Combined integrated fault diagnosis method and system

Abstract

The invention relates to the technical field of fault diagnosis, and discloses a combined and integrated fault diagnosis method and system. The method comprises the following steps: acquiring sensor data from vehicle, ground and air equipment, and constructing a K-fold training data set, a test data set and corresponding label sets; using the K-fold training data set to train a plurality of basic fault diagnosis algorithms, and constructing a basic fault diagnosis model; respectively testing the basic fault diagnosis models by using the K-fold test data set; training a set meta-classification algorithm by using the position coding metadata set, and constructing a meta-classifier model; sensor data from vehicle, ground and air equipment in a regional rail transit road network are collected in real time, and a regional rail transit road network operation state monitoring / diagnosis result is obtained by using the constructed combined and integrated fault diagnosis model, so that the advantages of each single fault diagnosis model are effectively integrated; and the diagnosis performance and the overall efficiency of integrated detection and monitoring of the running state of the regional rail transit road network are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a combined integrated fault diagnosis method and system. Technical Background

[0002] With the rapid development and complexity of regional rail transit networks, traditional single fault diagnosis models can no longer meet the needs of integrated detection and monitoring of data in multiple professional fields. Existing single models, such as fault diagnosis models based on convolutional neural networks (CNN), long short-term memory networks (LSTM), and multi-layer perceptrons (MLP), usually focus on processing a certain type of data or a specific task. The CNN-based fault diagnosis model is good at extracting local fault features, but has weak adaptability to dynamic time series data; the LSTM-based fault diagnosis model has strong time series memory and can handle long-distance dependencies in time series data, but performs poorly when facing multi-source heterogeneous data; and the MLP-based fault diagnosis model can classify fault features extracted manually, but has poor adaptability and generalization ability under complex working conditions. Although these single fault diagnosis models have certain advantages in specific scenarios, they are difficult to adapt to the complex and changeable passenger flow, working conditions, road conditions and environment when faced with multi-source, multi-dimensional, heterogeneous and heterogeneous detection or monitoring data of equipment operation status in regional rail transit networks from vehicles (cars), ground infrastructure (ground) and environmental conditions (air), and often cannot provide high-performance diagnostic results. Therefore, there is an urgent need for an integrated fault diagnosis solution that can combine multiple fault diagnosis models / methods and give full play to the advantages of each model to improve the diagnostic performance and overall efficiency of integrated detection and monitoring of the operation status of regional rail transit networks. Summary of the invention

[0003] The purpose of the present invention is to disclose a combined integrated fault diagnosis method and system, so as to realize the combined integration of multiple fault diagnosis models in rail transit and to rapidly diagnose multiple faults.

[0004] To achieve the above objectives, in a first aspect, the present invention discloses a combined integrated fault diagnosis method and system, comprising:

[0005] S1: Collect sensor data from vehicles, ground infrastructure and equipment in the regional rail transit network to construct a K-fold training dataset and its corresponding label set, a K-fold test dataset and its corresponding label set;

[0006] S2: Use the K-fold training data set and the corresponding label set to train, build a basic fault diagnosis model of the basic fault diagnosis algorithm and an optimal parameter basic fault diagnosis model; use the K-fold test data set and the corresponding label set to obtain the primary fault diagnosis test result of the optimal parameter basic fault diagnosis model;

[0007] S3: constructing the primary fault diagnosis test results of all basic fault diagnosis algorithms with position coding based on the primary fault diagnosis test results; and constructing the meta-feature data sets with position coding and the corresponding meta-feature label sets under all arrangement modes;

[0008] S4: training based on the meta-feature data set with position encoding and the corresponding meta-feature label set to construct an optimal parameter meta-classifier model under all arrangement modes, wherein the optimal parameter basic fault diagnosis model and the optimal parameter meta-classifier model are combined to form an integrated fault diagnosis model;

[0009] S5: collecting device sensor data in real time and constructing an inference data set, inputting the inference data set into the optimal parameter basic fault diagnosis model for inference, and obtaining the average primary fault diagnosis inference results of all J columns of output;

[0010] S6: Based on the average primary fault diagnosis reasoning results of all J columns, construct the average primary fault diagnosis reasoning results of all J columns with position coding; and construct the reasoning meta-feature data set with position coding under the u-th arrangement; input the reasoning meta-feature data set with position coding into the optimal parameter meta-classifier model under the u-th arrangement to obtain the final combined integrated fault diagnosis reasoning result.

[0011] To achieve the above-mentioned purpose, in a second aspect, the present invention further discloses a combined integrated fault diagnosis method and system, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method described in the first aspect when executing the computer program.

[0012] The present invention has the following beneficial effects

[0013] The present application provides a combined integrated fault diagnosis method. In the training stage: the method first uses sensor data collected from vehicles (cars), ground infrastructure (ground) and environmental conditions (air) in the regional rail transit network to construct a K-fold training data set and a test data set and their corresponding label sets; uses the K-fold training data set to train multiple basic fault diagnosis algorithms and construct a basic fault diagnosis model; uses the K-fold test data set to test the basic fault diagnosis model; uses the position coding meta-data set to train the set meta-classification algorithm and construct a meta-classifier model; the basic fault diagnosis model and the meta-classifier model are combined to form an integrated fault diagnosis model; in the reasoning stage: real-time collection of sensor data from vehicles, ground and air equipment in the regional rail transit network, using the constructed combined integrated fault diagnosis model to obtain the monitoring / diagnosis results (abnormal / fault location, classification identification and degree identification) of the regional rail transit network operation status, so that the advantages of each single fault diagnosis model are effectively integrated, and the diagnostic performance and overall efficiency of the integrated detection and monitoring of the regional rail transit network operation status are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A flowchart of a combined integrated fault diagnosis method according to a preferred embodiment of the present invention;

[0015] Figure 2(a) , 2(b) , 2(c), 2(d), and 2(e) are respectively the 1st, 2nd, 3rd, 4th, and 5th row matrices in the confusion matrix diagram of the primary fault diagnosis reasoning results of all optimal parameter-based fault diagnosis models of a combined integrated fault diagnosis method of a preferred embodiment of the present invention;

[0016] Figure 3 A confusion matrix diagram of integrated fault diagnosis reasoning results of all optimal parameter meta-classifier models of a combined integrated fault diagnosis method of a preferred embodiment of the present invention; DETAILED DESCRIPTION

[0017] The technical solution of the present invention is described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] Unless otherwise defined, the technical terms or scientific terms used in the present invention shall have the usual meanings understood by persons with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. Similarly, words such as "one" or "one" do not indicate quantity restrictions, but indicate the existence of at least one. Words such as "connect" or "connected" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the object being described changes, the relative positional relationship also changes accordingly.

