Railcar door system anomaly discrimination method and system based on multi-resolution windows

Abstract

The application discloses a kind of track car door system abnormality discrimination method and system based on multi-resolution window, the car door operation data of each line is divided into source domain and target domain, different length windows are set to the car door operation data is windowed and extracted and first feature extraction is carried out;The first feature is sorted by importance, and a certain number of second features are screened out;The second feature is fused and the third feature common in the source domain and the target domain is screened out;The third feature is used for classification model training, and the abnormal state discrimination result of each sample in the target domain is obtained;The application weakens the time scale sensitivity of feature extraction caused by single window size, and improves the model accuracy by screening out redundant features that have little or no contribution to the accuracy of model prediction results, improves the generalization performance and applicability of the established model on different individual car door system data.

Description

TECHNICAL FIELD

[0001] The application relates to an abnormality discrimination method and system for a railcar door system, in particular to an abnormality discrimination method and system for a railcar door system based on a multi-resolution window. BACKGROUND

[0002] The door system is one of the most important subsystems in a rail transit system, and its health state directly affects the normal operation of the train and the personal safety of passengers. Therefore, accurately discriminating the abnormal type of the door system and timely repairing and maintaining the door system are of great significance to improving the safety and reliability of the door system. CN202010326103.2 discloses a device fault diagnosis method based on multi-source monitoring data fusion, which performs data segmentation processing on the multi-source monitoring data based on multiple different time window lengths, and directly inputs the features extracted from the data in different windows into a fault diagnosis model. However, the number of features obtained by calculating the data in multiple different windows is very large, and there are a large number of redundant features, which will increase the computational load of the model and negatively interfere with the accuracy of the final model. SUMMARY

[0003] The purpose of the application is to provide an abnormality discrimination method and system for a railcar door system based on a multi-resolution window, which has strong generalization ability and high accuracy.

[0004] The abnormality discrimination method for a railcar door system based on a multi-resolution window divides the door operation data of each line into a source domain and a target domain, sets different length windows to window the door operation data and perform first feature extraction;

[0005] The first features are sorted by importance, and a certain number of second features are selected; the second features are fused and a third feature common to the source domain and the target domain is selected;

[0006] The third feature is used for classification model training to obtain an abnormal state discrimination result of each sample in the target domain.

[0007] Further, the first feature sorting by importance and selecting a certain number of second features include: using a random forest algorithm to sort the first features of the source domain and the target domain by importance, and selecting a certain number of second features according to the importance order.

[0008] Further, setting different length windows to window the door operation data includes: windowing the door operation data and its baseline curve, the baseline curve being an average value sequence of the door operation data in a normal operating state.

[0009] Further, the first feature includes a DTW distance between the door operation data and its reference curve, and a time domain feature of the door operation data.

[0010] Further, the classification model training using the third feature to obtain the abnormal state discrimination result of each sample in the target domain includes: using one-hot encoding to encode the normal and abnormal type names as the label of each type, and combining the third feature of the source domain and the target domain samples with the corresponding label to form a modeling data set; using the KNN algorithm to train the classification model using the source domain and the target domain as the training set and the test set respectively, to obtain the abnormal state discrimination result of each sample in the test set.

[0011] Further, before the classification model training using the third feature to obtain the abnormal state discrimination result of each sample in the target domain, the method further includes: using the SMOTE algorithm to expand the minority class samples in the source domain and the target domain, and using the balanced distribution adaptation method to calculate the maximum mean difference for data distribution alignment.

[0012] Further, before the setting different length windows to window the door operation data and extract the first feature, the method further includes: preprocessing the door operation data, eliminating abnormal length data and unifying the data length.

[0013] The abnormal discrimination system of the rail door system based on the multi-resolution window includes:

[0014] A data acquisition and preprocessing unit is configured to divide the door operation data of each line into a source domain and a target domain.

