Speed sensor interference source identification method based on GWO-CFDP algorithm

Abstract

The invention relates to the technical field of train signal monitoring, and discloses a speed sensor interference source identification method based on a GWO-CFDP algorithm, which comprises the following steps of: acquiring time domain data of a TCU speed sensor under an interference working condition when a train runs, and performing data preprocessing including data segmentation, normalization processing, spectral analysis, envelope extraction and data splicing; obtaining an initial feature vector of the time domain data, inputting the initial feature vector into a multi-layer sparse self-encoding network for sparse encoding to obtain a low-dimensional feature data set, mapping the low-dimensional feature data set to a high-dimensional space in a nonlinear manner by adopting a kernel principal component analysis method, and extracting and screening nonlinear principal components to obtain a final dimension-reduced feature data set; according to the final dimension reduction feature data set, adopting a self-adaptive density peak value clustering method improved based on a grey wolf algorithm to carry out clustering identification, and obtaining an identified interference source type; according to the method, classification and identification of various interference working conditions are realized, and the accuracy of classification and identification under the interference working conditions is improved.

Description

Technical Field

[0001] The invention relates to the technical field of train signal monitoring, and in particular to a method for identifying interference sources of a speed sensor based on a GWO-CFDP algorithm. Background Art

[0002] During the operation of the EMU, the TCU speed sensor is often interfered with. The characteristics of various types of interference signals have high similarity, and distinguishing interference from different sources becomes a major challenge. The characteristics of these interference signals are difficult to fix, making it difficult to achieve accurate identification using conventional feature extraction and recognition methods. Summary of the invention

[0003] In view of the above-mentioned deficiencies in the prior art, the present invention provides a speed sensor interference source identification method based on the GWO-CFDP algorithm, which is used to solve the problem of low accuracy in identifying interference features by conventional feature extraction and identification methods, thereby improving the accuracy and applicability of classification and identification of interference conditions.

[0004] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0005] The method for identifying interference sources of a speed sensor based on the GWO-CFDP algorithm comprises the following steps:

[0006] S1. collecting time domain data of the TCU speed sensor under interference conditions when the train is running, and performing data preprocessing including data segmentation, normalization, spectrum analysis, envelope extraction and data splicing on the time domain data to obtain the initial feature vector of the time domain data;

[0007] S2. Input the initial feature vector of the time domain data into a multi-layer sparse autoencoder network for sparse encoding to obtain a low-dimensional feature data set, and use the kernel principal component analysis method to nonlinearly map the low-dimensional feature data set to a high-dimensional space, and extract and screen the nonlinear principal components to obtain the final reduced-dimensional feature data set;

[0008] S3. According to the final dimension reduction feature data set, the adaptive density peak clustering method based on the improved Grey Wolf Algorithm is used to perform clustering identification to obtain the identified interference source type.

[0009] The present invention has the following beneficial effects:

[0010] The speed sensor interference source identification method based on the GWO-CFDP algorithm proposed in the present invention first performs feature analysis on the actual interference signal of the train; then, the self-extraction and dimension reduction of the features of the pre-processed signal are completed by adopting a multi-layer sparse autoencoder network together with a kernel principal component analysis; finally, the GWO-CFDP algorithm, i.e., an adaptive density peak clustering method improved based on the grey wolf algorithm, is used to realize the classification and identification of various interference conditions, thereby improving the accuracy and effectiveness of the classification and identification under interference conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a flow chart of a method for identifying interference sources of a speed sensor based on the GWO-CFDP algorithm;

[0012] Figure 2 It is a schematic diagram of the structure of a multi-layer sparse autoencoder network;

[0013] Figure 3 Schematic diagram of decision making in a normal situation in the embodiment;

[0014] Figure 4 This is a decision diagram for introducing indicators in the embodiment. DETAILED DESCRIPTION

[0015] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0016] like Figure 1 As shown, the speed sensor interference source identification method based on the GWO-CFDP algorithm includes the following steps S1-S3:

[0017] S1. Collect the time domain data of the TCU speed sensor under interference conditions when the train is running, and perform data preprocessing on the time domain data including data segmentation, normalization, spectrum analysis, envelope extraction and data splicing to obtain the initial feature vector of the time domain data.

[0018] In this embodiment, in order to improve the efficiency of feature extraction in subsequent steps, a series of data preprocessing is first performed on the time domain data under interference conditions, including data segmentation, normalization, spectrum analysis, envelope extraction, and data splicing, so as to use the initial feature vector of the spliced ​​time domain data to perform deep feature sub-extraction using a multi-layer sparse autoencoder network in subsequent steps, thereby improving data accuracy. In addition, the initial feature vector of the time domain data is the feature data of the preprocessed bow network offline, in-out phase separation, main circuit breaker disconnection, and main circuit breaker closing.

[0019] Specifically, step S1 includes S11-S12:

[0020] S11. Collect time domain data of the TCU speed sensor under interference conditions when the train is running; wherein the interference conditions are pantograph-network offline, in-and-out phase separation, main circuit breaker disconnection, and main circuit breaker closing.

