A fault diagnosis method and system based on partial domain adaptation of PDA-BLS model

Abstract

The present invention discloses a fault diagnosis method and system for partial domain adaptation based on the PDA-BLS model. First, the labels of target domain samples are predicted by a probabilistic labeling method to obtain the weights of source domain samples. Then, within the BLS framework, the weighted maximum mean difference of source domain samples and the weighted classification loss of source domain samples are integrated, and the selection bias of source domain shared samples is corrected by a weighted strategy, and source domain samples that are not related to the target domain fault category are filtered out. Only the knowledge of source domain samples related to the target domain fault category is transferred, thereby suppressing the negative transfer caused by source domain private category samples in the domain adaptation process. In addition, the target domain sample classification loss method is used to further promote positive transfer, so as to learn the BLS classifier, use the learned BLS classifier to realize the diagnosis of target domain samples, and recalculate the weights of source domain samples based on the diagnosis results. The BLS classifier is iteratively updated repeatedly to realize cross-domain fault diagnosis under partial domain adaptation.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a fault diagnosis method and system based on partial domain adaptation of a PDA-BLS model. Background Art

[0002] In the field of mechanical fault diagnosis, affected by factors such as speed or load, variable operating conditions of mechanical equipment are common, and the vibration signals collected under different operating conditions often have different nonlinear characteristics, which makes the fault characteristics between different operating conditions have large data distribution differences. In addition, some fault categories of the working conditions in the source domain may not appear in the working conditions in the target domain. There is a situation where the fault category space of the target domain is a subset of the fault category space of the source domain, which leads to the problem of partial domain adaptation fault diagnosis. The source domain and target domain under partial domain adaptation not only have different data distributions, but also have inconsistent label spaces. Influenced by the private categories of the source domain, the samples of the target domain may be misclassified as private categories of the source domain. Directly aligning the feature distribution between the source domain and the target domain may lead to serious negative transfer, resulting in performance degradation of the domain adaptation method. Summary of the invention

[0003] The present invention provides a fault diagnosis method and system for partial domain adaptation based on a PDA-BLS model, thereby effectively improving the cross-domain fault diagnosis performance under partial domain adaptation.

[0004] The present invention provides a fault diagnosis method based on partial domain adaptation of a PDA-BLS model, comprising:

[0005] Based on source domain samples And the probability labeling method, predict the target domain samples The initial label of s is the number of source domain samples, n t is the number of samples in the target domain;

[0006] Based on the initial label through the formula Calculate the weight of the qth sample in the source domain category c when the PDA-BLS model is iterated to the iterth time in, The number of samples in the target domain predicted as category c when the PDA-BLS model iterates to iter-1th time;

[0007] By formula Calculate the source domain sample weight vector ω of the iter-th iteration of the PDA-BLS model iter ;

[0008] By formula Calculate the element (V0) in the i-th row and j-th column of the marginal distribution SWMMD matrix V0 ij ;

[0009] By formula Calculate the conditional distribution SWMMD matrix V c The element in the i-th row and j-th column of c ) ij ;

[0010] By formula Calculate the SWMMD matrix V;

[0011] The input data X = [X s ,X t ] is mapped to the feature layer of the PDA-BLS model to form a feature node Then generate the enhanced node H through the nonlinear transformation of the activation function m =[H1,...,H m ], get the actual input signal in, α(x i ) T is the i-th sample x of the input data X i Output after feature layer transformation and enhancement layer transformation;

[0012] By formula Calculate the weighted maximum mean difference term of the source domain samples; where β is the output weight matrix;

[0013] By formula The classification loss of the PDA-BLS model is calculated; where is the diagonal matrix of source domain sample weights for the iterth iteration of the PDA-BLS model, is the expected output of the i-th sample in the source domain, is the predicted output of the jth sample in the target domain, δ∈[0,1] is the classification loss balance factor;

[0014] Apply L2 regularization to the output weight matrix β and construct the objective function of the PDA-BLS model Among them, θ is the regular term coefficient, γ∈[0,1] is the balance factor between data distribution difference and classification loss;

[0015] Solve the objective function and obtain the output weight matrix β;

[0016] When the number of training samples n s +n tWhen the number of nodes is greater than or equal to nk+m of the PDA-BLS model, the output weight matrix β of the PDA-BLS model is solved based on the left pseudo-inverse algorithm;

[0017] When the number of training samples n s +n t When the number of nodes is less than nk+m of the PDA-BLS model, the output weight matrix β of the PDA-BLS model is solved based on the right pseudo-inverse algorithm;

[0018] Based on the solved output weight matrix β, the formula f(x) = α(x) β = [f1(x), f2(x), ..., f C (x)] to obtain the PDA-BLS model output of the target domain sample x with unknown labels;

[0019] By the formula Label(x) = argmaxf i (x),i=1,...,C obtains the predicted label of the target domain sample x with unknown label.

[0020] Specifically, the output weight matrix β of the PDA-BLS model is solved based on the left pseudo-inverse algorithm, including:

[0021] By formula Calculate the output weight matrix β; where Ι nk+m It is the identity matrix of (nk+m)×(nk+m).

[0022] Specifically, the output weight matrix β of the PDA-BLS model is solved based on the right pseudo-inverse algorithm, including:

[0023] By formula Calculate the output weight matrix β; where Y s is the label matrix of the source domain samples, is the label matrix of the predicted target domain samples.