[0019] It is worth noting that the combined integrated fault diagnosis method provided by the present application is mainly aimed at the field of rail transit, and is applicable to the equipment operation status detection, monitoring and fault diagnosis in vehicles (cars), ground infrastructure (ground) and environmental conditions (air) in rail transit networks. Rail transit vehicles (cars) include but are not limited to high-speed trains / EMUs, ordinary trains, subway trains, light rail trains, monorail trains, maglev trains, heavy-load locomotives, etc.; equipment in vehicles (cars) includes but is not limited to pantographs, traction transformers, four-quadrant rectifiers, traction inverters, traction motors, wheelsets, bogies, etc. Equipment in ground infrastructure (ground) includes tracks, contact lines, power supply equipment, electrical equipment, railway signal systems, switch machines, etc. Equipment in environmental conditions (air) includes but is not limited to anemometers, rain gauges, snow depth gauges, foreign object intrusion collection equipment, road and bridge monitoring equipment, tunnel monitoring equipment, various disaster prevention monitoring equipment, etc.

[0020] In addition, it is worth noting that the basic fault diagnosis model, the optimal parameter basic fault diagnosis model and the corresponding output results in this application can all be expressed in matrices, but this is only one form of expression. This application is only an example and not a limitation. However, no matter how it is transformed, it is within the scope of protection of this application.

[0021] like Figure 1 As shown, this embodiment provides a combined integrated fault diagnosis method, including:

[0022] S1: Collect sensor data from vehicles, ground infrastructure and equipment in the regional rail transit network to construct a K-fold training dataset and its corresponding label set, a K-fold test dataset and its corresponding label set;

[0023] S2: Use the K-fold training data set and the corresponding label set to train, build a basic fault diagnosis model of the basic fault diagnosis algorithm and an optimal parameter basic fault diagnosis model; use the K-fold test data set and the corresponding label set to obtain the primary fault diagnosis test result of the optimal parameter basic fault diagnosis model;

[0024] S3: constructing the primary fault diagnosis test results of all basic fault diagnosis algorithms with position coding based on the primary fault diagnosis test results; and constructing the meta-feature data sets with position coding and the corresponding meta-feature label sets under all arrangement modes;

[0025] S4: training based on the meta-feature data set with position encoding and the corresponding meta-feature label set to construct an optimal parameter meta-classifier model under all arrangement modes, wherein the optimal parameter basic fault diagnosis model and the optimal parameter meta-classifier model are combined to form an integrated fault diagnosis model;

[0026] S5: collecting device sensor data in real time and constructing an inference data set, inputting the inference data set into the optimal parameter basic fault diagnosis model for inference, and obtaining the average primary fault diagnosis inference results of all J columns of output;

[0027] S6: Based on the average primary fault diagnosis reasoning results of all J columns, construct the average primary fault diagnosis reasoning results of all J columns with position coding; and construct the reasoning meta-feature data set with position coding under the u-th arrangement; input the reasoning meta-feature data set with position coding into the optimal parameter meta-classifier model under the u-th arrangement to obtain the final combined integrated fault diagnosis reasoning result.

[0028] In this application, sensor data includes but is not limited to electrical signals / timing sequences, images, video signals, log or message information, audio information, etc.

[0029] The above-mentioned combined integrated fault diagnosis method, in the training stage: uses sensor data collected from vehicles (cars), ground infrastructure (ground) and environmental conditions (air) in the regional rail transit network to construct a K-fold training data set and a test data set and their corresponding label sets; constructs a K-fold training data set to train multiple basic fault diagnosis algorithms and construct a basic fault diagnosis model matrix; uses a K-fold test data set to test the basic fault diagnosis model matrix; uses a position-coded meta-data set to train the set meta-classification algorithm and construct a meta-classifier model; the basic fault diagnosis model and the meta-classifier model are combined to form an integrated fault diagnosis model; in the reasoning stage: real-time collection of operating status detection and monitoring data from vehicles, ground and air equipment in the regional rail transit network, and uses the constructed combined integrated fault diagnosis model to obtain the regional rail transit network operating status diagnosis results (abnormal / fault location, classification identification and degree identification), which effectively integrates the advantages of each single fault diagnosis model and improves the integrated detection and monitoring efficiency and diagnostic performance of the regional rail transit network operating status.

[0030] It should be noted that the device described in this embodiment is a traction motor of an onboard device in a rail transit vehicle, and the sensor data collected is a stator current signal of the traction motor. The main hyperparameters of the model training in this embodiment are shown in Table 1.

[0031] Table 1 Model training hyperparameters

[0032] Learning Rate Iterations Batch size Optimizer Basic fault diagnosis model 0.01 10 128 SGD Meta-classifier model 6e-4 10 128 SGD

[0033] Further, based on the above parameters, the steps of the combined integrated fault diagnosis method provided by the present application are described in detail as follows:

[0034] Optionally, the S1 includes:

[0035] S11: Collect sensor data from vehicles (cars), ground infrastructure (ground), and environmental conditions (air) in the regional rail transit network to form an original data set D = {d1,…,d n ,…,d N},in is a sample, n=1,2,…,N, N is the total number of samples, C is the total number of sensor channels, and W is the length of the sensor data sequence; for sample d n After labeling, the samples and labels satisfy the mapping relationship:

[0036] Λ:d n →y n (1)

[0037] In the formula, y n∈{0,1,2,…,L-1}, L is the total number of labels, which constitutes the label set Y corresponding to the original data set D = {y1,…,y n ,…y N}, Λ represents the mapping from samples to labels.