[0021] In this embodiment, the TCU speed sensor shield layer signal, the secondary side signal of the high-voltage transformer, the differential mode signal of the traction converter input line, the common mode signal of the traction converter input line, the differential mode signal of the traction converter output line, the common mode signal of the traction converter output line, the traction converter ground line signal and the traction transformer ground line signal of a certain type of high-speed EMU are monitored for a long time, and the data acquisition mode is set to threshold trigger storage. When the amplitude of the TCU speed sensor shield layer signal exceeds the threshold, all monitoring point data 0.5ms before the trigger and 9.5ms after the trigger are stored. The signal of half a cycle length of the power frequency signal is selected for storage because the category of the interference signal can be distinguished by using 10ms of data, and the pressure of data storage can be greatly reduced, which is convenient for the needs of long-term vehicle monitoring. By analyzing the characteristics of the long-term monitoring data and combining the working condition table recorded on the vehicle, it is found that there are four working conditions that cause interference to the TCU speed sensor, namely, bow-net offline, in-out phase separation, main circuit breaker disconnection, and main circuit breaker closing; therefore, the time domain data under these four working conditions can be collected and used for the identification of interference sources in the subsequent steps.

[0022] S12, performing data segmentation on the time domain data and dividing it into three segments to obtain three segments of time domain data.

[0023] In this embodiment, the purpose of data segmentation is to reduce random variables of data.

[0024] S13, using the minimum-maximum normalization method to normalize the segmented three-segment time domain data to obtain normalized three-segment time domain data.

[0025] In this embodiment, the purpose of the normalization process is to adjust the data to the interval [0, 1] to reduce the possibility of large feature differences among similar data due to a large amplitude fluctuation range of the disturbance pulse.

[0026] S14, performing spectrum analysis on the normalized three-segment time domain data, and obtaining envelope features of the normalized three-segment time domain data by extracting the spectrum envelope of the normalized three-segment time domain data.

[0027] S15. Perform data splicing on the envelope features of the three normalized time domain data segments to obtain an initial feature vector of the time domain data.

[0028] In this embodiment, steps S14-S15 reduce the differences between signals of the same type by acquiring the spectrum envelope. Each piece of data has three-segment envelope features. Finally, the three-segment envelope features of each piece of data are concatenated to obtain an initial feature vector. This feature vector reduces the randomness between data of the same type while containing the inherent features of the signal as much as possible, which is beneficial to feature extraction in subsequent steps.

[0029] S2. Input the initial feature vector of the time domain data into a multi-layer sparse autoencoder network for sparse encoding to obtain a low-dimensional feature data set, and use the kernel principal component analysis method to nonlinearly map the low-dimensional feature data set to a high-dimensional space, and extract and screen the nonlinear principal components to obtain the final reduced-dimensional feature data set.

[0030] In this embodiment, since the initial feature vectors of the preprocessed time domain data, i.e., the four types of feature data (feature data of bow network offline, in and out phase separation, main circuit breaker disconnection, and main circuit breaker closing) are obviously different, but the feature dimensions are too many, a multi-layer sparse autoencoder network is used to extract the deep features of the initial feature vectors of the time domain data to obtain a low-dimensional feature data set. At the same time, in order to more accurately screen out the most useful features for classification and recognition, the kernel principal component analysis method is used to calculate the contribution rate of each feature, and the features whose cumulative variance contribution rate exceeds the predetermined extraction rate are selected as the input of the classification and recognition algorithm in the subsequent steps.

[0031] Specifically, step S2 specifically includes S21-S23:

[0032] S21. Construct a multi-layer sparse autoencoder network formed by stacking single-layer sparse autoencoder networks; wherein the single-layer sparse autoencoder network includes a hidden layer; the multi-layer sparse autoencoder network includes an input layer, several hidden layers and an output layer.

[0033] like Figure 2 As shown, Figure 2The multi-layer sparse autoencoder network includes an input layer, two hidden layers, and one output layer; and the entire training process of the multi-layer sparse autoencoder network can be expressed as finding a mapping relationship for the input data through multiple adjustments of network parameters, encoding and decoding processes, so that the output data obtained after decoding is as similar to the input data as possible. Compared with ordinary encoders, sparse autoencoders add sparsity constraints. In conjunction with the minimum loss function, they can extract the feature expression of the input data even when there are more neurons in the hidden layer.

[0034] S22. Input the initial feature vector of the time domain data into a multi-layer sparse autoencoder network for encoding and decoding operations to obtain a low-dimensional feature data set.

[0035] Specifically, step S22 specifically includes S221-S225:

[0036] S221, input the initial feature vector of the time domain data into the input layer of the multi-layer sparse autoencoder network, and input it into the hidden layer for encoding operation to obtain the output of the hidden layer, that is:

[0037] o=f1(W1x+b1)

[0038] Among them, o represents the output of the hidden layer, f1 represents the activation function of the hidden layer, W1 represents the weight coefficient from the input layer to the hidden layer, x represents the input data of the input layer, and b1 represents the bias vector of the hidden layer.

[0039] Among them, the activation function of the hidden layer is the sigmoid activation function.

[0040] In this embodiment, the sigmoid activation function is selected as the activation function of the hidden layer because of its high stability and strong applicability.

[0041] S222, the output of the hidden layer is used as the input of the output layer for decoding operation to obtain the output of the output layer, that is:

[0042] y=f2(W2o+b2)

[0043] Among them, y represents the output of the output layer, f2 represents the activation function of the output layer, W2 represents the weight coefficient from the hidden layer to the output layer, and b2 represents the bias vector of the hidden layer.