[0024] The present invention also provides a partial domain adaptation fault diagnosis system based on the PDA-BLS model, comprising:

[0025] Initial label prediction module, used to predict labels based on source domain samples And the probability labeling method, predict the target domain samples The initial label of s is the number of source domain samples, n t is the number of samples in the target domain;

[0026] The weight acquisition module is used to obtain the weight based on the initial label through the formula Calculate the weight of the qth sample in the source domain category c when the PDA-BLS model is iterated to the iterth time in, The number of samples in the target domain predicted as category c when the PDA-BLS model iterates to iter-1th time;

[0027] The weight vector calculation module is used to calculate the weight vector through the formula Calculate the source domain sample weight vector ω of the iter-th iteration of the PDA-BLS model iter ;

[0028] Edge distribution matrix operation module, used to pass the formula Calculate the element (V0) in the i-th row and j-th column of the marginal distribution SWMMD matrix V0 ij ;

[0029] Conditional distribution matrix operation module, used to calculate the matrix through the formula Calculate the conditional distribution SWMMD matrix V c The element in the i-th row and j-th column of c ) ij ;

[0030] SWMMD matrix operation module is used to calculate the matrix of Calculate the SWMMD matrix V;

[0031] The actual input signal acquisition module is used to convert the input data X=[X s ,X t ] is mapped to the feature layer of the PDA-BLS model to form a feature node Then generate the enhanced node H through the nonlinear transformation of the activation function m =[H1,...,H m ], get the actual input signal in, α(x i ) T is the i-th sample x of the input data X i Output after feature layer transformation and enhancement layer transformation;

[0032] The weighted maximum mean difference calculation module is used to calculate the weighted maximum mean difference through the formula Calculate the weighted maximum mean difference term of the source domain samples; where β is the output weight matrix;

[0033] The classification loss calculation module is used to calculate the loss through the formula The classification loss of the PDA-BLS model is calculated; where is the diagonal matrix of source domain sample weights for the iterth iteration of the PDA-BLS model, is the expected output of the i-th sample in the source domain, is the predicted output of the jth sample in the target domain, δ∈[0,1] is the classification loss balance factor;

[0034] Objective function construction module, used to apply L2 regularization to the output weight matrix β and construct the objective function of the PDA-BLS model

[0035] Among them, θ is the regular term coefficient, γ∈[0,1] is the balance factor between data distribution difference and classification loss;

[0036] The objective function solving module is used to solve the objective function and obtain the output weight matrix β;

[0037] Output weight matrix first solution module, used when the number of training samples is n s +n t When the number of nodes is greater than or equal to nk+m of the PDA-BLS model, the output weight matrix β of the PDA-BLS model is solved based on the left pseudo-inverse algorithm;

[0038] Output weight matrix second solution module, used when the number of training samples is n s +n t When the number of nodes is less than nk+m of the PDA-BLS model, the output weight matrix β of the PDA-BLS model is solved based on the right pseudo-inverse algorithm;

[0039] The model output module is used to calculate the output weight matrix β through the formula f(x)=α(x)β=[f1(x),f2(x),...,f C (x)] to obtain the PDA-BLS model output of the target domain sample x with unknown labels;

[0040] Fault label prediction module, used to predict the fault label using the formula Label(x) = argmaxf i (x),i=1,...,C obtains the predicted label of the target domain sample x with unknown label.

[0041] Specifically, the first solution module of the output weight matrix is ​​used when the number of training samples is n s +n t When the number of nodes is greater than or equal to nk+m of the PDA-BLS model, the formula Calculate the output weight matrix β; where Ι nk+m It is the identity matrix of (nk+m)×(nk+m).

[0042] Specifically, the second solution module of the output weight matrix is ​​used when the number of training samples is n s +n t When the number of nodes is less than nk+m of the PDA-BLS model, the formula Calculate the output weight matrix β; where Y s is the label matrix of the source domain samples, is the label matrix of the predicted target domain samples.

[0043] One or more technical solutions provided in the present invention have at least the following technical effects or advantages:

[0044] First, the labels of the target domain samples are predicted by the probability labeling method to obtain the weights of the source domain samples. Then, in the BLS framework, the maximum mean difference of the weighted source domain samples and the weighted classification loss of the source domain samples are integrated. The selection bias of the source domain shared samples is corrected by the weighting strategy, and the source domain samples that are not related to the target domain fault category are filtered out. Only the knowledge of the source domain samples related to the target domain fault category is transferred, thereby suppressing the negative transfer caused by the source domain private category samples in the domain adaptation process. In addition, the target domain sample classification loss method is used to further promote positive transfer, so as to learn the BLS classifier, use the learned BLS classifier to realize the diagnosis of the target domain samples, and recalculate the weights of the source domain samples according to the diagnosis results. In this way, the BLS classifier is repeatedly iterated and updated to realize cross-domain fault diagnosis under partial domain adaptation. The present invention makes full use of the sample similarity of the shared category, while overcoming the problems of domain data distribution differences and label inconsistency, and effectively improves the cross-domain fault diagnosis performance under partial domain adaptation. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A flowchart of a fault diagnosis method based on partial domain adaptation of a PDA-BLS model provided in an embodiment of the present invention;

[0046] Figure 2 A module diagram of a fault diagnosis system based on partial domain adaptation of a PDA-BLS model provided in an embodiment of the present invention;

[0047] Figure 3 This is a site photo of the SpectraQuest power transmission fault diagnosis experimental platform in an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The embodiment of the present invention provides a fault diagnosis method and system based on partial domain adaptation of the PDA-BLS model, thereby effectively improving the cross-domain fault diagnosis performance under partial domain adaptation.