[0038] S12: Randomly divide the original data set D into K subsets in is the kth data subset, k = 1, 2, ..., K, N s =N / K represents the total number of samples in the kth data subset; all subsets meet the condition i=1,2,…,K; correspondingly, the label set Y is divided into K subsets is the label subset corresponding to the kth data subset; data subset N s samples and label subsets N s The labels satisfy the mapping relationship of formula (1).

[0039] Construct K-fold training data sets separately K-fold validation dataset in are the k-fold training data set and the k-fold validation data set, respectively, k = 1, 2, ..., K, With the kth data subset The relationship satisfies the following relationship:

[0040]

[0041] In the formula, Indicates that in the data set D K Remove the kth data subset from The set consisting of the K-1 data subsets remaining after the difference operation,

[0042] Construct label sets corresponding to K-fold training data sets respectively The label set corresponding to the K-fold validation dataset in are the label sets corresponding to the k-th fold training data set and the k-th fold validation data set, respectively. With the kth label subset The relationship satisfies the following relationship:

[0043]

[0044] In the formula, Indicates that in the label set Y K Remove the k-th label subset from The set consisting of the K-1 label subsets remaining after the difference operation,

[0045] Optionally, the S2 includes:

[0046] S21: Constructing basic fault diagnosis model:

[0047] Assume that J different basic fault diagnosis algorithms are represented as M J ={M1,…,M j ,…,M J}, where M j represents the jth basic fault diagnosis algorithm, j = 1, 2, ..., J, J is the total number of basic fault diagnosis algorithms; using the K-fold training data set and the label set of the K-fold training dataset For the jth basic fault diagnosis algorithm M j Train K times respectively to obtain K basic fault diagnosis models {M 1,j ,…,M k,j ,…,M K,j}; for J basic fault diagnosis algorithms {M1,…,M j ,…,M J} are trained K times respectively, and a total of K×J basic fault diagnosis models are obtained, which are expressed in matrices as follows:

[0048]

[0049] In the formula, are J basic fault diagnosis algorithms {M1,…,M j ,…,M J The basic fault diagnosis model matrix obtained after K trainings, M k,j Represents the k-th row and j-th column element of the basic fault diagnosis model matrix M, that is, the k-th fold data set is used for the j-th basic fault diagnosis algorithm And the corresponding label set The basic fault diagnosis model obtained through training.

[0050] As a preferred implementation of this embodiment, two different fault diagnosis algorithms M1 and M2 (J=2) are set. M1 is a manual feature extraction and multi-layer perceptron classifier algorithm. The manually extracted features include six groups of features: time domain mean, root mean square, standard deviation, peak, kurtosis, and skewness of the three-phase current; frequency domain features include spectral energy, spectral center, spectral bandwidth, spectral kurtosis, and spectral entropy, a total of 5 groups of features; phase features include phase difference mean, phase difference standard deviation, phase difference peak, phase difference square sum and correlation coefficient, a total of 5 groups, a total of 48 features; the classifier uses a four-layer multi-layer perceptron network. M2 is an 18-layer one-dimensional residual neural network algorithm, which includes 5 basic units, among which the feature extractor basic unit 1 includes 1 one-dimensional convolution, 1 batch normalization layer and 1 maximum pooling layer, each basic unit in the feature extractor basic unit 2 to basic unit 4 includes only 1 residual block, and the feature extractor basic unit 5 includes 1 residual block, 1 average pooling layer and 1 flattening layer; the classifier is a single-layer fully connected network. The original data set is divided into 5 folds (K=5), and a 5×2 basic fault diagnosis model matrix is ​​obtained in total. Two different fault diagnosis algorithms train a total of 10 basic fault diagnosis models.

[0051] S22: Construct an optimal parameter-based fault diagnosis model matrix.

[0052] For the jth basic fault diagnosis algorithm M j The kth basic fault diagnosis model M k,j Then perform E rounds of training and update to obtain K optimal parameter basic fault diagnosis models Taking j = 1, 2, ..., J, we get K × J optimal parameter basic fault diagnosis models, which can be expressed as:

[0053]

[0054] In the formula, represents J basic fault diagnosis algorithms {M1,…,M j ,…,M J After K trainings, K×J basic fault diagnosis models M are obtained. Then, each basic fault diagnosis model M is trained separately. k,j The optimal parameter basic fault diagnosis model matrix obtained by E rounds of training; Represents the optimal parameter-based fault diagnosis model matrix The kth row and jth column element of j Use the k-fold dataset And the corresponding label set The kth basic fault diagnosis model M obtained by training k,j , and then the optimal parameter basic fault diagnosis model is obtained through E rounds of training; The optimal parameter-based fault diagnosis model The optimal parameters of The optimization function is:

[0055]

[0056] Where, L ce (·,·) represents the cross entropy loss function, φ k,j The basic fault diagnosis model M k,j Parameters, The parameter is φ k,j Basic fault diagnosis model M k,j In the k-th fold training data set The model output when trained with input, The parameter is φ k,j Basic fault diagnosis model M k,j In the k-th fold training data set Model output when trained for input and Corresponding label set The cross entropy loss function value between ; Represents the cross entropy loss function Returns the model parameter φ when the minimum value is taken k,j Operations for obtaining values; is the k-fold training data set The corresponding label set.

[0057] The stochastic gradient descent method is used to transform the basic fault diagnosis model M in formula (4) k,j Perform the e-th round of training update. The formula for updating the model parameters in the e-th round is:

[0058]

[0059] In the formula, φ k,j,e Update the model for round e The parameter, φ k,j,e-1 Update model for round e-1 Parameters, Update model parameters φ for round e k,j,e The learning rate, The parameter is φ k,j,e-1 The e-1th round update model In the k-th fold training data set Model output when trained for input; Represents the cross entropy loss function About φ k,j,e-1 The partial derivative of .

[0060] Specifically, when e=1, the basic fault diagnosis model M in formula (4) is k,j The formula for updating the first round of model parameters is:

[0061]

[0062] where φ k,j,0 To use the k-fold data set for the j-th basic fault diagnosis algorithm And the corresponding label set Basic fault diagnosis model obtained through training (model not updated) Initialization parameters; Indicates that the initialization parameter is φ k,j,0 The unupdated model In the k-th fold training data set Model output when trained for input; Denotes the cross entropy loss function About the basic fault diagnosis model (no updated model) Initialization parameter φ k,j,0 The partial derivative of .