[0044] Among them, the activation function of the output layer is the sigmoid activation function.

[0045] S223. Introduce sparsity constraints as the loss function of the multi-layer sparse autoencoder network, that is:

[0046]

[0047] Among them, J(W,b) represents the loss function, W represents the weight coefficient, b represents the bias vector, p represents the number of training samples, and h W,b represents the prediction function of the hidden layer, λ1 represents the coefficient of the regularization term, n l represents the number of neurons in the lth layer, s l represents the parameters of the lth layer, W jk represents the weight between the jth neuron and the kth neuron, β represents the weight of the penalty factor, and n h represents the number of hidden layers, KL represents the divergence function, ρ represents the sparsity parameter, and ρ j1 represents the average activation value of neurons in the j1th hidden layer, and log represents the logarithmic function.

[0048] In this embodiment, the sparsity constraint is introduced as the loss function of the multi-layer sparse autoencoder network to reduce the difference between the input data and the output data, so that the multi-layer sparse autoencoder network can be represented compactly and meaningfully, which helps to improve the efficiency of feature extraction of the multi-layer sparse autoencoder network and prevent the occurrence of overfitting. In addition, the first two terms in the loss function, namely: It is often called the data fitting term or reconstruction error term, which measures the output h of the autoencoder W,b The difference between x and the actual target data y; The penalty term of the loss function is also called the weight decay term, which is used to penalize large weight values ​​to prevent the weight of the autoencoder from becoming too large, thereby helping to prevent overfitting. At the same time, the divergence function KL is used to limit the difference between the sparsity parameter and the average activation value of the hidden layer neural unit.

[0049] S224. Use the loss function to calculate the gradient of the weight coefficient and the bias vector, and then perform a forward operation to calculate the deviation between the output of the output layer of the multi-layer sparse autoencoder network and the label value. When the deviation is the smallest, the current weight coefficient and bias vector are used as the optimal parameters of the multi-layer sparse autoencoder network.

[0050] In this embodiment, in order to achieve effective self-extraction of the features of the TCU speed sensor disturbance data, it is necessary to minimize the loss function, that is, use the back propagation algorithm to calculate the gradient of the loss function in the autoencoder network to the parameters and, and then feed the calculation results back to the optimization algorithm such as gradient descent to update the values ​​of the parameters and so that the loss function is as small as possible. The training process of the back propagation algorithm is as follows: first initialize the parameters W and b, then perform forward operations, calculate the deviation between the prediction result of the output layer of the sparse autoencoder network and the label value, and when the deviation is minimized, obtain the optimal parameters.

[0051] S225. The multi-layer sparse autoencoder network is trained again based on the optimal parameters of the multi-layer sparse autoencoder network to obtain a trained multi-layer sparse autoencoder network, and a low-dimensional feature data set is output.

[0052] S23. Use Gaussian kernel function to nonlinearly map the low-dimensional feature data set to high-dimensional space to obtain a high-dimensional feature data set, and use principal component analysis to calculate the eigenvalues ​​and eigenvectors of the high-dimensional feature data set, then perform nonlinear principal component extraction and screening to obtain the final reduced-dimensionality feature data set.

[0053] In this embodiment, principal component analysis (PCA) is a common data dimensionality reduction method, which mainly realizes data dimensionality reduction by maximizing the projection variance of the data set (n dimension) on the new dimension (m dimension, and n>m), and the n-dimensional orthogonal data is the principal component of the m-dimensional data set. However, PCA is only applicable to linearly separable data, and the improved method of PCA is commonly used for dimensionality reduction of nonlinear data-kernel principal component analysis (KPCA). KPCA is an extended method of PCA, which combines kernel function and PCA. This method nonlinearly maps low-dimensional input data to high-dimensional space through kernel function, and then uses PCA to calculate the eigenvalues ​​and eigenvectors of high-dimensional data, and finally extracts the principal components. Among them, the kernel function selects Gaussian kernel function, which can map the original low-dimensional input data to high-dimensional space, and this mapping is nonlinear. This enables KPCA to handle complex structures and patterns in nonlinear data sets; by applying PCA in high-dimensional space and then extracting principal components, Gaussian kernel function helps to reduce the dimension of data while retaining important information of data as much as possible.

[0054] Specifically, step S23 includes:

[0055] S231. Use a Gaussian kernel function to nonlinearly map the low-dimensional feature data set to a high-dimensional space to obtain a high-dimensional feature data set.

[0056] Among them, the low-dimensional feature dataset is defined as:

[0057] Y={y1,y2,...,y π}, Y∈n.

[0058] The nonlinear mapping relationship is:

[0059] φ:Y→F.

[0060] The high-dimensional feature dataset is:

[0061] Φ(y)=(Φ(y1), Φ(y2),..., Φ(y i ), ..., Φ(y q )),Φ(y)∈m,m<n.

[0062] Among them, n and m represent dimensions, Y represents a low-dimensional feature dataset, y1, y2, y π They represent the first, second, and π low-dimensional feature data in the low-dimensional feature data set, respectively. → represents the mapping relationship. F represents the high-dimensional space. φ and Φ(y) both represent high-dimensional feature data sets. Φ(y1), Φ(y2), Φ(y i )、Φ(y q ) represent the first, second, i, and q high-dimensional feature data in the high-dimensional dataset respectively.