[0049] The specific steps of the PDA-BLS cross-domain fault diagnosis method provided by the embodiment of the present invention are as follows:

[0050] Step 1: For the vibration signals collected under different working conditions and different units, extract multi-statistics and multi-scale fuzzy measurement entropy as fault features, and obtain labeled source domain data and unlabeled target domain data;

[0051] Step 2: Initialize the parameters of the PDA-BLS model, including the maximum number of iterations T, the number of feature nodes, the number of enhanced nodes, the regularization parameter θ, and the balance factors γ, δ;

[0052] Step 3: Combined with the source domain samples, an auxiliary classifier based on probability annotation is used to predict the initial labels of each sample in the target domain. Fault samples with higher probability in the target domain are given larger weights, and fault samples with lower probability are given smaller weights, and the initial weighted vector of the source domain samples is constructed.

[0053] Step 4: Randomly initialize the feature layer weight parameter W ei and β ei , enhancement layer weight parameter W hj and β hj , based on the source domain and target domain data, the feature layer output of the PDA-BLS model is constructed, and the feature layer output data is used as the enhancement layer input data to generate the enhancement layer output data under the nonlinear transformation of the activation function.

[0054] Step 5: Merge the output data of the feature layer and the enhancement layer, and project the merged data through the mapping function. Combined with the weighted vector of the source domain fault samples, the maximum mean difference term (i.e., SWMMD term) of the source domain samples weights is constructed to achieve data distribution alignment between the source domain and the target domain.

[0055] Specifically, in order to reduce the negative transfer caused by source domain private category samples during the domain adaptation process, by increasing the weight of source domain samples corresponding to the fault categories with higher probability in the target domain and reducing the weight of source domain samples corresponding to the fault categories with lower probability in the target domain, the alignment degree of source domain and target domain data is improved, the negative transfer caused by source domain private category samples is suppressed, and the data distribution difference between the source domain and the target domain is reduced. Specifically, the weighted marginal distribution difference of source domain samples is calculated as follows:

[0056]

[0057] Where V0 is the marginal distribution SWMMD matrix, (V0) ij Represents the element in row i and column j of V0, calculated as follows:

[0058]

[0059] The weighted conditional distribution difference of source domain samples is calculated as follows:

[0060]

[0061] Among them, V c is the conditional distribution SWMMD matrix, (V c )ij V c The element in row i and column j is calculated as follows:

[0062]

[0063] in, is the data set belonging to fault category c in the source domain. i ) is the sample x i The real label, For the dataset The number of samples in . Accordingly, is the dataset belonging to fault category c in the target domain, For sample x j The predicted label of It indicates the number of samples in the target domain that are predicted to be of category c when the model is iterated to iter-1th time. Represents the weight of the i-th sample in the source domain when the model is iterated to the iterth time. By combining the source domain samples, the initial labels of the target domain samples are obtained using the probability labeling method, and the number of samples of each category in the initial label is counted Then, the initial weight vector ω1 of the source domain samples is calculated by formula (5) and formula (6). Similarly, the predicted labels of the target domain samples when the PDA-BLS model is iterated to the iter-1th time, and the number of samples of each category in the target domain is counted. Similarly, the source domain sample weight vector ω of the iter-th iteration is calculated by formula (5) and formula (6): iter (iter∈[2,T]), and the source domain data weight vector is continuously updated in this way.

[0064]

[0065]

[0066] In the formula, represents the weight of the qth sample in the source domain category c when the model iterates to the iterth time. The value indicates that the probability that the qth sample in the source domain category c appears in the target domain is high, which can also be regarded as the probability that the fault category of the sample belongs to the shared category. By applying weight ω to the source domain samples iter , to a certain extent, it can identify and distinguish the shared categories in the source domain and the private categories in the source domain, and indirectly filter out the samples of the private categories in the source domain. In particular, considering that normal state samples in the target domain are easy to obtain, the weight of the normal state samples in the source domain is always assigned to 1.

[0067] Considering formula (1) and formula (3) comprehensively, the calculation formula of SWMMD is as follows:

[0068]

[0069] Where V is the SWMMD matrix.

[0070] Step 6: Use the weighted classification loss of source domain samples and the classification loss of target domain samples to reduce the classification error after domain alignment;

[0071] Specifically, in the process of aligning the data distribution of the source domain and the target domain using the SWMMD method, the classification loss of samples in each field is minimized at the same time to further improve the cross-domain fault diagnosis effect under partial domain adaptation. Due to the existence of private category samples in the source domain, it is difficult to effectively reduce the classification loss of shared category samples in the source domain by treating the classification loss of each sample equally. In order to reduce the classification loss to a greater extent, it is considered to increase the weight of the classification loss of shared category samples in the source domain and reduce the weight of the classification loss of private category samples in the source domain. By weighting the classification loss of source domain samples, the uncertainty in the migration process is suppressed to promote the positive migration of the model. In addition, the classification loss of target domain samples is fully utilized to further improve the cross-domain fault diagnosis performance. Specifically, the weighted classification loss of source domain samples and target domain sample classification are integrated into BLS, and the calculation formula is as follows:

[0072]

[0073] Among them, the first term in formula (9) represents the classification loss of source domain samples, Represents the diagonal matrix of source domain sample weights for the iterth iteration of the PDA-BLS model. The second term represents the target domain sample classification loss, is the expected output of the i-th sample in the source domain, is the predicted output of the jth sample in the target domain. The parameter δ∈[0,1] is used to control the importance of the classification loss of source domain samples and the classification loss of target domain samples.