[0063] Take e = 1, 2, ..., E, and repeat formula (7) to update until the maximum number of iterations E. At this time, the E-th round of updating model The parameter is φ k,j,E , let the optimal parameter That is, the jth basic fault diagnosis algorithm uses the kth fold data set And the corresponding label set The basic fault diagnosis model M obtained by training k,j , and then the optimal parameter basic fault diagnosis model obtained by E rounds of training is

[0064] Take k = 1, 2, ..., K, j = 1, 2, ..., J, and construct the optimal parameter basic fault diagnosis model matrix of formula (5)

[0065] S23: Use the k-fold validation dataset Test the optimal parameter basic fault diagnosis model matrix separately The kth row of the J optimal parameter-based fault diagnosis model The primary fault diagnosis test results of the k-th row J optimal parameter basic fault diagnosis model are obtained, which can be expressed as:

[0066]

[0067] in, Indicates that the k-th row of J optimal parameter basic fault diagnosis models are input into the k-th fold validation data set respectively Primary fault diagnosis test results obtained during the test; Represents the optimal parameter-based fault diagnosis model for the kth row and jth column Input is the k-fold validation dataset The primary fault diagnosis test result obtained when is expressed as:

[0068]

[0069] In the formula, Indicates input of the k-fold validation dataset When testing, the corresponding nth s The primary fault diagnosis test results obtained by samples, n s =1,2,…,N s , For the nth s The test result of the l-th class label obtained by the samples is the probability of being judged as the l-th class label, l = 0, 1, 2, ..., L, L is the total number of labels.

[0070] Take k = 1, 2, ..., K, and get the primary fault diagnosis test result matrix obtained by testing the J basic fault diagnosis models trained by all K-fold training data sets and the K-fold validation data sets, expressed as:

[0071]

[0072] in, Represents the optimal parameter-based fault diagnosis model matrix The primary fault diagnosis test result matrix obtained by testing the K-fold validation data set for all optimal parameter basic fault diagnosis models in , N = KN s ;

[0073] Primary Fault Diagnosis Test Results Matrix The corresponding true label matrix is ​​expressed as:

[0074]

[0075] in, Represents the primary fault diagnosis test result matrix The corresponding true label matrix is ​​obtained from the K-fold validation dataset The corresponding label set Build, where is the k-fold validation dataset The corresponding label set.

[0076] Optionally, the S3 includes:

[0077] S31: Primary fault diagnosis test result matrix The jth column All elements are position-coded to obtain the jth basic fault diagnosis algorithm M with position coding. j Primary fault diagnosis test results The formula is:

[0078]

[0079] in, Denotes the jth basic fault diagnosis algorithm M with position coding j The primary fault diagnosis test results of cat(·,·) represent the result of horizontal (column-wise) stacking operation on the two matrices in the brackets. represents an N×1 column matrix of all 1 elements; onehot(·) represents the result of the one-hot encoding operation, which is expressed as:

[0080] onehot(j)=[O1(j) … O t (j) … O J (j)] (14)

[0081] in, represents the one-hot encoded column vector about j, O t (j) is the value of the tth element in the one-hot encoding vector of j, t = 1, 2, 3, ..., J, satisfying the condition:

[0082]

[0083] In formula (9) Kronecker product is an operation that multiplies all elements of the first matrix by the second matrix in two matrices of arbitrary size, and replaces the results in blocks to fill in all elements of the first matrix to synthesize a larger matrix. The result is:

[0084]

[0085] In the formula, Representation Matrix With the matrix The result of the Kronecker product operation; in formula (9) For the matrix With the matrix After the horizontal stacking operation, a new matrix with a dimension of N×(L+J) is obtained;

[0086] Primary fault diagnosis test result matrix All elements in the jth column of the , j = 1, 2, ..., J, and J basic fault diagnosis algorithms with position coding {M1, ..., M j ,…,M J Primary fault diagnosis test results

[0087] S32: Construct the full permutation set of the natural number sequence {1,2,…,j,…,J}, expressed as:

[0088] S J ={σ1,σ1,…,σ u ,…,σ U} (17)

[0089] Among them, S J is the set of all permutations of the natural number sequence {1,2,…,j,…,J}, σ u represents the full permutation set S J The u-th arrangement, σ u =(σ u (1),σ u (2),…,σ u (j)…,σ u (J)), u = 1, 2, 3, ..., U, U is the total number of permutations, U = J!, where J! represents the total number of permutations of J, σ u (j) represents the jth value of the uth arrangement, j = 1, 2, 3, ..., J; for any σ v ,σ n (v≠n)∈S J ,have

[0090] When J = 3, that is, the natural number sequence is {1,2,3}, the full permutation set S3, a total of U = 3! = 6 permutations, expressed as: S3 = {σ1,σ2,σ3,σ4,σ5,σ6}, the first, second, third, fourth, fifth, and sixth permutations are σ1 = (1,2,3), σ2 = (1,3,2), σ3 = (2,1,3), σ4 = (2,3,1), σ5 = (3,1,2), σ6 = (3,2,1); the fourth permutation is denoted as σ4 = (σ4(1),σ4(2),σ4(3)), σ4(1) = 2, σ4(2) = 3, σ4(3) = 1 respectively represent the first, second, and third values ​​of the fourth permutation.

[0091] For the u-th arrangement σ u ∈S J , construct the meta-feature dataset matrix with position encoding under the u-th arrangement, expressed as:

[0092]

[0093] In the formula represents the matrix of meta-feature data set with position encoding under the u-th arrangement; represents the column matrix of the jth meta-feature data set with position encoding under the uth arrangement;

[0094] The meta-feature dataset matrix with position encoding in all u-th (u=1,2,3,…,U) permutations Stacking vertically (by row) obtains the meta-feature dataset matrix with position encoding, expressed as:

[0095]

[0096] In the formula, It means that the meta-feature dataset matrices with position coding under all U kinds of arrangement modes are stacked vertically (by row) to obtain the meta-feature dataset matrix with position coding;

[0097] Corresponding meta-feature dataset matrix with positional encoding The meta-feature label set matrix is ​​expressed as:

[0098]

[0099] In the formula, Represents the corresponding feature dataset matrix with position encoding The meta-feature label set matrix of ;

[0100] Optionally, the S4 includes:

[0101] S41: Set the meta-classifier algorithm to F; use the meta-feature dataset matrix with position encoding And the corresponding meta-feature label set matrix Perform T rounds of training optimization on the meta-classifier algorithm F to obtain the optimal parameter meta-classifier model is the optimal parameter meta-classifier model The optimal parameters of .