[0063] S232. Calculate the covariance matrix of the high-dimensional feature data set, that is:

[0064]

[0065] Among them, C represents the covariance matrix of the high-dimensional feature data set, T represents the transpose, and q represents the total number of high-dimensional feature data.

[0066] S233, calculate the eigenvalues ​​and eigenvectors of the covariance matrix of the high-dimensional feature data set, that is:

[0067]

[0068] Among them, λ i represents the eigenvalue of the i-th high-dimensional feature data, P i ′ represents the feature vector of the i-th high-dimensional feature data, a i Represents the linear coefficient matrix of the i-th high-dimensional feature data.

[0069] S234, based on the eigenvector of the covariance matrix of the high-dimensional feature data set, extract the nonlinear principal components of the low-dimensional feature data set projected in the high-dimensional space, that is:

[0070]

[0071] Among them, α i Represents the variance contribution rate of the i-th high-dimensional feature data.

[0072] S235. Based on the eigenvalues ​​and eigenvectors of the covariance matrix of the high-dimensional feature data set, the reduced-dimensional feature data set and the cumulative variance contribution rate are obtained, namely:

[0073]

[0074] Among them, Y′ represents the reduced dimension feature data set, and L represents the cumulative variance contribution rate.

[0075] S236 , selecting a reduced-dimensionality feature data set whose cumulative variance contribution rate is greater than a preset extraction rate as the final reduced-dimensionality feature data set.

[0076] S3. According to the final dimension reduction feature data set, the adaptive density peak clustering method based on the improved Grey Wolf Algorithm is used to perform clustering identification to obtain the identified interference source type.

[0077] In this embodiment, the fast clustering by fast search and find of density peaks (CFDP) algorithm is a method suitable for clustering non-spherical distribution data. This method is a method based on fast search of density peaks. It does not need to know the number of clustering categories in advance, and meets the premise of using the disturbed working condition identification in the actual work of TCU speed sensor. The algorithm is based on two assumptions. One is that the cluster center is surrounded by neighboring points with lower density, and the other is that the distance from the neighboring points of this cluster center to its cluster center is smaller than that to other cluster centers. Therefore, the CFDP algorithm calculates the local density of the data set and the minimum distance between a point in the data set and other data points with higher density. Generally, there are two methods to measure the local density of a data set. The first is truncated kernel density and the second is Gaussian kernel density. Since the fast clustering method based on density peak is sensitive to data density, when the amount of data is small, the density calculated by truncated kernel density may produce statistical errors, thereby affecting the clustering effect, so the Gaussian kernel function should be used to calculate the local density of the data point. At the same time, the minimum distance is defined as: In addition, it is difficult to ensure that the local density and minimum distance between data points in the data set used for classification each time remain consistent. Data points with large density values ​​are likely to dominate the clustering results, and data points with small density values ​​are ignored. Therefore, the local density and minimum distance can be normalized to unify data points of different densities on the same scale, thereby reducing the impact of large differences in density values ​​and minimum distances on the clustering effect.

[0078] Usually, for any point in the data set, the corresponding normalized local density and minimum distance can be calculated. The normalized local density value is the horizontal axis variable, and the normalized minimum distance value is the vertical axis variable. A scatter plot is drawn in the plane rectangular coordinate system, such as Figure 3 As shown in , points with relatively large two parameters are selected as cluster centers, and unassigned points are assigned to the category where the nearest high-density points are located, so that several clusters are obtained. Each cluster represents a disturbance condition. The selection of cluster centers affects the final clustering results. Therefore, Figure 3 It is called a decision diagram.

[0079] However, in order to reduce the number of variables that need to be determined in advance, the indicator γ is introduced i″ , the calculation formula is The decision diagram corresponds to Figure 4As shown in Figure 2, the CFDP algorithm can determine the number of cluster centers, but the algorithm also has limitations, that is, the selection of the number of cluster centers requires prior information and the need to pre-set γ i″ The lower limit of the value γ judge and d c If the value is not selected properly, the number of clusters will be wrong, which will eventually affect the clustering effect and the accuracy of the results. Therefore, it is particularly important to realize the adaptive selection of the number of clusters. c The optimization is performed to construct an adaptive density peak clustering method based on the improved Grey Wolf Algorithm to cluster and identify interference sources. The algorithm is as follows:

[0080] Specifically, step S3 specifically includes S31-S37:

[0081] S31. According to the final dimension reduction feature data set, an initial cutoff distance is set, and the initial cutoff distance is optimized using the gray wolf algorithm to obtain the optimal cutoff distance.

[0082] Specifically, step S31 includes S311-S315:

[0083] S311, initialize the parameters of the gray wolf group, specifically:

[0084] Set the number of gray wolves and the boundary of the gray wolf search space.

[0085] The maximum and minimum values ​​of the data points in the final dimension reduction feature data set are used as the upper and lower bounds of the gray wolf search space.

[0086] Randomly initialize the position of each gray wolf, that is:

[0087]

[0088] in, represents the initial position of the gray wolf j2 in dimension i2, L i2 represents the lower bound of the gray wolf search space initially set under dimension i2, rand represents a random function used to generate a random number between 0 and 1, and U j2 , L j2 They respectively represent the search range of the gray wolf j2 in a certain dimension.