[0074] Step 7: Construct the objective function of the PDA-BLS model through the feature layer nodes, enhancement layer nodes, weighted maximum mean difference term of source domain samples, weighted classification loss term of source domain samples and classification loss term of target domain samples of the BLS framework;

[0075] Step 8: Solve the objective function of the PDA-BLS model, obtain the output weight matrix of the PDA-BLS model, input the target domain samples into the PDA-BLS model for testing, determine the fault category to which the target domain samples belong, and update the source domain sample weights and SWMMD matrix.

[0076] Specifically, combining formula (7) and formula (9), the optimization objective function of the PDA-BLS model is obtained, which is expressed as:

[0077]

[0078] Among them, the parameter γ∈[0,1] is used to balance the data distribution difference term and the sample classification loss term between the source domain and the target domain. According to the Karush–Kuhn–Tucker (KKT) theory, the dual optimization problem of formula (10) can construct the Lagrangian function as follows:

[0079]

[0080] Where η i and λ j is the Lagrange multiplier. PDA-BLS For variables β, ξ i , j , η i , j The partial derivative of and set it equal to 0, the KKT optimal condition is obtained:

[0081]

[0082] After simplification, we can get

[0083]

[0084] Solve the above equation by ridge regression generalized inverse, when the number of training samples n s +n t When it is greater than or equal to the number of network nodes nk+m, the output connection matrix β of the PDA-BLS model is solved based on the left pseudo-inverse algorithm:

[0085]

[0086] In the formula, 1 nk+m It is the identity matrix of (nk+m)×(nk+m).

[0087] When the number of training samples n s +n t When the number of network nodes is less than nk+m, the output connection matrix β of the PDA-BLS model is solved based on the right pseudo-inverse algorithm:

[0088]

[0089] Step 9: Repeat steps (3-8) until the termination condition is met and output the classification results of the target domain samples.

[0090] Specifically, for a target domain sample x with an unknown label, the output of the PDA-BLS model is as shown in formula (16), and then the predicted label of x is obtained by formula (17).

[0091] f(x)=α(x)β=[f1(x),f2(x),…,f C (x)] (16)

[0092] Label(x)=argmaxf i (x),i=1,…,C (17)

[0093] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0094] See also Figure 1 The fault diagnosis method based on partial domain adaptation of the PDA-BLS model provided by the embodiment of the present invention includes:

[0095] Step S110: Set the maximum number of iterations T, the number of feature nodes, the number of enhanced nodes, the regularization parameter θ and the balance factors γ, δ; based on the source domain samples And the probability labeling method, predict the target domain samples The initial label of s is the number of source domain samples, n t is the number of samples in the target domain;

[0096] Step S120: Based on the initial label, the formula Calculate the weight of the qth sample in the source domain category c when the PDA-BLS model is iterated to the iterth time in, The number of samples in the target domain predicted as category c when the PDA-BLS model iterates to iter-1th time;

[0097] Step S130: By formula Calculate the source domain sample weight vector ω of the iter-th iteration of the PDA-BLS model iter ;

[0098] Step S140: By formula Calculate the element (V0) in the i-th row and j-th column of the marginal distribution SWMMD matrix V0 ij ;

[0099] Step S150: By formula Calculate the conditional distribution SWMMD matrix V c The element in the i-th row and j-th column of c ) ij ;

[0100] Step S160: By formula Calculate the SWMMD matrix V;

[0101] Step S170: Randomly initialize feature layer weight parameter W ei and β ei , enhancement layer weight parameter W hj and β hj ; Through the linear transformation function, the input data X = [X s ,X t ] is mapped to the feature layer of the PDA-BLS model to form a feature node Then generate the enhanced node H through the nonlinear transformation of the activation function m =[H1,...,H m ], get the actual input signal in, α(x i ) T is the i-th sample x of the input data X i Output after feature layer transformation and enhancement layer transformation;

[0102] Step S180: By formula Calculate the weighted maximum mean difference term of the source domain samples; where β is the output weight matrix;

[0103] Step S190: By formula The classification loss of the PDA-BLS model is calculated; the first item represents the source domain sample classification loss, the second term represents the target domain sample classification loss, is the diagonal matrix of source domain sample weights for the iterth iteration of the PDA-BLS model, is the expected output of the i-th sample in the source domain, is the predicted output of the jth sample in the target domain, δ∈[0,1] is the classification loss balance factor, which is used to control the importance of the classification loss of the source domain sample and the classification loss of the target domain sample;

[0104] Step S200: Apply L2 regularization to the output weight matrix β to construct the objective function of the PDA-BLS model

[0105] Among them, θ is the regularization term coefficient, γ∈[0,1] is the balance factor between data distribution difference and classification loss, which is used to balance the data distribution difference term and sample classification loss term between the source domain and the target domain;

[0106] Step S210: solving the objective function to obtain an output weight matrix β;

[0107] Step S220: When the number of training samples n s +n tWhen the number of nodes is greater than or equal to nk+m of the PDA-BLS model, the output weight matrix β of the PDA-BLS model is solved based on the left pseudo-inverse algorithm;

[0108] Specifically, the output weight matrix β of the PDA-BLS model is solved based on the left pseudo-inverse algorithm, including:

[0109] By formula Calculate the output weight matrix β; where Ι nk+m It is the identity matrix of (nk+m)×(nk+m).