[0015] A feature extraction unit is configured to set different length windows to window the door operation data and extract the first feature, sort the importance of the first feature, and select a certain number of second features; and perform feature fusion on the second features and select third features common to the source domain and the target domain.

[0016] An abnormal discrimination unit is configured to train a classification model using the third feature to obtain the abnormal state discrimination result of each sample in the test set.

[0017] The electronic device includes a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the computer program is loaded into the processor, the method for discriminating the abnormality of the rail door system based on the multi-resolution window is implemented.

[0018] The computer readable storage medium of the present application stores a computer program, wherein the computer program is executed by a processor to implement the abnormality discrimination method of the rail car door system based on the multi-resolution window.

[0019] Advantages: Compared with the prior art, the present application has the following advantages: (1) using multiple windows of different sizes to truncate data and extract features from different scales, weakening the time scale sensitivity of feature extraction caused by a single window size, and enhancing the identification and matching performance of key sensitive intervals such as the closing section and the locking section during the operation of the car door; (2) further screening and fusing the features extracted by the multi-resolution window, screening out redundant features that have little or no contribution to the accuracy of the model prediction result, retaining the most valuable features, improving the ability of the constructed features to reflect different abnormal modes of the car door system, thereby optimizing the feature set input to the classification model and improving the accuracy of the model; (3) through the feature unification of the source domain and the target domain feature set, improving the generalization performance and applicability of the established model on different individual car door system data, and improving the accuracy of the model in practical application. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 The abnormality discrimination method flowchart of the present application.

[0021] Figure 2 The abnormality discrimination result comparison chart in the embodiment of the present application. DETAILED DESCRIPTION

[0022] The technical solutions of the present application will be further described below with reference to the accompanying drawings.

[0023] As shown in Figure 1 The multi-resolution window rail car door system abnormality discrimination method of the present application includes the following steps:

[0024] Step 1, data acquisition.

[0025] Obtain rail car door operation data of different lines, which includes normal operation data and abnormal operation data, and the abnormal operation data includes four types of abnormalities, i.e. V-shaped size abnormality, centering size abnormality, lower blocking pin interference abnormality, and pressure roller abnormality. Extract the motor current, rotation angle and rotation speed signal data in each type of data when the door is opened and closed, and divide the car door system of each line into a source domain and a target domain.

[0026] Step 2, data preprocessing.

[0027] Remove the data with abnormal length in the collected signal data; unify the length of the collected signal data.

[0028] Step 3, reference curve calculation.

[0029] The average value sequence of the opening and closing door current, rotation angle and rotation speed sequence of all samples of the vehicle door in the normal operation state after the source domain and target domain data are preprocessed is taken as the reference curve of the opening and closing door current, rotation angle and rotation speed of the vehicle door in the source domain and target domain respectively. The calculation formula of the average value sequence is:

[0030]

[0031] wherein (z1, z2, …, z f ) is the reference curve, z pq represents the value of the qth point of the pth normal sample sequence, f is the sample sequence length, and g is the total number of normal samples.

[0032] Step 4, windowing of the multi-resolution window data.

[0033] According to the length of the signal data, windows of different lengths are set to window and intercept the opening and closing door current, rotation angle and rotation speed signal data of the rail vehicle door in the source domain and target domain after preprocessing and the reference curve, and the windows do not overlap.

[0034] Step 5, DTW distance feature and time domain feature extraction.