[0089] S312. Divide the social hierarchy, that is, use the Davidson-Boulding index as the fitness function to calculate the distance between each gray wolf and its prey, and sort the distance between each gray wolf and its prey from small to large, and divide the three gray wolves closest to the prey into α′ wolf, β′ wolf, and δ′ wolf, and the other gray wolves into ω′ wolf.

[0090] Among them, using the Davidson-Bodin index as the fitness function, the formula for calculating the distance between each gray wolf and its prey is:

[0091]

[0092] Among them, DBI represents the distance between each gray wolf and its prey, K represents the number of categories, max represents the maximum value, S i′ represents the standard error between the data point and the cluster center in the i′th cluster, S j′ represents the standard error between the data point and the cluster center in the j′th cluster, d i′j′ represents the distance between the cluster centers of the i′th cluster and the j′th cluster, Z i′ represents the number of data points in the i′th cluster, v i′ represents the cluster center, and x1 represents a single data point within the cluster.

[0093] S313, surround the prey, that is, when the position of the prey no longer changes, update the position of the gray wolf group based on the nonlinear decreasing convergence factor, that is:

[0094]

[0095] in, represents the distance between the gray wolf and its prey. represents the position of prey P at iteration number t, denotes the coefficient vector, represents the position of the gray wolf at iteration number t, represents the convergence factor of nonlinear decrease, Both represent random vectors, T′ represents the maximum number of iterations, and exp represents the empirical function.

[0096] In this embodiment, since the gray wolf algorithm is prone to fall into a local optimal solution, which affects the prediction accuracy, it is necessary to improve the gray wolf algorithm. That is, the convergence factor of the gray wolf algorithm is changed to a convergence factor based on nonlinear decreasing, that is: In order to improve the accuracy and efficiency of finding the optimal solution and balance local convergence and global search capabilities.

[0097] S314, group hunting stage, that is, based on the position of the gray wolf group, the α′ wolf first determines the hunting target, and then the β′ wolf and the δ′ wolf lead the other ω′ wolves to surround the hunting target, and at the same time continuously obtain the position of the prey to update the position of the α′ wolf, the β′ wolf, and the δ′ wolf, that is:

[0098]

[0099] in, Respectively represent the distance between α′ wolf, β′ wolf, δ′ wolf and the prey, are coefficient vectors, They represent the positions of α′ wolf, β′ wolf, and δ′ wolf at iteration number t, respectively. They represent the disturbed positions of α′ wolf, β′ wolf, and δ′ wolf respectively.

[0100] Based on the updated positions of α′ wolf, β′ wolf, and δ′ wolf, in the next iteration, the position of the gray wolf group is adaptively updated, that is:

[0101]

[0102] in, represents the position of the j3rd gray wolf at iteration t+1;

[0103] In this embodiment, in order to solve the problem that the α′ wolf, β′ wolf, and δ′ wolf have different influences in the wolf pack, the above formula is used to adaptively update the position of the gray wolf group.

[0104] At the same time, according to the importance of α′ wolf, β′ wolf, and δ′ wolf, dynamic weights are introduced to readjust the position of the prey, namely:

[0105]

[0106] Among them, X P (t+1) represents the position of the prey P at iteration number t+1, W j3 represents the weight of the j3rd gray wolf, represents the position of the j3rd gray wolf, X α′ , X β′ , X δ′ Respectively represent the positions of α′ wolf, β′ wolf, and δ′ wolf, W α′ , W β′ , W δ′ They represent the dynamic weights of α′ wolf, β′ wolf, and δ′ wolf respectively.

[0107] In this embodiment, the dynamic weights allow the algorithm to adjust the search strategy according to the importance or fitness of different individuals in the wolf pack, thereby searching for the optimal solution more efficiently and ensuring that more important individuals have a greater impact on the location of the prey.

[0108] S315, when When , the gray wolf group will attack the prey and take the current position of the wolf α′ as the optimal cutoff distance. When , return to step S313 to perform iterative calculation until the maximum number of iterations is reached or the optimal truncation distance is found

[0109] S32. According to the optimal cutoff distance and the final reduced dimension feature data set, the local density and the minimum distance of the data points in the final reduced dimension feature data set are calculated, that is:

[0110]

[0111] in, represents the local density of data point i″, e represents the exponential function, d i″j″ represents the Euclidean distance from data point i″ to data point j″, d c represents the cutoff distance, δ i″ represents the minimum distance from data point i″ to data point j″, represents the local density of data point j", mmin represents the minimum value, represents the maximum local density.

[0112] S33, normalizing the local density and the minimum distance of the data points to obtain the normalized local density and the minimum distance of the data points.

[0113] S34. Based on the normalized local density and minimum distance of the data points, the decision index of the data points is calculated, namely:

[0114]

[0115] in, represents the normalized local density of data point i″, represents the minimum distance of normalized data points, γ i″ represents the decision indicator of data point i″, and α1 represents the compensation parameter.

[0116] S35. After normalizing the data points and the decision indicators of the data points, a decision diagram of the decision indicator values ​​of the normalized data points is drawn with the normalized data points as the horizontal axis and the decision indicators of the normalized data points as the vertical axis.

