variables in equation (5) can be expressed as:

[0066]

[0067]

[0068] in, and Let represent the probability vectors of the i-th sample output by the Softmax function of the k-th classifier from the source and target domains, respectively. and Represents the i-th source domain sample and target domain samples The element corresponding to the tag index position:

[0069]

[0070] For source and target domain sample pairs and probability vector pairs The following formula can be obtained:

[0071]

[0072] Where, θ FE and θ MTC Let G be the feature extractor. FE (·) and multi-scale top-level classifier G MTC The network parameters are (·). To avoid class weights (w) S (c) and w T The values ​​of (c) are too small, so they are normalized to:

[0073]

[0074] 3) Optimization process

[0075] The Softmax loss function is used to minimize the classification error of source domain samples, thereby ensuring excellent diagnostic capabilities of EWSAN:

[0076]

[0077] Where I(·) is an indicator function, when When I = 1, otherwise, when When I = 0.

[0078] To reduce computational complexity, the EJDA mechanism is placed on the last fully connected layer of the multi-scale top-level classifier. By combining the EWMMD metric with the EJDA mechanism, equation (4) can be expressed as:

[0079]

[0080] in, and Let represent the output vectors of the i-th source domain sample and the target domain sample on the k-th classifier in the c-th class, respectively. This class prior knowledge can be computed:

[0081]

[0082] Therefore, the final EJDA mechanism used to optimize EWSAN is shown in the following equation:

[0083]

[0084] Furthermore, this embodiment uses an edge distribution alignment loss function L. MDA and information entropy loss function L IE To enhance the reliability of obtaining pseudo-labels (i.e., the probability vector output from the target domain in formula (14)), these loss functions are expressed as follows:

[0085]

[0086]

[0087] Integrating the above four loss functions (L C ,L EJDA ,L MDA ,L IE The final loss function can be expressed as:

[0088] L W =L C +λ(L IE +LEJDA )+(1-λ)L MDA (17)

[0089] in, ρ is the ratio of the current number of network iterations to the total number of iterations.

[0090] Using the RMSprop optimizer and equation (17), the network parameters (θ) FE and θ MTC It can be optimized by the following formula:

[0091]

[0092]

[0093] Where ε is the learning rate.

[0094] Step 3: Input the segmented training samples into the constructed Ensemble Weighted Subdomain Adaptive Network (EWSAN), and use the classification loss function L for labeled samples in the source domain. c (Equation 11) The EJDA domain adaptation loss function L between the source domain and the unlabeled target domain EJDA (Equation 14) The edge distribution alignment loss L between the source domain and the unlabeled target domain MDA and the unlabeled target domain information entropy loss function L IE The transfer diagnostic network is iteratively updated and trained (Equations 18 and 19).

[0095] Step 4: After multiple iterations of training, the error curve tends to stabilize, the model training is complete, and the trained transfer diagnostic network will be used for closed set and partial set transfer diagnostics across bearings.

[0096] The effectiveness of the intelligent diagnostic method of the present invention will be demonstrated through comparative experiments below.

[0097] Verification Experiment: The raw bearing vibration signals collected in this experiment came from three test benches, specifically the CWRU standard bearing dataset publicly available from Case Western Reserve University. A schematic diagram of the CWRU bearing dataset is shown below. Figure 2 As shown, it consists of a motor, bearings at both ends of the motor, a torque sensor, and a power meter; and a self-testing RDS bearing dataset. A schematic diagram of the RDS test bench is shown below. Figure 3 As shown, it consists of a servo motor, a coupling, a rotor, and bearings and bearing housings at both ends. The IMS bearing dataset from the University of Cincinnati is also included, and a schematic diagram of the test bench is shown below. Figure 4As shown, it consists of a motor, support bearings, faulty bearings used for testing, and a loading system. Information such as fault type, fault size, and speed for each bearing dataset is shown in Table 1, where NC, IF, BF, and OF represent normal state, inner ring fault, rolling element fault, and outer ring fault, respectively. Using these three bearing datasets, we can establish six cross-bearing closed-set migration diagnostic tasks (A→B, B→A, A→C, C→A, C→B, and B→C) and six cross-bearing partial-set migration diagnostic tasks (P1, P2, P3, P4, P5, and P6). Detailed information for the six cross-bearing partial-set migration diagnostic tasks is shown in Table 2.

[0098] Table 1. Detailed information for the three bearing datasets.