[0035] The DTW distance between the opening and closing door current, rotation angle and rotation speed data of the source domain and target domain vehicle door in the normal and four abnormal operation states after preprocessing and the respective reference curves and the mean value, peak value, variance and time center of gravity features of the opening and closing door current, rotation angle and rotation speed data are calculated when the data is intercepted by windows of different lengths. The calculation method of the DTW distance is:

[0036] 1) Given the windowed and intercepted signal data sequence K = [k1, k2, …, k y ] and the corresponding windowed and intercepted reference curve sequence L = [l1, l2, …, l y ];

[0037] 2) The Euclidean distance between the value of each point in the signal data sequence and the value of each point in the reference curve sequence is calculated, and a y x y matrix D is formed according to the calculated distance values, wherein the (v, w)th element corresponds to the Euclidean distance between k v and l w ;

[0038] 3) Starting from the starting point D 00 of the matrix and ending at the end point D uu , the path with the smallest sum of the matrix is searched. The sum of the path elements is the DTW distance between the signal and the corresponding reference curve. The formula is:

[0039]

[0040] where dis m represents the distance element corresponding to the mth point on the path, and M represents the total number of elements on the path;

[0041] The calculation formulas of the mean value, peak value, variance, and time center of gravity are as follows:

[0042] Mean value:

[0043] Peak value: K peak = max |K|;

[0044] Variance:

[0045] Time center of gravity:

[0046] where t d is the time, and k d is the absolute value of the signal amplitude corresponding to the time.

[0047] Step 6, random forest feature screening.

[0048] The features extracted by different length windows from the source domain data and a small amount of labeled data of the target domain are sorted in terms of feature importance using the random forest algorithm, and the features are screened. The random forest feature importance sorting process is as follows:

[0049] 1) Use the bootstrap resampling technique to extract b samples from the data set, and establish b decision tree models;

[0050] 2) Calculate the out-of-bag data error errOOB n1 corresponding to the b extracted sample data sets, respectively;

[0051] 3) Add random noise interference to the b extracted sample data sets, and recalculate the out-of-bag data error errOOB n2 ;

[0052] 4) Calculate the feature importance according to the following formula:

[0053]

[0054] where E represents the number of trees;

[0055] 5) Sort IM b in descending order, that is, the feature importance, and the larger the value, the more important the feature.

[0056] Step 7, feature fusion.

[0057] The features screened out by the window length of 5 and 10 are fused for the source domain and the target domain;

[0058] Step 8, feature unification.

[0059] Unify the features of the source domain and the target domain, that is, select the common features in the source domain and the target domain from the fusion features;

[0060] Step 9, data class balancing.

[0061] The SMOTE algorithm is used to analyze and simulate the few-class samples, and the artificially simulated new samples are added to the data set, so that the class imbalance in the original data is no longer serious. Through the SMOTE algorithm, the number of less-class samples in the data set after the above processing is increased for the source domain data and the target domain few-labeled data, so that the number of all types of samples is balanced as much as possible. The calculation process of the SMOTE algorithm is as follows:

[0062] 1) For each sample a in each less-class, calculate the Euclidean distance from it to all samples in the corresponding class sample set;

[0063] 2) For each sample a, select several nearest samples a of the same type according to the calculated Euclidean distance h ;

[0064] 3) For each selected sample a h , select a random number between 0 and 1, multiply the difference between the sample a and the selected sample a h , and then add a to obtain the newly generated sample, which is expressed by the formula:

[0065] a new =a+rand(0,1)×|a-a h |

[0066] Where a h is the selected sample, a new is the generated new sample, and a is the original sample.

[0067] Step 10, balanced distribution adaptation.

[0068] The balanced distribution adaptation method is used to align the data distribution of the feature data of the source domain and the target domain obtained after the above processing, and to reduce the distribution difference between the feature data of the source domain and the target domain. The balanced distribution adaptation expression is:

[0069]

[0070] Where MMD(A T x s ,A T x t ) is the balanced distribution adaptation MMD(maximum mean difference) distance, represent a source domain data set, represent a target domain data set, s , n t respectively represent the number of samples of the source domain and the target domain, ||·|| H represents a reproducing Hilbert space, A is a transformation matrix, and u is a balancing factor.

[0071] Step 11, model training test.

[0072] The normal and four abnormal type names are encoded using one-hot encoding as the label of each type. The source domain and target domain sample features are combined with the corresponding labels to form a modeling data set; the source domain and target domain are used as the training set and the test set respectively, and the KNN algorithm is used for classification model training to obtain the state discrimination result of each sample in the test set, i.e. the target domain.