[0117] S36. According to the decision diagram of the decision index value of the normalized data point, the vertical distance from each data point to the straight line is calculated, and the point with the largest vertical distance is selected as the knee point.

[0118] Among them, the straight line is the connection between the first data point and the last data point in the decision graph of the decision indicator value of the normalized data point.

[0119] S37, selecting w data points before the knee point as cluster centers, and assigning unassigned data points to the category where the nearest high-density point is located, to obtain classified clusters;

[0120] Among them, each cluster represents a disturbance condition.

[0121] In this embodiment, steps S35-S27, since the number of clusters depends on γ judge The selection of the value directly affects the clustering effect. judge The value is adaptively selected as the data set changes, introducing the concept of knee point. The knee point is usually used to describe a curve with a clear direction change at a certain point, but the sign of the slope does not necessarily change. The knee point is usually used as the best compromise point between balancing performance and cost, and is often selected as the optimal point in specific problems. Therefore, the knee point can also be used to select the clustering center of the decision graph.

[0122] In addition, the accuracy rate ACC and F-measure are used to evaluate the clustering effect; the definition of the accuracy rate ACC is: Among them, TP is the number of samples whose predicted category and true category are both positive, TN is the number of samples whose predicted category is positive but the actual category is negative, FN is the number of samples whose predicted category and true category are both negative, FP is the number of samples whose predicted category is negative but the actual category is positive; F-measure is an external evaluation indicator used to evaluate the clustering effect, which can comprehensively evaluate the effectiveness of the clustering results. Originally, F-measure was only applicable to binary classification problems. Therefore, when applied to clustering evaluation, it can only evaluate the quality of a single cluster. In order to comprehensively evaluate the effect of the entire clustering, the average F-measure of all clusters in the clustering results is used as a measure of the overall clustering quality. The calculation formula of F-measure is: Among them, Precision represents the proportion of the predicted positive class in the true positive class, and Recall represents the proportion of the predicted positive class in the true positive class; in clustering evaluation, θ can be selected according to the needs of the specific application scenario to reflect the different importance of precision and recall. For example, if you want to pay more attention to precision, you can choose a larger θ value; if you want to pay more attention to recall, you can choose a smaller θ value or θ = 1. At this time, F-measure is simplified to F1-measure, and its calculation formula is: That is, when the accuracy ACC and F1-measure values ​​are closer to 1, the clustering effect is better.

[0123] In summary, the speed sensor interference source identification method based on the GWO-CFDP algorithm proposed in the present invention performs feature analysis on the measured interference signal of the train; then, the self-extraction and dimension reduction of the features of the preprocessed signal are completed by adopting a multi-layer sparse autoencoder network together with kernel principal component analysis; finally, the GWO-CFDP algorithm, i.e., the adaptive density peak clustering method improved based on the grey wolf algorithm, is used to realize the classification and identification of four interference conditions, and the experimental results show that the interference source identification method proposed in the present invention has an identification accuracy of ACC to 0.990 for the four interference conditions, and F1-measure reaches 0.991, which verifies the effectiveness and practical value of the present invention in train interference source identification.

[0124] The present invention uses specific embodiments to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

[0125] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A speed sensor interference source identification method based on the GWO-CFDP algorithm is characterized in that: The following steps are involved: S1. collecting time domain data of the TCU speed sensor under interference conditions when the train is running, and performing data preprocessing including data segmentation, normalization, spectrum analysis, envelope extraction and data splicing on the time domain data to obtain the initial feature vector of the time domain data; S2. Input the initial feature vector of the time domain data into a multi-layer sparse autoencoder network for sparse encoding to obtain a low-dimensional feature data set, and use the kernel principal component analysis method to nonlinearly map the low-dimensional feature data set to a high-dimensional space, and extract and screen the nonlinear principal components to obtain the final reduced-dimensional feature data set; S3. According to the final dimension reduction feature data set, the adaptive density peak clustering method based on the improved Grey Wolf Algorithm is used to perform clustering identification to obtain the identified interference source type.

2. The method for identifying interference sources of a speed sensor based on the GWO-CFDP algorithm according to claim 1, characterized in that: Step S1 specifically includes: S11, collecting time domain data of the TCU speed sensor under interference conditions when the train is running; Among them, the interference conditions are pantograph-network offline, phase separation in and out, main circuit breaker off, and main circuit breaker on; S12, performing data segmentation on the time domain data and segmenting the time domain data into three segments to obtain three segments of time domain data; S13, using the minimum-maximum normalization method to normalize the three segments of time domain data to obtain normalized three segments of time domain data; S14, performing spectrum analysis on the normalized three-segment time domain data, and obtaining envelope features of the normalized three-segment time domain data by extracting spectrum envelopes of the normalized three-segment time domain data; S15. Perform data splicing on the envelope features of the three normalized time domain data segments to obtain an initial feature vector of the time domain data.

3. The method for identifying interference sources of a speed sensor based on the GWO-CFDP algorithm according to claim 2 is characterized in that: Step S2 specifically includes: S21, constructing a multi-layer sparse autoencoder network stacked by a single-layer sparse autoencoder network; Among them, the single-layer sparse autoencoder network includes a hidden layer; The multi-layer sparse autoencoder network includes an input layer, several hidden layers and an output layer; S22, inputting the initial feature vector of the time domain data into a multi-layer sparse autoencoder network for encoding and decoding operations to obtain a low-dimensional feature data set; S23. Use Gaussian kernel function to nonlinearly map the low-dimensional feature data set to high-dimensional space to obtain a high-dimensional feature data set, and use principal component analysis to calculate the eigenvalues ​​and eigenvectors of the high-dimensional feature data set, then perform nonlinear principal component extraction and screening to obtain the final reduced-dimensionality feature data set.

