A machine learning-based pantograph and catenary fault detection method

Abstract

The application provides a train pantograph fault detection method based on machine learning, comprising: obtaining pantograph historical monitoring data, clustering the historical monitoring data, labeling each cluster sample data set, and labeling the fault type of each cluster sample data set; forming an initial training database by summarizing, training a fault classification model using the initial training database, and obtaining an initial fault classification model; identifying a newly collected monitoring sample using the fault classification model, checking whether the identification result is correct; if not, correcting the category of the current monitoring sample and storing it in an optimized training database; training the current fault classification model again using the optimized training database to obtain optimized model parameters, saving the fault classification model when the accuracy meets the requirements, and ending the model training. The fault classification model trained by the method can identify more comprehensive fault types and is suitable for safety detection of train pantographs.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of fault detection, in particular to a train pantograph-catenary fault detection method based on machine learning. BACKGROUND

[0002] The train pantograph-catenary system (train pantograph-catenary system) as a train traction device, its operation state directly affects the safety of train operation, for fault diagnosis of the pantograph-catenary system, mainly includes the following ways:

[0003] 1) Artificial maintenance method: the maintenance personnel need to climb, and judge the fault type by observing the state of the pantograph. This maintenance method is not only low in efficiency, but also needs to rely on the past experience of the maintenance personnel to further locate the fault cause. For the maintenance personnel with little experience, the result of fault diagnosis often cannot guarantee accuracy. In addition, the artificial detection method needs to be carried out in the parking maintenance stage, which leads to the fact that this method cannot realize real-time fault diagnosis and analysis of the pantograph, and the timely fault maintenance may directly affect the safe operation of the train.

[0004] 2) Visual detection method: such as patent document CN113859312A proposes a pantograph fault alarm method based on vehicle-mounted PHM; it collects pantograph images through image acquisition equipment, analyzes and compares image information, judges the real-time appearance information of the pantograph, and judges the fault type. This method can solve the problem of low efficiency of artificial detection, but it needs to accurately locate the structural features of the pantograph. If there are many disturbances near the pantograph, or in rainy weather, the quality of the collected images is poor, and the accuracy of fault recognition will be greatly reduced, leading to misjudgment of the fault.

[0005] 3) Deep learning-based method: such as patent document CN112991318A proposes a high-speed train pantograph fault detection method, which identifies whether there is foreign matter intrusion in the pantograph through a fault classification model. This method has the following disadvantages:

[0006] First, the fault diagnosis type is limited, and only whether there is a fault can be judged, and the fault type cannot be further analyzed. The fault diagnosis positioning type has insufficient granularity, and thus cannot provide effective maintenance guidance for train pantograph maintenance;

[0007] Second, when training the fault classification model, the data samples need to be obtained through artificial marking, that is, each sample data is identified and labeled. Such processing method is low in efficiency and has large fault label classification error.

[0008] Third, the lack of model optimization process, the model training result is limited to the labeled fault type. For practical application process, the newly added fault type, the fault type not appeared in the sample or the fault type rarely appeared lack of effective training, and then lead to fault classification model in the later use misjudgment. SUMMARY

[0009] In order to solve the above technical problems, the present application provides a kind of based on machine learning train pantograph fault detection method, and the fault classification model training process of this method is fast and effective;The fault classification model trained can identify more comprehensive fault type, can assist train operation personnel to obtain accurate fault type, and is suitable for the safety detection of train pantograph.

[0010] Technical scheme as follows:

[0011] A kind of based on machine learning train pantograph fault detection method, the temperature data of each monitoring point under the use state of train pantograph is collected, which is input into the fault classification model that has been trained to detect, and the model outputs whether train pantograph has fault and the type of fault;

[0012] The fault classification model is trained using the following steps:

[0013] S1, obtain the historical monitoring data of the same model train pantograph as the train pantograph to be detected, and the historical monitoring data includes multiple groups of monitoring samples under the use state of pantograph;

[0014] Each group of monitoring samples includes the position coordinates of each monitoring point on the train pantograph and the temperature data corresponding to each point;

[0015] S2, cluster the historical monitoring data to divide it into fault data set and no fault data set;

[0016] S3, further cluster the fault data set to divide it into multiple cluster sample data sets, and label each cluster sample data set and mark the fault type of each cluster sample data set;

[0017] S4, the no fault data set and each labeled fault type data set are summarized to form an initial training database, and the fault classification model is trained using the initial training database to obtain an initial fault classification model;

[0018] S5, collect monitoring samples under the use state of train pantograph;

[0019] The fault classification model is used to identify the newly collected monitoring samples, and the identification result is output, which is no fault or fault type;

[0020] S6, check whether the identification result is correct:

[0021] If correct, the number of identification times is added by one, and the number of correct times is added by one;

[0022] If incorrect, the number of identification times is added by one; meanwhile, the category of the current monitoring sample is corrected, and it is stored to the optimized training database;

[0023] S7, judge whether the number of identification times is less than a preset number of times:

[0024] If less, directly jump to execute step S5;

[0025] If not less, calculate the proportion of the number of correct times in the number of identification times:

[0026] If the proportion is lower than a preset proportion, use the optimized training database to retrain the current fault classification model to obtain optimized model parameters, use the optimized model parameters to update the fault classification model, clear the number of identification times and the number of correct times, and jump to execute step S5;

[0027] If the proportion is not lower than the preset proportion, save the current fault classification model, and the model training is ended.