[0110] Step S230: When the number of training samples n s +n t When the number of nodes is less than nk+m of the PDA-BLS model, the output weight matrix β of the PDA-BLS model is solved based on the right pseudo-inverse algorithm;

[0111] Specifically, the output weight matrix β of the PDA-BLS model is solved based on the right pseudo-inverse algorithm, including:

[0112] By formula Calculate the output weight matrix β; where Y s is the label matrix of the source domain samples, is the label matrix of the predicted target domain samples.

[0113] Step S240: Based on the solved output weight matrix β, the formula f(x)=α(x)β=[f1(x),f2(x),...,f C (x)] to obtain the PDA-BLS model output of the target domain sample x with unknown labels;

[0114] Step S250: By formula Label(x)=argmaxf i (x),i=1,...,C obtains the predicted label of the target domain sample x with unknown label.

[0115] Update the source domain sample weight vector ω based on the predicted label iter And the source domain sample weight matrix W iter , SWMMD matrix V, and continue to calculate the predicted label until the model reaches the number of iterations or converges.

[0116] See also Figure 2 The partial domain adaptation fault diagnosis system based on the PDA-BLS model provided by the embodiment of the present invention includes:

[0117] The initial label prediction module 100 is used to set the maximum number of iterations T, the number of feature nodes, the number of enhanced nodes, the regularization parameter θ and the balance factors γ, δ; based on the source domain samples And the probability labeling method, predict the target domain samples The initial label of s is the number of source domain samples, n t is the number of samples in the target domain;

[0118] The weight acquisition module 200 is used to obtain the weight based on the initial label through the formula Calculate the weight of the qth sample in the source domain category c when the PDA-BLS model is iterated to the iterth time in, The number of samples in the target domain predicted as category c when the PDA-BLS model iterates to iter-1th time;

[0119] The weight vector calculation module 300 is used to calculate the weight vector by formula Calculate the source domain sample weight vector ω of the iter-th iteration of the PDA-BLS model iter ;

[0120] The edge distribution matrix operation module 400 is used to calculate the edge distribution matrix through the formula Calculate the element (V0) in the i-th row and j-th column of the marginal distribution SWMMD matrix V0 ij ;

[0121] The conditional distribution matrix operation module 500 is used to calculate the conditional distribution matrix through the formula Calculate the conditional distribution SWMMD matrix V c The element in the i-th row and j-th column of c ) ij ;

[0122] SWMMD Matrix Operation Module 600 is used to calculate the matrix Calculate the SWMMD matrix V;

[0123] The actual input signal acquisition module 700 is used to randomly initialize the feature layer weight parameter W ei and β ei , enhancement layer weight parameter W hj and β hj ; Through the linear transformation function, the input data X = [X s ,X t ] is mapped to the feature layer of the PDA-BLS model to form a feature node Then generate the enhanced node H through the nonlinear transformation of the activation function m =[H1,...,H m ], get the actual input signal in, α(x i ) Tis the i-th sample x of the input data X i Output after feature layer transformation and enhancement layer transformation;

[0124] The weighted maximum mean difference term calculation module 800 is used to calculate the weighted maximum mean difference term by formula Calculate the weighted maximum mean difference term of the source domain samples; where β is the output weight matrix;

[0125] The classification loss calculation module 900 is used to calculate the classification loss through the formula The classification loss of the PDA-BLS model is calculated; the first item represents the source domain sample classification loss, the second term represents the target domain sample classification loss, is the diagonal matrix of source domain sample weights for the iterth iteration of the PDA-BLS model, is the expected output of the i-th sample in the source domain, is the predicted output of the jth sample in the target domain, δ∈[0,1] is the classification loss balance factor, which is used to control the importance of the classification loss of the source domain sample and the classification loss of the target domain sample;

[0126] The objective function construction module 1000 is used to apply L2 regularization to the output weight matrix β to construct the objective function of the PDA-BLS model.

[0127] Among them, θ is the regularization term coefficient, γ∈[0,1] is the balance factor between data distribution difference and classification loss, which is used to balance the data distribution difference term and sample classification loss term between the source domain and the target domain;

[0128] The objective function solving module 1100 is used to solve the objective function and obtain the output weight matrix β;

[0129] The first solution module 1200 outputs the weight matrix, which is used when the number of training samples is n s +n t When the number of nodes is greater than or equal to nk+m of the PDA-BLS model, the output weight matrix β of the PDA-BLS model is solved based on the left pseudo-inverse algorithm;

[0130] Specifically, the first solution module 1200 for outputting the weight matrix is ​​specifically used when the number of training samples n s +n t When the number of nodes is greater than or equal to nk+m of the PDA-BLS model, the formula Calculate the output weight matrix β; where Ι nk+m It is the identity matrix of (nk+m)×(nk+m).

[0131] The second solution module 1300 outputs the weight matrix, which is used when the number of training samples is n s +n t When the number of nodes is less than nk+m of the PDA-BLS model, the output weight matrix β of the PDA-BLS model is solved based on the right pseudo-inverse algorithm;

[0132] Specifically, the second weight matrix solving module 1300 is used when the number of training samples n is s +n t When the number of nodes is less than nk+m of the PDA-BLS model, the formula Calculate the output weight matrix β; where Y s is the label matrix of the source domain samples, is the label matrix of the predicted target domain samples.

[0133] The model output module 1400 is used to output the weight matrix β obtained by the formula f(x)=α(x)β=[f1(x),f2(x),...,f C (x)] to obtain the PDA-BLS model output of the target domain sample x with unknown labels;

[0134] Fault label prediction module 1500 is used to predict the fault label using the formula Label(x)=argmaxf i (x), i=1,...,C obtains the predicted label of the target domain sample x with unknown label. Update the source domain sample weight vector ω based on the predicted label iter And the source domain sample weight matrix W iter , SWMMD matrix V, and continue to calculate the predicted label until the model reaches the number of iterations or converges.