[0102] S42: Constructing optimal parameters The optimization function is:

[0103]

[0104] In the formula, is the optimal parameter meta-classifier model The optimal parameters of the meta-classifier model F θ Parameters, Denotes the meta-classifier model F with parameters θ θ In the dataset matrix D with positional encoding meta-features meta,pos is the model output when training with input, L ce (F θ (D meta,pos ),Y meta,pos ) represents the meta-classifier model F with parameter θ θ In the dataset matrix D with positional encoding meta-features meta,pos Model output when trained for input With D meta,pos Corresponding label set Y meta,pos The cross entropy loss function value between ; Denotes the cross entropy loss function L ce (F θ (D meta,pos ),Y meta,pos ) returns the value of the model parameter θ when it takes the minimum value;

[0105] Satisfies the following expression:

[0106]

[0107] in, Denotes the meta-classifier model F with parameters θ θ The matrix of the feature dataset with position encoding in the u-th arrangement The output of the model when trained on the input.

[0108] S43: Meta-classifier model using stochastic gradient descent Perform the tth round of training update, and the formula for updating the model parameters in the tth round is:

[0109]

[0110] In the formula, Update the model for round t Parameters, Update the model for round t-1 Parameters, Update model parameters for round t The learning rate; Indicates that the parameter is The t-1th round update model The matrix of the feature dataset with position encoding in the u-th arrangement Model output when trained for input; Indicates that the parameter is The t-1th round update model The matrix of the feature dataset with position encoding in the u-th arrangement Model output when trained for input and Corresponding label set The cross entropy loss function value between ; Represents the cross entropy loss function about The partial derivative of .

[0111] Specifically, when t = 1, the meta-classifier model F θ The formula for updating the first round of model parameters is:

[0112]

[0113] In the formula, Update the model for round 1 Parameters, The matrix of the meta-feature dataset with position encoding under the u-th arrangement is used for the meta-classifier algorithm And the corresponding label set Meta classifier model obtained through training (model not updated) Initialization parameters; Indicates that the initialization parameters are The unupdated model The matrix of the feature dataset with position encoding in the u-th arrangement Model output when trained for input; Denotes the cross entropy loss about The partial derivative of .

[0114] Take t = 1, 2, ..., T, repeat formula (23) to update until the maximum number of iterations T, then the Tth round of updating model The parameters are Let the optimal parameters That is, the meta-classifier algorithm uses the meta-feature dataset matrix with position encoding under the u-th arrangement And the corresponding label set The meta-classifier model obtained by training The optimal parameter meta-classifier model obtained after T rounds of training is:

[0115] Take u=1,2,…,U to form the optimal parameter meta-classifier model matrix

[0116] As a preferred implementation of this embodiment, the meta-classifier algorithm is set to a multi-layer perceptron algorithm including a three-layer multi-layer perceptron, and a meta-classifier model matrix containing two models is obtained by training in total.

[0117] It is worth emphasizing that the optimal parameter basic fault diagnosis model and the optimal parameter meta-classifier model are combined to form an integrated fault diagnosis model.

[0118] Optionally, the S5 includes:

[0119] S51: In the inference phase, collect device sensor data in real time and build an inference data set Where N infer is the total number of samples in the inference dataset;

[0120] S52: Using the Inference Dataset Input the optimal parameter basic fault diagnosis model matrix of formula (5) respectively All K×J optimal parameter basic fault diagnosis models in the , and obtain the output results of the K×J optimal parameter basic fault diagnosis models, that is, the primary fault diagnosis reasoning result matrix, expressed as:

[0121]

[0122] In the formula, Represents the optimal parameter-based fault diagnosis model matrix Inference dataset The primary fault diagnosis reasoning result matrix obtained as input, Represents the optimal parameter-based fault diagnosis model matrix The optimal parameter-based fault diagnosis model for the kth row and jth column in Input inference dataset D infer The primary fault diagnosis reasoning result obtained when is expressed as:

[0123]

[0124] In the formula, Denotes the input inference dataset D infer When making inferences, it corresponds to the nth infer The primary fault diagnosis reasoning results obtained by samples, n infer =1,2,…,N infer , For the nth infer The primary fault diagnosis inference result about the lth class label obtained by samples is the probability size of the lth class label, l = 0, 1, 2, ..., L-1, L is the total number of labels.

[0125] S53: Optimal parameter-based fault diagnosis model matrix The K optimal parameter basic fault diagnosis model in the jth column (the jth basic fault diagnosis algorithm) Input the inference dataset D respectively infer When reasoning, the primary fault diagnosis reasoning result of the j-th column K optimal parameter basic fault diagnosis model is obtained as follows:

[0126]

[0127] As a preferred embodiment of the present invention, the 5×2 optimal parameter basic fault diagnosis model matrix, the 1st, 2nd, 3rd, 4th, and 5th row matrices of the confusion matrix diagram of the primary fault diagnosis reasoning results of all optimal parameter basic fault diagnosis models are respectively as follows: Figure 2(a) , 2(b) , 2(c), 2(d), and 2(e).

[0128] All K elements in formula (27) are averaged to obtain the average primary fault diagnosis reasoning result of the primary fault diagnosis reasoning results of the K optimal parameter basic fault diagnosis model in the jth column, and the formula is:

[0129]

[0130] In the formula, is the average value of the primary fault diagnosis reasoning results of the j-th column K optimal parameter basic fault diagnosis models, that is, the average primary fault diagnosis reasoning result of the j-th column of the optimal parameter basic fault diagnosis model;

[0131] Take j = 1, 2, ..., J, and obtain the j-th column average primary fault diagnosis reasoning result matrix of all J basic fault diagnosis algorithms, expressed as:

[0132]

[0133] In the formula, is the average primary fault diagnosis inference result matrix of all J columns of the optimal parameter basic fault diagnosis model.