[0073] The abnormal discrimination system of the rail car door system based on the multi-resolution window according to the application comprises:

[0074] A data acquisition and preprocessing unit is configured to divide the car door operation data of each line into a source domain and a target domain,

[0075] A feature extraction unit is configured to set different length windows to window and intercept the car door operation data and perform first feature extraction, sort the first features according to importance, and select a certain number of second features; and perform feature fusion on the second features and select third features common to the source domain and the target domain.

[0076] An abnormal discrimination unit is configured to train a classification model using the third features to obtain an abnormal state discrimination result of each sample in the test set.

[0077] The electronic device according to the application comprises a memory, a processor, and a computer program stored on the memory and executable on the processor, and the computer program realizes the abnormal discrimination method of the rail car door system based on the multi-resolution window when loaded into the processor.

[0078] The computer readable storage medium according to the application stores a computer program, and the computer program realizes the abnormal discrimination method of the rail car door system based on the multi-resolution window when executed by a processor.

[0079] The computer readable storage medium can include RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, flash memory or any other medium that can be used to store desired program codes in the form of instructions or data structures and can be accessed by a computer.

[0080] The processor is used for executing the computer program stored in the memory to realize each step in the method related to the above-mentioned embodiments.

[0081] The application will be further described below in combination with specific experiments.

[0082] Step 1, obtaining railcar door operation data.

[0083] Collect the door operation data of 5 different subway lines, which contains normal operation data and abnormal operation data, and the abnormal operation data includes 4 types of abnormality, i.e. V-shaped size abnormality, centering size abnormality, lower blocking pin interference abnormality and pressure wheel abnormality.

[0084] Extract the current, angle and speed signal data at the opening and closing of each type of data, and divide the railcar door system of different lines into source domain and target domain.

[0085] Step 2, data set pre-processing.

[0086] Remove the abnormal sequence data with length less than 300 and greater than 450 in the opening and closing speed data, and unify the sequence data length to 310.

[0087] Step 3, reference curve calculation.

[0088] Calculate the reference curves of the current, angle and speed of the opening and closing of the door in the source domain and the target domain:

[0089] Step 4, data windowing.

[0090] According to the signal data length 310, set the window length to 5 and 10 respectively to window and intercept the opening and closing current, angle and speed signal data of the pre-processed door operation data of the source domain and the target domain and the reference curves, and the windows do not overlap.

[0091] Step 5, DTW distance feature and time domain feature extraction.

[0092] Calculate the DTW distance between the opening and closing current, angle and speed data of the pre-processed normal and 4 types of abnormal operation data of the source domain and the target domain and their respective reference curves when the window length is 5 and 10 respectively, as well as the mean value, peak value, variance and time center of gravity features of the opening and closing current, angle and speed data. Among them, when the window length is 5, a total of 1860 features are extracted, and when the window length is 10, a total of 930 features are extracted.

[0093] Step 6, random forest feature screening.

[0094] Feature importance ranking is performed on the features extracted from the source domain data and the small amount of labeled data of the target domain using a random forest algorithm with window lengths of 5 and 10. The top 20 features in the source domain and the target domain are selected according to the feature importance ranking.

[0095] Step 7, feature fusion.

[0096] The features selected from the source domain and the target domain with window lengths of 5 and 10 are fused to obtain 40 features.

[0097] Step 8, feature unification.

[0098] The features of the source domain and the target domain are unified, and 14 features common to the source domain and the target domain are selected from the 40 features.

[0099] Step 9, data class balancing.

[0100] The number of samples of the less common classes in the data set after the above processing is increased for the source domain data and the small amount of labeled data of the target domain, so that the number of samples of all types is as balanced as possible.