[0099]

[0100] Table 2 Six Cross-Bearing Partial Set Migration Diagnostic Tasks

[0101]

[0102] Comparative experiment:

[0103] To demonstrate the superiority of the transfer diagnostic method based on the Integrated Weighted Subdomain Adaptive Network (EWSAN) of this invention, experimental results for six cross-bearing closed-set transfer diagnostic tasks and six cross-bearing partial-set transfer diagnostic tasks are compared with current typical edge distribution alignment network models (DDC, DCORAL, DANN, MCD), classic joint distribution alignment (JDA, DJDA), and well-known partial-set transfer diagnostic models (PADA, SAN). Tables 3 and 4 show that the proposed Integrated Weighted Subdomain Adaptive Network has higher transfer diagnostic accuracy and stronger generalization ability (bold indicates the highest diagnostic accuracy in each transfer diagnostic task). Specifically, EWSAN achieves an average diagnostic accuracy of over 90% on the six cross-bearing closed-set transfer diagnostic task and over 80% on the six cross-bearing partial-set transfer diagnostic task; compared with other transfer diagnostic methods, it improves the overall diagnostic accuracy by approximately 5%.

[0104] Table 3. Experimental results of six cross-bearing closed set migration diagnostic tasks.

[0105]

[0106] Table 4. Experimental results of six cross-bearing partial set migration diagnostic tasks.

[0107]

[0108] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A migration diagnostic method based on an integrated weighted subdomain adaptive network, characterized in that, The method specifically includes the following steps: S1: Collect raw vibration signals from mechanical equipment using sensors; then, expand the samples using sliding window sampling technology and divide the expanded samples into training samples and test samples. S2: Construct an Integrated Weighted Subdomain Adaptive Network, abbreviated as EWSAN; the constructed EWSAN uses a one-dimensional convolutional neural network as the backbone network to adaptively extract fault features; the EWSAN includes a feature extractor and a multi-scale top-level classifier; the feature extractor consists of three convolutional blocks; the multi-scale top-level classifier is an integrated structure composed of multiple classifiers with different network parameters and normalization layers, each single classifier consists of a convolutional block and two fully connected layers; wherein, each convolutional block contains a convolutional layer, a max pooling layer and a batch normalization layer; S3: Input the training samples into the constructed EWSAN and use the classification loss function of labeled samples in the source domain. EJDA domain adaptation loss function between the source domain and the unlabeled target domain The edge distribution alignment loss function between the source domain and the unlabeled target domain and the unlabeled target domain information entropy loss function The EWSAN is iteratively updated and trained; after multiple iterations, the error curve tends to stabilize, and the EWSAN training is completed; where EJDA represents enhanced joint distribution alignment. EJDA domain adaptation loss function between the source domain and the unlabeled target domain The expression is: in, m Indicates the number of kernel functions. K This indicates the number of classifiers within the top-level classifier of a multi-scale system. It is the number of categories; and They represent the first c The number of samples in the source and target domains of the class. ; and They represent the first c Class weights of the source and target domains of a class; Indicates the first p A mapping function, Indicates the first p One kernel function; and They represent the first i The source domain samples and target domain samples in the th... k On the classifier at the th th , c Output vector on the class; Classification loss function for labeled samples in the source domain The expression is: in, It is an indicator function, when hour Conversely, when hour ; Edge distribution alignment loss function between source domain and unlabeled target domain The expression is: in, Denotes the Frobenius norm in the RKHS space; Unlabeled target domain information entropy loss function The expression is: S4: Input the test samples into the trained EWSAN for cross-bearing closed set and partial set transfer diagnosis.

2. The migration diagnosis method according to claim 1, characterized in that, In step S3, the total loss function of EWSAN training is iteratively updated. The expression is: in, , This is the ratio of the current number of network iterations to the total number of iterations.

3. The migration diagnosis method according to claim 2, characterized in that, In step S3, the RMSprop optimizer and loss function are used. The feature extractor is optimized according to the following formula. and multi-scale top-level classifier parameters and ; in, This refers to the learning rate.

4. The migration diagnosis method according to claim 1, characterized in that, In step S3, the first c Class weights of the source and target domains of a class and The expression is: in, and They respectively represent the source domain and the target domain at the th... k The output of Softmax on the classifier is the first i The probability vector of each sample; j Indicates the index position of the element in the vector; and Indicates the first i Source domain samples and target domain samples The element corresponding to the tag index position; For source and target domain sample pairs ( , ) and probability vector pairs ( , The expression is: in, and These represent feature extractors. and multi-scale top-level classifier Network parameters; to avoid class weights and The values ​​are too small, so they are normalized to: 。

Images