[0134] Optionally, S6 includes:

[0135] S61: Average primary fault diagnosis reasoning results for the jth column Perform position encoding to obtain the j-th column average primary fault diagnosis reasoning result with position encoding, the formula is:

[0136]

[0137] In the formula is the average primary fault diagnosis reasoning result for the jth column Perform position encoding to obtain the j-th column average primary fault diagnosis reasoning result with position encoding; Represents the matrix With the matrix Stack horizontally to get a dimension of N infer ×(L+J) operation of the new matrix;

[0138] Construct the average primary fault diagnosis reasoning result matrix of all J columns with position coding, the formula is:

[0139]

[0140] in, is the average primary fault diagnosis inference result matrix of all J columns with position encoding.

[0141] S62: For any u-th arrangement σ u ∈S J ,(u∈1,2,3,…,U), average all J columns of the primary fault diagnosis reasoning result matrix with position coding Stacking is performed horizontally (by column) to obtain the matrix of the meta-feature dataset with positional coding inference under the u-th arrangement, which is expressed as:

[0142]

[0143] In the formula, It represents the average primary fault diagnosis reasoning result matrix of all J columns with position coding under the u-th arrangement mode Stack horizontally (by column) to obtain the matrix of the meta-feature dataset with positional coding inference in the u-th arrangement;

[0144] S63: The inference meta-feature dataset matrix with position encoding under the u-th arrangement Input to the optimal parameter meta-classifier model under the u-th arrangement The optimal parameter meta-classifier model output is obtained, which is the combined integrated fault diagnosis reasoning result:

[0145]

[0146] In the formula, The optimal parameter is The optimal meta-classifier model for The matrix of the inference meta-feature dataset with position encoding in the u-th arrangement The model output when reasoning with the input, that is, N infer Inference dataset of samples The combined integrated fault diagnosis reasoning results.

[0147] Taking u=1,2,…,U, the combined integrated fault diagnosis reasoning results are as follows:

[0148]

[0149] As a preferred embodiment of the present invention, the optimal parameter meta-classifier model matrix of the two models is included, and the confusion matrix of the combined integrated fault diagnosis reasoning result of each optimal parameter meta-classifier model is as follows: Figure 3 shown.

[0150] The present application also provides a combined integrated fault diagnosis system, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the computer program. The combined integrated fault diagnosis system can implement various embodiments of the above combined integrated fault diagnosis method and achieve the same beneficial effects, which will not be described in detail here.

[0151] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A combined integrated fault diagnosis method, characterized in that: include: S1: Collect sensor data from vehicles, ground infrastructure and equipment in the regional rail transit network to construct a K-fold training dataset and its corresponding label set, a K-fold test dataset and its corresponding label set; S2: Use the K-fold training data set and the corresponding label set to train, build a basic fault diagnosis model of the basic fault diagnosis algorithm and an optimal parameter basic fault diagnosis model; use the K-fold test data set and the corresponding label set to obtain the primary fault diagnosis test result of the optimal parameter basic fault diagnosis model; S3: constructing primary fault diagnosis test results of all basic fault diagnosis algorithms with position coding based on the primary fault diagnosis test results; And construct the meta-feature dataset with position encoding and the corresponding meta-feature label set under all arrangements; S4: training based on the meta-feature data set with position encoding and the corresponding meta-feature label set to construct an optimal parameter meta-classifier model under all arrangement modes, wherein the optimal parameter basic fault diagnosis model and the optimal parameter meta-classifier model are combined to form an integrated fault diagnosis model; S5: collecting device sensor data in real time and constructing an inference data set, inputting the inference data set into the optimal parameter basic fault diagnosis model for inference, and obtaining the average primary fault diagnosis inference results of all J columns of output; S6: Based on the average primary fault diagnosis reasoning results of all J columns, construct the average primary fault diagnosis reasoning results of all J columns with position coding; and construct the reasoning meta-feature data set with position coding under the u-th arrangement; input the reasoning meta-feature data set with position coding into the optimal parameter meta-classifier model under the u-th arrangement to obtain the final combined integrated fault diagnosis reasoning result.