[0135] Experimental verification and analysis

[0136] (1) Description of experimental dataset

[0137] This experiment uses the bearing data set and gear data set collected by the SpectraQuest power transmission fault diagnosis experimental platform to verify the effectiveness and superiority of the proposed method. The structure of the experimental platform is as follows: Figure 3As shown in the figure. The experimental platform consists of a drive motor, a coupling, a first-stage planetary gearbox, a second-stage parallel gearbox, a brake, a speed controller, a load regulator and a vibration signal acquisition system. The faulty bearing and the faulty gear are installed on the output shaft of the second-stage parallel gearbox. The bearing model is ER-16K, the rolling element diameter is 7.94 mm, the pitch diameter is 38.51 mm, and the number of rolling elements is 9. By adjusting the speed controller and the load regulator, the vibration signals of the unit under different speeds and loads are collected. The bearing data set contains 8 different working states: normal state (NC), inner ring fault (IF), outer ring fault (OF), ball fault (BF), outer ring + ball composite fault (OBF), inner ring + outer ring composite fault (IOF), inner ring + ball composite fault (IBF) and inner ring + outer ring + ball composite fault (IOBF). For each bearing state, the vibration signals under three different working conditions (2990rpm / 2V, 2863rpm / 4V and 3152rpm / 6V) are collected respectively. The gear dataset contains four different working states: normal state (HG), gear surface wear (SWG), gear missing teeth (MTG), and gear eccentricity fault (EG). For each gear state, vibration signals under four different working conditions (2994rpm / 0V, 2994rpm / 4V, 3175rpm / 0V, and 3175rpm / 4V) are collected respectively. Among them, 2994rpm / 0V means that the motor running speed is 2994rpm and the load voltage is 0V. Based on the collected original vibration signal, it is sliced ​​into equal lengths in sequence, and each segment contains 2500 sampling points, among which 100 segments are obtained in each healthy state of the gear dataset, and 200 segments are obtained in each healthy state of the gear dataset, and 46 multi-statistical multi-scale fuzzy measurement entropy are extracted from each segment to form a fault feature set. Based on the above-mentioned bearing data and gear fault data, this experiment designed multiple groups of migration diagnosis tasks B1~B8 and G1~G6 under partial domain adaptation, as shown in Tables 1 and 2, to verify the effectiveness and superiority of the PDA-BLS model provided by the embodiment of the present invention. It can be seen from Tables 1 and 2 that the fault state types in the source domain sample set include the fault state types in the target domain sample set, which is a typical migration fault diagnosis problem under partial domain adaptation.

[0138] Table 1 Description of some domain diagnosis tasks for bearings

[0139]

[0140]

[0141] Table 2 Description of some domain diagnosis tasks for gears

[0142]

[0143] (2) Diagnostic results and comparative analysis of the PDA-BLS model

[0144] In order to verify the superiority of the PDA-BLS model provided by the embodiment of the present invention, it is compared with four existing algorithms, including: (1) K-nearest neighbor (KNN); (2) easy transfer learning (Easy Transfer Learning, EasyTL); (3) joint distribution adaptation method (Joint Distribution Adaptation, JDA); (4) cross-domain extreme learning machine method CDELM-C (Cross-Domain Extreme Learning Machines) based on conditional probability distribution; KNN, as a supervised machine learning method, uses source domain samples to train the model, and uses the trained model to diagnose target domain samples, and the number of neighbors is set to 1; the JDA method uses a Gaussian function as the kernel function, and the regularization term parameter is selected from [0.001, 0.005, 0.01, 0.05, 0.1, 0.5, 1, 5]; the number of hidden layer nodes of the CDELM-C method is set to 500, and other parameters are selected from [10 -4 ,10 -3 ,10 -2 ,10 -1 ,1,10], the manifold regularization term is not considered, the number of iterations is set to 20, and the final diagnosis result is the average diagnostic accuracy of the model repeated 10 times.

[0145] Table 3 Diagnostic results of bearing experiments (%)

[0146] Task Name KNN EasyTL JDA CDELM-C PDA-BLS B1 50 49 50 41.5 99.7 B2 96.33 49 57 61.6 100 B3 66.7 87.67 58.33 78.53 100 B4 78.5 75.67 53.37 60.35 100 B5 81.8 70 88.2 85.6 100 B6 87.33 84 93.33 88.63 100 B7 81.83 90.33 61.5 83.03 99.6 B8 86 80.5 90.29 93.17 96.83 average value 78.56 73.27 69.00 74.05 99.46

[0147] The diagnosis results of all the above methods in the bearing and gear experiments are shown in Table 3 and Table 4 respectively. As can be seen from Table 3: for the KNN method, due to the inconsistent distribution of training data and test data, the diagnosis results are not ideal, and the average diagnosis accuracy is 78.56%; the EasyTL, JDA and CDELM-C methods have negative transfer phenomena in most diagnosis tasks, and the average diagnosis accuracy is 73.27%, 69% and 74.05% respectively. The reason is that the source domain under partial domain adaptation contains fault categories that are irrelevant to the target domain, resulting in the target domain samples being misclassified as source domain private categories, thereby causing the degradation of model performance. For the PDA-BLS method, due to the consideration of the source domain sample weight, the importance of the source domain private category samples can be reduced to a certain extent, so that the diagnosis accuracy of this method is greatly improved. The average diagnosis accuracy of the 8 migration tasks exceeds 96%, and the average diagnosis accuracy of the PDA-BLS method is as high as 99.46%, which shows that the comprehensive consideration of the source domain sample weight and the target domain sample classification loss can promote the improvement of the PDA-BLS model performance. Based on the above analysis, it is shown that the PDA-BLS method can effectively realize fault diagnosis under partial domain adaptation, and its diagnostic performance is better than some existing domain adaptation methods.