[0101] Step 10, balanced distribution adaptation:

[0102] The balanced distribution adaptation method is used to align the data distribution of the feature data of the source domain and the target domain obtained after the above processing, and to reduce the distribution difference between the source domain and the target domain data.

[0103] Step 11, model training and testing.

[0104] The normal and four types of abnormal names are encoded using one-hot encoding as the label of each type. The source domain and target domain sample features are combined with the corresponding labels as input samples. The source domain and target domain samples are used as training set and test set respectively. The training set and test set data are normalized according to the following formula:

[0105]

[0106] where X norm is the normalized value, X max , X min is the maximum and minimum value of the data before normalization, and X is the original numerical value.

[0107] The KNN algorithm is used for classification model training. The state discrimination results of each sample in the test set are obtained, and the present application is compared with the method using window length 5 and 10 alone. The comparison results are shown in Figure 2 It can be seen that the accuracy of the present application is significantly improved after using multi-resolution window for feature extraction and feature selection, fusion and unification.

Claims

1. A method for discriminating abnormality of a railcar door system based on a multi-resolution window, characterized by, The door operation data of each line is divided into a source domain and a target domain, different length windows are set to window the door operation data and perform first feature extraction; The first features are sorted by importance, a certain number of second features are screened out, the second features are fused and third features common to the source domain and the target domain are screened out; The third features are used for classification model training to obtain an abnormal state discrimination result of each sample in the target domain.

2. The multi-resolution window based railcar door system anomaly discrimination method of claim 1, wherein, The first features are sorted by importance, a certain number of second features are screened out, including: using a random forest algorithm to sort the first features of the source domain and the target domain by importance, and screening out a certain number of second features according to the importance order.

3. The multi-resolution window based railcar door system anomaly discrimination method of claim 1, wherein, The different length windows are set to window the door operation data, including: windowing the door operation data and its baseline curve, the baseline curve being an average value sequence of the door operation data in a normal operation state.

4. The multi-resolution window based railcar door system anomaly discrimination method of claim 3, wherein, The first features include a DTW distance between the door operation data and its baseline curve, and a time domain feature of the door operation data.

5. The multi-resolution window based railcar door system anomaly discrimination method of claim 1, wherein, The third features are used for classification model training to obtain an abnormal state discrimination result of each sample in the target domain, including: using one-hot encoding to encode normal and abnormal type names as labels of each type, combining the third features of the source domain and the target domain samples with the corresponding labels to form a modeling data set; using a KNN algorithm for classification model training with the source domain and the target domain as a training set and a test set respectively to obtain an abnormal state discrimination result of each sample in the test set.

6. The multi-resolution window based railcar door system anomaly discrimination method of claim 1, wherein, Before the third features are used for classification model training to obtain an abnormal state discrimination result of each sample in the target domain, there is also including: using a SMOTE algorithm to expand the minority class samples in the source domain and the target domain, and using a balanced distribution adaptation method to calculate the maximum mean difference for data distribution alignment.

7. The multi-resolution window based railcar door system anomaly discrimination method of claim 1, wherein, Before the different length windows are set to window the door operation data and perform first feature extraction, there is also including: preprocessing the door operation data to eliminate length-abnormal data and unify the data length.

8. A multi-resolution window based railcar door system anomaly discrimination system, characterized by, Including: A data acquisition and preprocessing unit for dividing the door operation data of each line into a source domain and a target domain, A feature extraction unit for setting different length windows to window the door operation data and perform first feature extraction, sorting the first features by importance, screening out a certain number of second features, fusing the second features and screening out third features common to the source domain and the target domain; An abnormality discrimination unit for using the third features for classification model training to obtain an abnormal state discrimination result of each sample in the test set.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The computer program is loaded into the processor to implement the abnormality discrimination method of the rail door system based on the multi-resolution window according to any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. The computer program is executed by the processor to implement the abnormality discrimination method of the rail door system based on the multi-resolution window according to any one of claims 1-7.

Images