2. The combined integrated fault diagnosis method according to claim 1, characterized in that: The S1 includes: S11: Collect sensor data of equipment in rail transit vehicles to form an original data set D = {d1,…,d n ,…,d N },in is the sample, n=1,2,…,N, N is the total number of samples, C is the total number of sensor channels, W is the length of the sensor data sequence, is the real number domain; for sample d n After labeling, the mapping relationship between samples and labels is as follows: L:d n →y n (1) In the formula, y n ∈{0,1,2,…,L-1}, L is the total number of labels, which constitutes the label set Y corresponding to the original data set D = {y1,…,y n ,…y N }, Λ represents the mapping from sample to label; S12: Randomly divide the original data set D into K subsets in is the kth data subset, k = 1, 2, ..., K, N s =N / K represents the total number of samples in the kth data subset; all subsets meet the condition Divide the label set Y into K subsets is the label subset corresponding to the kth data subset; data subset N s The samples are respectively associated with the label subset N s The labels satisfy the mapping relationship of formula (1); Construct K-fold training data sets separately K-fold validation dataset in are the k-fold training data set and the k-fold validation data set, respectively, k = 1, 2, ..., K, With the kth data subset The relationship satisfies the following relationship: In the formula, Indicates that in the data set D K Remove the kth data subset from The set consisting of the K-1 data subsets remaining after the difference operation, Construct label sets corresponding to K-fold training data sets respectively The label set corresponding to the K-fold validation dataset in are the label sets corresponding to the k-th fold training data set and the k-th fold validation data set, respectively. With the kth label subset The relationship satisfies the following relationship: In the formula, Indicates that in the label set Y K Remove the kth label subset from The set consisting of the K-1 label subsets remaining after the difference operation, 3. The combined integrated fault diagnosis method according to claim 1, characterized in that: Assume that J different basic fault diagnosis algorithms are represented as M J ={M1,…,M j ,…,M J }, where M j represents the jth basic fault diagnosis algorithm, j = 1, 2, ..., J, J is the total number of basic fault diagnosis algorithms; using the K-fold training data set and the label set of the K-fold training dataset For the jth basic fault diagnosis algorithm M j Train K times respectively to obtain K basic fault diagnosis models {M 1,j ,…,M k,j ,…,M K,j }; for J basic fault diagnosis algorithms {M1,…,M j ,…,M J } Train K times respectively, and get K×J basic fault diagnosis models, which are expressed as follows in matrix: In the formula, are J basic fault diagnosis algorithms {M1,…,M j ,…,M J The basic fault diagnosis model obtained after K trainings, M k,j represents the k-th row and j-th column element of the basic fault diagnosis model M; S22: Constructing the optimal parameter-based fault diagnosis model: For the jth basic fault diagnosis algorithm M j The kth basic fault diagnosis model M k,j Then perform E rounds of training and update to obtain K optimal parameter basic fault diagnosis models Taking j = 1, 2, ..., J, we get K × J optimal parameter basic fault diagnosis models, which can be expressed as: In the formula, represents J basic fault diagnosis algorithms {M1,…,M j ,…,M J After K trainings, K×J basic fault diagnosis models M are obtained. Then, each basic fault diagnosis model M is trained separately. k,j The optimal parameter-based fault diagnosis model obtained by E rounds of training; Represents the optimal parameter-based fault diagnosis model The k-th row and j-th column element of ; The optimal parameter-based fault diagnosis model The optimal parameters of The optimization function satisfies the following relationship: Where, L ce (·,·) represents the cross entropy loss function, φ k,j The basic fault diagnosis model M k,j Parameters, The parameter is φ k,j Basic fault diagnosis model M k,j In the k-th fold training data set The model output when trained with input, The parameter is φ k,j Basic fault diagnosis model M k,j In the k-th fold training data set Model output when trained for input and Corresponding label set The cross entropy loss function value between ; Represents the cross entropy loss function Returns the model parameter φ when the minimum value is taken k,j Operations for obtaining values; is the k-fold training data set The corresponding label set; The stochastic gradient descent method is used to transform the basic fault diagnosis model M in formula (4) k,j Perform the e-th round of training update. The formula for updating the model parameters in the e-th round is as follows: In the formula, φ k,j,e Update the model for round e The parameter, φ k,j,e-1 Update model for round e-1 Parameters, Update model parameters φ for round e k,j,e The learning rate, The parameter is φ k,j,e-1 The e-1th round update model In the k-th fold training data set Model output when trained for input; Represents the cross entropy loss function About φ k,j,e-1 The partial derivative of It means to find partial derivative; Take e = 1, 2, ..., E, and repeat formula (7) to update until the maximum number of iterations E. At this time, the E-th round of updating model The parameter is φ k,j,E , let the optimal parameter That is, the jth basic fault diagnosis algorithm uses the kth fold data set And the corresponding label set The basic fault diagnosis model M obtained by training k,j , and then the optimal parameter basic fault diagnosis model obtained by E rounds of training is Take k = 1, 2, ..., K, j = 1, 2, ..., J, and construct the optimal parameter-based fault diagnosis model of formula (5) S23: Use the k-fold validation dataset Test the optimal parameter-based fault diagnosis model separately The kth row of the J optimal parameter-based fault diagnosis model The primary fault diagnosis test results of the k-th row J optimal parameter basic fault diagnosis model are obtained, which are expressed as follows: in, Indicates that the k-th row of J optimal parameter basic fault diagnosis models are input into the k-th fold validation data set respectively Primary fault diagnosis test results obtained during the test; Represents the optimal parameter-based fault diagnosis model for the kth row and jth column Input is the k-fold validation dataset The primary fault diagnosis test result obtained when is expressed as: In the formula, Indicates input of the k-fold validation dataset When testing, it corresponds to the nth s The primary fault diagnosis test results obtained by samples, n s =1,2,…,N s , For the nth s The test results of the lth class label obtained by samples, l = 0, 1, 2, ..., L, L is the total number of labels; Take k = 1, 2, ..., K, and get the primary fault diagnosis test results of the J basic fault diagnosis models trained by all K-fold training data sets and tested by the K-fold validation data sets, which are expressed as: in, Represents the optimal parameter-based fault diagnosis model The primary fault diagnosis test results obtained by testing all the optimal parameter basic fault diagnosis models with K-fold validation data sets; Primary troubleshooting test results The corresponding true label is expressed as: in, Indicates the results of the primary fault diagnosis test The corresponding true label.

4. The combined integrated fault diagnosis method according to claim 1, characterized in that: The S3 includes: S31: Test results for primary fault diagnosis The jth column All elements are position-coded to obtain the jth basic fault diagnosis algorithm M with position coding. j Primary fault diagnosis test results Satisfies the following relationship: in, Denotes the jth basic fault diagnosis algorithm M with position coding j The primary fault diagnosis test results of cat(·,·) represent the result of horizontal stacking operation on the two matrices in brackets. represents an N×1 column matrix of all 1 elements, Kronecker product is an operation that multiplies all elements of the first matrix by the second matrix in two matrices of arbitrary size, and replaces the result in blocks to fill in all elements of the first matrix to synthesize a larger matrix; onehot(·) represents the result of one-hot encoding operation; Primary fault diagnosis test results All elements in the jth column of the , j = 1, 2, ..., J, and J basic fault diagnosis algorithms with position coding {M1, ..., M j ,…,M J Primary fault diagnosis test results S32: Construct the full permutation set of the natural number sequence {1,2,…,j,…,J}, expressed as: S J ={σ1,σ1,…,σ u ,…,s U } (16) Among them, S J is the set of all permutations of the natural number sequence {1,2,…,j,…,J}, σ u Represents the full permutation set S J The u-th arrangement, σ u =(σ u (1),σ u (2),…,σ u (j)…,σ u (J)), u = 1, 2, 3, ..., U, U is the total number of permutations, U = J!, where J! represents the total number of permutations of J, σ u (j) represents the jth value of the uth arrangement, j = 1, 2, 3, ..., J; for any σ v ,σ n (v≠n)∈S J ,have For the u-th arrangement σ u ∈S J , construct the meta-feature dataset with position encoding under the u-th arrangement, expressed as: In the formula, represents the meta-feature dataset with position encoding under the u-th arrangement; represents the jth meta-feature dataset column with positional encoding under the uth arrangement; The meta-feature dataset with position encoding for all u-th arrangement modes By stacking vertically, we can obtain a meta-feature dataset with position encoding, which is expressed as: Corresponding meta-feature dataset with position encoding The meta-feature label set is expressed as: In the formula Indicates the results of the primary fault diagnosis test The corresponding true label.