[0148] Table 4 Diagnostic results of gear test (%)

[0149] Task Name KNN EasyTL JDA CDELM-C PDA-BLS G1 74.33 69.67 76.5 82.93 99.45 G2 76 78.67 65.33 78.27 98.55 G3 79.67 70.67 82.67 85.23 97.7 G4 73.25 68.5 68.75 54.7 100 G5 85 85.75 94.75 61.95 100 G6 41.5 40.5 49.25 37.5 90.18 average value 71.63 68.96 72.88 66.76 97.65

[0150] In addition, it can be seen from Table 4 that the average diagnostic accuracy obtained by the KNN method is only 71.63%; the EasyTL and CDELM-C methods also have negative transfer phenomena in most migration tasks, with average diagnostic accuracy rates of 68.96% and 66.76% respectively; compared with the KNN method, the diagnostic effect of the JDA method has been improved, with an average diagnostic accuracy of 72.88%. For the PDA-BLS method, the average diagnostic accuracy has been greatly improved, and the average diagnostic accuracy of the six migration tasks has exceeded 90%, and the average diagnostic accuracy of the PDA-BLS method is as high as 97.65%. Based on the above analysis, it is shown that the PDA-BLS method can effectively realize fault diagnosis under partial domain adaptation, and its diagnostic performance is better than some existing domain adaptation methods.

[0151] In view of the cross-domain fault diagnosis problem of the multi-operating condition characteristics of the unit and inconsistent fault categories, an embodiment of the present invention proposes a novel PDA-BLS method for realizing cross-domain fault diagnosis under partial domain adaptation. This method simultaneously considers the weighted maximum mean difference of source domain samples, the weighted classification loss of source domain samples, and the classification loss of target domain samples within the BLS framework. The weighting strategy of source domain samples can correct the selection bias of source domain shared samples and suppress the negative transfer problem that may be caused by source domain private category samples during the domain adaptation process. By reducing the weight of private category samples in the source domain, filtering out source domain samples that are not related to the target domain fault category, and converting the partial domain adaptation problem into a standard domain adaptation problem as much as possible, only the knowledge of source domain samples related to the target domain fault category is transferred, thereby improving the matching degree of source domain and target domain data under partial domain adaptation. In addition, the classification loss of target domain samples is used to further promote positive transfer, and the learned BLS is used to diagnose the target domain samples. The weights of the source domain fault samples are recalculated based on the diagnosis results. This is repeated and the BLS classifier is continuously updated, which improves the cross-domain fault diagnosis effect under partial domain adaptation and provides a new method for the fault diagnosis problem under partial domain adaptation.

[0152] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0153] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0154] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0155] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0156] The parts not described in detail in the embodiments of the present invention are all known technologies to those skilled in the art. Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention is described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions of the present invention, which should be included in the scope of the claims of the present invention.

Claims

1. A fault diagnosis method based on partial domain adaptation of PDA-BLS model, characterized in that: include: Based on source domain samples And the probability labeling method, predict the target domain samples The initial labels of the source domain samples and the target domain samples are obtained based on the vibration signals of the acquisition unit; where n s is the number of source domain samples, n t is the number of samples in the target domain; Based on the initial label through the formula c=1,2,...C, Calculate the weight of the qth sample in the source domain category c when the PDA-BLS model is iterated to the iterth time in, The number of samples in the target domain predicted as category c when the PDA-BLS model iterates to iter-1th time; By formula Calculate the source domain sample weight vector ω of the iter-th iteration of the PDA-BLS model iter ; Represents the weight of the i-th sample in the source domain when the model iterates to the iter-th time; By formula Calculate the element (V0) in the i-th row and j-th column of the marginal distribution SWMMD matrix V0 ij ; By formula Calculate the conditional distribution SWMMD matrix V c The element in the i-th row and j-th column of c ) ij ; By formula Calculate the SWMMD matrix V; The input data X = [X s ,X t ] is mapped to the feature layer of the PDA-BLS model to form a feature node Then generate the enhanced node H through the nonlinear transformation of the activation function m =[H1,...,H m ], get the actual input signal in, α(x i ) T is the i-th sample x of the input data X i Output after feature layer transformation and enhancement layer transformation; By formula Calculate the weighted maximum mean difference of source domain samples; where β is the output weight matrix; By formula The classification loss of the PDA-BLS model is calculated; where is the diagonal matrix of source domain sample weights for the iterth iteration of the PDA-BLS model, is the expected output of the i-th sample in the source domain, is the predicted output of the jth sample in the target domain, δ∈[0,1] is the classification loss balance factor; Apply L2 regularization to the output weight matrix β and construct the objective function of the PDA-BLS model Among them, θ is the regular term coefficient, γ∈[0,1] is the balance factor between data distribution difference and classification loss; Solve the objective function and obtain the output weight matrix β; Based on the solved output weight matrix β, the formula f(x) = α(x) β = [f1(x), f2(x), ..., f C (x)] to obtain the PDA-BLS model output of the target domain sample x with unknown labels; By the formula Label(x) = argmaxf i (x),i=1,...,C obtains the predicted label of the target domain sample x with unknown label.