4. The method for identifying interference sources of a speed sensor based on the GWO-CFDP algorithm according to claim 3 is characterized in that: Step S22 specifically includes: S221, input the initial feature vector of the time domain data into the input layer of the multi-layer sparse autoencoder network, and input it into the hidden layer for encoding operation to obtain the output of the hidden layer, that is: o=f1(W1x+b1) Among them, o represents the output of the hidden layer, f1 represents the activation function of the hidden layer, W1 represents the weight coefficient from the input layer to the hidden layer, x represents the input data of the input layer, and b1 represents the bias vector of the hidden layer; Among them, the activation function of the hidden layer is the sigmoid activation function; S222, the output of the hidden layer is used as the input of the output layer for decoding operation to obtain the output of the output layer, that is: y=f2(W2o+b2) Among them, y represents the output of the output layer, f2 represents the activation function of the output layer, W2 represents the weight coefficient from the hidden layer to the output layer, and b2 represents the bias vector of the hidden layer; Among them, the activation function of the output layer is the sigmoid activation function; S223. Introduce sparsity constraints as the loss function of the multi-layer sparse autoencoder network, that is: Among them, J(W,b) represents the loss function, W represents the weight coefficient, b represents the bias vector, p represents the number of training samples, and h W,b represents the prediction function of the hidden layer, λ1 represents the coefficient of the regularization term, n l represents the number of neurons in the lth layer, s l represents the parameters of the lth layer, W jk represents the weight between the jth neuron and the kth neuron, β represents the weight of the penalty factor, and n h represents the number of hidden layers, KL represents the divergence function, ρ represents the sparsity parameter, and ρ j1 represents the average activation value of neurons in the j1th hidden layer, and log represents the logarithmic function; S224, using the loss function to calculate the gradient of the weight coefficient and the bias vector, and then performing a forward operation to calculate the deviation between the output of the multi-layer sparse autoencoder network output layer and the label value, and when the deviation is the smallest, using the current weight coefficient and the bias vector as the optimal parameters of the multi-layer sparse autoencoder network; S225. The multi-layer sparse autoencoder network is trained again based on the optimal parameters of the multi-layer sparse autoencoder network to obtain a trained multi-layer sparse autoencoder network, and a low-dimensional feature data set is output.

5. The method for identifying interference sources of a speed sensor based on the GWO-CFDP algorithm according to claim 4 is characterized in that: Step S23 specifically includes: S231, using a Gaussian kernel function to nonlinearly map the low-dimensional feature data set to a high-dimensional space to obtain a high-dimensional feature data set; Among them, the low-dimensional feature dataset is defined as: Y={y1,y2,…,y π },Y∈n; The nonlinear mapping relationship is: φ:Y→F; The high-dimensional feature dataset is: Φ(y)=(Φ(y1),Φ(y2),…,Φ(y i ),…,Φ(y q )),Φ(y)∈m,m <n; Among them, n and m represent dimensions, Y represents a low-dimensional feature dataset, y1, y2, y π Respectively represent the first, second, and π low-dimensional feature data in the low-dimensional feature data set, → represents the mapping relationship, F represents the high-dimensional space, φ and Φ(y) both represent high-dimensional feature data sets, Φ(y1), Φ(y2), Φ(y i )、Φ(y q ) represent the first, second, i, and q high-dimensional feature data in the high-dimensional data set respectively; S232. Calculate the covariance matrix of the high-dimensional feature data set, that is: Where C represents the covariance matrix of the high-dimensional feature data set, T represents the transpose, and q represents the total number of high-dimensional feature data; S233, calculate the eigenvalues ​​and eigenvectors of the covariance matrix of the high-dimensional feature data set, that is: Among them, λ i represents the eigenvalue of the i-th high-dimensional feature data, P i ′ represents the feature vector of the i-th high-dimensional feature data, a i Represents the linear coefficient matrix of the i-th high-dimensional feature data; S234, based on the eigenvector of the covariance matrix of the high-dimensional feature data set, extract the nonlinear principal components of the low-dimensional feature data set projected in the high-dimensional space, that is: Among them, α i Represents the variance contribution rate of the i-th high-dimensional feature data; S235. Based on the eigenvalues ​​and eigenvectors of the covariance matrix of the high-dimensional feature data set, the reduced-dimensional feature data set and the cumulative variance contribution rate are obtained, namely: Among them, Y′ represents the feature data set with reduced dimension, and L represents the cumulative variance contribution rate; S236 , selecting a reduced-dimensionality feature data set whose cumulative variance contribution rate is greater than a preset extraction rate as the final reduced-dimensionality feature data set.