5. The combined integrated fault diagnosis method according to claim 1, characterized in that: The S4 includes: S41: Set the meta-classifier algorithm to F; use the meta-feature dataset with position encoding: and the corresponding meta-feature tag set Perform T rounds of training optimization on the meta-classifier algorithm F to obtain the optimal parameter meta-classifier model is the optimal parameter meta-classifier model The optimal parameters of S42: Constructing optimal parameters The optimization function satisfies the following relationship: Where θ is the meta-classifier model F θ Parameters, Denotes the meta-classifier model F with parameters θ θ In the dataset D with positional encoding meta-features meta,pos is the model output when training with input, L ce (F θ (D meta,pos ),Y meta,pos ) represents the meta-classifier model F with parameter θ θ In the dataset D with positional encoding meta-features meta,pos Model output when trained for input With D meta,pos Corresponding label set Y meta,pos The cross entropy loss function value between ; Denotes the cross entropy loss function L ce (F θ (D meta,pos ),Y meta,pos ) returns the value of the model parameter θ when it takes the minimum value; S43: Meta-classifier model using stochastic gradient descent Perform the tth round of training update. The formula for updating the model parameters in the tth round is as follows: In the formula, Update the model for round t Parameters, Update the model for round t-1 Parameters, Update model parameters for round t The learning rate; Indicates that the parameter is The t-1th round update model Meta-feature dataset with position encoding in the u-th arrangement Model output when trained for input; Indicates that the parameter is The t-1th round update model Meta-feature dataset with position encoding in the u-th arrangement Model output when trained for input and Corresponding label set The cross entropy loss function value between ; Represents the cross entropy loss function about The partial derivative of Take t = 1, 2, ..., T, repeat formula (21) to update until the maximum number of iterations T, then the Tth round of updating model The parameters are Let the optimal parameters That is, the meta-classifier algorithm uses the position-encoded meta-feature dataset under the u-th arrangement And the corresponding label set The meta-classifier model obtained by training The optimal parameter meta-classifier model obtained after T rounds of training is: Take u=1,2,…,U to form the optimal parameter meta-classifier model 6. The combined integrated fault diagnosis method according to claim 1, characterized in that: The S5 includes: S51: In the inference phase, collect device sensor data in real time and build an inference data set Where N infer is the total number of samples in the inference dataset; S52: Using the Inference Dataset Input optimal parameters of basic fault diagnosis model respectively All K×J optimal parameter basic fault diagnosis models in the , and obtain the primary fault diagnosis reasoning results of K×J optimal parameter basic fault diagnosis models, which satisfy the following relationship: In the formula, Represents the optimal parameter-based fault diagnosis model The optimal parameter-based fault diagnosis model for the kth row and jth column in Input inference dataset D infer The primary fault diagnosis reasoning results obtained when , are expressed as follows: In the formula, Denotes the input inference dataset D infer When making inferences, it corresponds to the nth infer The primary fault diagnosis reasoning results obtained by samples, n infer =1,2,…,N infer , For the nth infer The primary fault diagnosis reasoning results about the l-th class label obtained by samples, l = 0, 1, 2, ..., L, L is the total number of labels; S53: Optimal parameter-based fault diagnosis model The jth column K optimal parameter basic fault diagnosis model Input the inference dataset D respectively infer When reasoning, the primary fault diagnosis reasoning result of the j-th column K optimal parameter basic fault diagnosis model is obtained as follows: All K elements in formula (24) are averaged to obtain the average primary fault diagnosis reasoning result of the primary fault diagnosis reasoning results of the K optimal parameter basic fault diagnosis model in the jth column, and the formula is as follows: In the formula, is the average value of the primary fault diagnosis reasoning results of the j-th column K optimal parameter basic fault diagnosis models; Take j = 1, 2, 3, ..., J, and obtain the j-th average primary fault diagnosis reasoning results of all J basic fault diagnosis algorithms, which are expressed as follows: In the formula, It is the average primary fault diagnosis reasoning result of all J columns of the optimal parameter basic fault diagnosis model.

7. The combined integrated fault diagnosis method according to claim 6, characterized in that: The S6 includes: S61: Average primary fault diagnosis reasoning results for the jth column Perform position encoding to obtain the j-th column average primary fault diagnosis reasoning result with position encoding, the formula is as follows: In the formula, is the average primary fault diagnosis reasoning result for the jth column Perform position encoding to obtain the j-th column average primary fault diagnosis reasoning result with position encoding; Represents the matrix With the matrix Stack horizontally to get a dimension of N infer ×(L+J) operation of the new matrix; Construct the average primary fault diagnosis reasoning results of all J columns with position encoding, the formula is as follows: in, Average primary fault diagnosis reasoning results for all J columns with position encoding; S62: For any u-th arrangement σ u ∈S J ,(u∈1,2,3,…,U), average the primary fault diagnosis reasoning results of all J columns with position encoding By stacking horizontally, we can obtain the meta-feature dataset with positional coding reasoning under the u-th arrangement, which is expressed as follows: In the formula, represents the average primary fault diagnosis reasoning result of all J columns with position coding under the u-th arrangement Stack horizontally to obtain the meta-feature dataset with positional coding inference in the u-th arrangement; S63: The reasoning meta-feature dataset with position encoding under the u-th arrangement Input to the optimal parameter meta-classifier model under the u-th arrangement Get the combined integrated fault diagnosis reasoning result output by the optimal parameter meta-classifier model: In the formula, The optimal parameter is The optimal meta-classifier model for Inference meta-feature dataset with position encoding in the u-th arrangement The model output when reasoning with the input, that is, N infer Inference dataset of samples Combined integrated fault diagnosis reasoning results; Taking u=1,2,…,U, the combined integrated fault diagnosis reasoning results are as follows:

8. A combined integrated fault diagnosis system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method described in any one of claims 1 to 7 are implemented.

Images