2. The fault diagnosis method based on partial domain adaptation of the PDA-BLS model as claimed in claim 1, characterized in that: When the number of training samples n s +n t When the number of nodes is greater than or equal to nk+m of the PDA-BLS model, the output weight matrix β of the PDA-BLS model is solved based on the left pseudo-inverse algorithm; When the number of training samples n s +n t When the number of nodes is less than nk+m of the PDA-BLS model, the output weight matrix β of the PDA-BLS model is solved based on the right pseudo-inverse algorithm.

3. The fault diagnosis method based on partial domain adaptation of the PDA-BLS model as claimed in claim 2, characterized in that: The output weight matrix β of the PDA-BLS model is solved based on the left pseudo-inverse algorithm, including: By formula Calculate the output weight matrix β; where Ι nk+m It is the identity matrix of (nk+m)×(nk+m).

4. The fault diagnosis method based on partial domain adaptation of the PDA-BLS model as claimed in claim 2, characterized in that: The output weight matrix β of the PDA-BLS model is solved based on the right pseudo-inverse algorithm, including: By formula Calculate the output weight matrix β; where Y s is the label matrix of the source domain samples, is the label matrix of the predicted target domain samples.

5. A fault diagnosis system based on partial domain adaptation of PDA-BLS model, characterized in that: include: Initial label prediction module, used to predict labels based on source domain samples And the probability labeling method, predict the target domain samples The initial labels of the source domain samples and the target domain samples are obtained based on the vibration signals of the acquisition unit; where n s is the number of source domain samples, n t is the number of samples in the target domain; The weight acquisition module is used to obtain the weight based on the initial label through the formula c=1,2,...C, Calculate the weight of the qth sample in the source domain category c when the PDA-BLS model is iterated to the iterth time in, The number of samples in the target domain predicted as category c when the PDA-BLS model iterates to iter-1th time; The weight vector calculation module is used to calculate the weight vector through the formula Calculate the source domain sample weight vector ω of the iter-th iteration of the PDA-BLS model iter ; Represents the weight of the i-th sample in the source domain when the model iterates to the iter-th time; Edge distribution matrix operation module, used to pass the formula Calculate the element (V0) in the i-th row and j-th column of the marginal distribution SWMMD matrix V0 ij ; Conditional distribution matrix operation module, used to calculate the matrix through the formula Calculate the conditional distribution SWMMD matrix V c The element in the i-th row and j-th column of c ) ij ; SWMMD matrix operation module is used to calculate the matrix of Calculate the SWMMD matrix V; The actual input signal acquisition module is used to convert the input data X=[X s ,X t ] is mapped to the feature layer of the PDA-BLS model to form a feature node Then generate the enhanced node H through the nonlinear transformation of the activation function m =[H1,...,H m ], get the actual input signal in, α(x i ) T is the i-th sample x of the input data X i Output after feature layer transformation and enhancement layer transformation; The weighted maximum mean difference calculation module is used to calculate the weighted maximum mean difference through the formula Calculate the weighted maximum mean difference term of the source domain samples; where β is the output weight matrix; The classification loss calculation module is used to calculate the loss through the formula The classification loss of the PDA-BLS model is calculated; where is the diagonal matrix of source domain sample weights for the iterth iteration of the PDA-BLS model, is the expected output of the i-th sample in the source domain, is the predicted output of the jth sample in the target domain, δ∈[0,1] is the classification loss balance factor; Objective function construction module, used to apply L2 regularization to the output weight matrix β and construct the objective function of the PDA-BLS model Among them, θ is the regular term coefficient, γ∈[0,1] is the balance factor between data distribution difference and classification loss; The objective function solving module is used to solve the objective function and obtain the output weight matrix β; The model output module is used to calculate the output weight matrix β through the formula f(x)=α(x)β=[f1(x),f2(x),...,f C (x)] to obtain the PDA-BLS model output of the target domain sample x with unknown labels; Fault label prediction module, used to predict the fault label using the formula Label(x) = argmaxf i (x),i=1,...,C obtains the predicted label of the target domain sample x with unknown label.

6. The fault diagnosis system based on partial domain adaptation of the PDA-BLS model as claimed in claim 5, characterized in that: Also includes: Output weight matrix first solution module, used when the number of training samples is n s +n t When the number of nodes is greater than or equal to nk+m of the PDA-BLS model, the output weight matrix β of the PDA-BLS model is solved based on the left pseudo-inverse algorithm; Output weight matrix second solution module, used when the number of training samples is n s +n t When the number of nodes is less than nk+m of the PDA-BLS model, the output weight matrix β of the PDA-BLS model is solved based on the right pseudo-inverse algorithm.

7. The fault diagnosis system based on partial domain adaptation of the PDA-BLS model as claimed in claim 6, characterized in that: Output weight matrix first solution module, specifically used when the number of training samples is n s +n t When the number of nodes is greater than or equal to nk+m of the PDA-BLS model, the formula Calculate the output weight matrix β; where Ι nk+m It is the identity matrix of (nk+m)×(nk+m).

8. The fault diagnosis system based on partial domain adaptation of the PDA-BLS model as claimed in claim 6, characterized in that: The second solution module of the output weight matrix is ​​specifically used when the number of training samples is n s +n t When the number of nodes is less than nk+m of the PDA-BLS model, the formula Calculate the output weight matrix β; where Y s is the label matrix of the source domain samples, is the label matrix of the predicted target domain samples.

Images