6. The method for identifying interference sources of a speed sensor based on the GWO-CFDP algorithm according to claim 5, characterized in that: Step S3 specifically includes: S31. According to the final dimension reduction feature data set, an initial cutoff distance is set, and the initial cutoff distance is optimized by using the gray wolf algorithm to obtain an optimal cutoff distance; S32. According to the optimal cutoff distance and the final reduced dimension feature data set, the local density and the minimum distance of the data points in the final reduced dimension feature data set are calculated, that is: in, represents the local density of data point i″″, e represents the exponential function, d i″′j′′ represents the Euclidean distance from data point i′′ to data point j′′, d c represents the cutoff distance, δ i′′ represents the minimum distance from data point i″′ to data point j′′, represents the local density of data point j″, min represents the minimum value, represents the maximum local density; S33, normalizing the local density and the minimum distance of the data points to obtain the normalized local density and the minimum distance of the data points; S34. Based on the normalized local density and minimum distance of the data points, the decision index of the data points is calculated, namely: in, represents the normalized local density of data point i′′, represents the minimum distance of normalized data points, γ i′′ represents the decision indicator of data point i″, α1 represents the compensation parameter; S35, after normalizing the data points and the decision indicators of the data points, a decision diagram of the decision indicator values ​​of the normalized data points is drawn with the normalized data points as the horizontal axis and the decision indicators of the normalized data points as the vertical axis; S36, calculating the vertical distance from each data point to the straight line according to the decision graph of the decision index value of the normalized data point, and selecting the point with the largest vertical distance as the knee point; Among them, the straight line is the connection between the first data point and the last data point in the decision graph of the decision indicator value of the normalized data point; S37, selecting w data points before the knee point as cluster centers, and assigning unassigned data points to the category where the nearest high-density point is located, to obtain classified clusters; Among them, each cluster represents a disturbance condition.

7. The method for identifying interference sources of a speed sensor based on the GWO-CFDP algorithm according to claim 6, characterized in that: Step S31 specifically includes: S311, initialize the parameters of the gray wolf group, specifically: Set the number of gray wolves and the boundary of the gray wolf search space; The maximum and minimum values ​​of the data points in the final dimension reduction feature data set are used as the upper and lower bounds of the gray wolf search space; Randomly initialize the position of each gray wolf, that is: in, represents the initial position of the gray wolf j2 in dimension i2, L i2 represents the lower bound of the gray wolf search space initially set under dimension i2, rand represents a random function used to generate a random number between 0 and 1, and U j2 , L j2 They represent the search range of the gray wolf j2 in a certain dimension respectively; S312, dividing the social hierarchy, that is, using the Davidson-Bodin index as a fitness function to calculate the distance between each gray wolf and its prey, and sorting the distance between each gray wolf and its prey from small to large, and classifying the three gray wolves closest to the prey as α′ wolf, β′ wolf, and δ′ wolf, and the other gray wolves as ω′ wolf; Among them, using the Davidson-Bodin index as the fitness function, the formula for calculating the distance between each gray wolf and its prey is: Among them, DBI represents the distance between each gray wolf and its prey, K represents the number of categories, max represents the maximum value, S i′ represents the standard error between the data point and the cluster center in the i′th cluster, S j′ represents the standard error between the data point and the cluster center in the j′th cluster, d i′j′ represents the distance between the cluster centers of the i′th cluster and the j′th cluster, Z i′ represents the number of data points in the i′th cluster, v i′ represents the cluster center, and x1 represents a single data point within the cluster. S313, surround the prey, that is, when the position of the prey no longer changes, update the position of the gray wolf group based on the nonlinear decreasing convergence factor, that is: in, represents the distance between the gray wolf and its prey. represents the position of prey P at iteration number t, denotes the coefficient vector, represents the position of the gray wolf at iteration number t, represents the convergence factor of nonlinear decrease, All represent random vectors, T′ represents the maximum number of iterations, and exp represents the empirical function; S314, group hunting stage, that is, based on the position of the gray wolf group, the α′ wolf first determines the hunting target, and then the β′ wolf and the δ′ wolf lead the other ω′ wolves to surround the hunting target, and at the same time continuously obtain the position of the prey to update the position of the α′ wolf, the β′ wolf, and the δ′ wolf, that is: in, Respectively represent the distance between α′ wolf, β′ wolf, δ′ wolf and the prey, are coefficient vectors, They represent the positions of α′ wolf, β′ wolf, and δ′ wolf at iteration number t, respectively. They represent the positions of α′ wolf, β′ wolf, and δ′ wolf after being disturbed; Based on the updated positions of α′ wolf, β′ wolf, and δ′ wolf, in the next iteration, the position of the gray wolf group is adaptively updated, that is: in, represents the position of the j3rd gray wolf at iteration t+1; At the same time, according to the importance of α′ wolf, β′ wolf, and δ′ wolf, dynamic weights are introduced to readjust the position of the prey, namely: Among them, X P (t+1) represents the position of the prey P at iteration number t+1, represents the weight of the j3rd gray wolf, represents the position of the j3rd gray wolf, X α′ , X β′ , X δ′ Respectively represent the positions of α′ wolf, β′ wolf, and δ′ wolf, W α′ , W β′ , W δ′ They represent the dynamic weights of α′ wolf, β′ wolf, and δ′ wolf respectively; S315, when When , the gray wolf group will attack the prey and take the current position of the wolf α′ as the optimal cutoff distance. When , return to step S313 to perform iterative calculation until the maximum number of iterations is reached or the optimal truncation distance is found.

Images