[0028] Further, the historical monitoring data is clustered and divided into a fault data set and a non-fault data set in the following manner:

[0029] 5% to 20% of the monitoring samples in the historical monitoring data are selected as a sampling sample set, and the remaining monitoring samples are recorded as a data set A;

[0030] The sampling sample set is clustered to obtain a plurality of cluster sample sets, the cluster sample set with the largest number of samples is recorded as a non-fault data set, and the other sets are recorded as fault data sets;

[0031] A single monitoring sample in the data set A is extracted as a to-be-processed sample, the sum of the Euclidean distances between the to-be-processed sample and each monitoring sample in the non-fault data set is calculated, if the sum is less than a threshold value, the to-be-processed sample is stored in the non-fault data set, otherwise it is stored in the fault data set;

[0032] Each monitoring sample in the data set A is traversed, and each monitoring sample is stored in the non-fault data set and the fault data set in the same manner.

[0033] Further, the fault data set is clustered again, divided into a plurality of cluster sample data sets, and each cluster sample data set is labeled in the following manner:

[0034] The fault data set is clustered to divide it into a plurality of cluster sample data sets, and each cluster sample data set is processed as follows:

[0035] Sampling 5% to 20% of the monitoring samples in the clustered sample dataset, and labeling the fault type of each sampled monitoring sample;

[0036] If the proportion of the number of monitored samples with the same fault type in the total number of sampled monitored samples exceeds the threshold B, then the label of the clustered sample dataset is recorded as the same fault type, and all monitored samples in the dataset are labeled as the fault type; the threshold B is 50% to 90%;

[0037] Otherwise, if the current clustered sample dataset is a set of rare fault samples, it is necessary to judge each monitoring sample in the clustered sample dataset separately and label the corresponding fault type for each monitoring sample.

[0038] Preferably, the clustering method is a density-based clustering algorithm; in step S2, the parameters are set as follows: neighborhood radius 0.2 to 0.8, minimum number of points 0.05 to 0.1;

[0039] In step S3, the parameters are set as follows: neighborhood radius 0.05 to 0.1, minimum number of points 0.005 to 0.02.

[0040] Preferably, in step S1, the historical monitoring data is obtained through the following methods:

[0041] Throughout the entire life cycle of the train pantograph-catenary system, temperature data at various monitoring points under the operating conditions of one or more train pantograph-catenary systems are continuously observed at preset sampling intervals, and the data collected at each sampling time point are summarized to form a historical monitoring sample.

[0042] The preset sampling interval ranges from 1 min to 60 min.

[0043] The pantograph-catenary system of the train under continuous observation is the same model as the pantograph-catenary system to be tested;

[0044] The monitoring points are: 10 to 2000 points pre-selected on the pantograph wire mesh, which are evenly distributed on the pantograph wire mesh or distributed in the key monitoring areas of the pantograph wire mesh.

[0045] Furthermore, the method for collecting temperature data at each monitoring point under the condition of the train pantograph-catenary system is as follows: temperature detection sensors are installed at the monitoring points, or thermal imagers or infrared imagers are fixed on the roof of the train to take pictures of the monitoring points on the train pantograph-catenary system from an upward angle.

[0046] Furthermore, the temperature data of each monitoring point in the monitoring sample were normalized to obtain the normalized temperature data tnormalized. ij :

[0047]

[0048] Among them, tij temperature data of the jth monitoring point in the ith monitoring sample, t i最大值 maximum temperature data in the ith monitoring sample, t i最小值 minimum temperature data in the ith monitoring sample.

[0049] Preferably, in step S4, before training the fault classification model, the monitoring samples of the fault types known artificially but not collected are also stored into the initial training database.

[0050] Further, the monitoring sample types contained in the initial training database and the optimized training database include: no fault, abnormal wear of the pantograph, crack in the middle section of the carbon slide plate, carbon slide plate chunk, carbon slide plate arc, pantograph lifting failure, electrical fault, absence of the goat horn, abnormal wear of the contact network, bending of the horn, pantograph center line deviation too large, contact line fracture and pantograph overall absence.

[0051] Preferably, in step S1, the historical monitoring data is pre-processed, including:

[0052] The monitoring samples with missing data are removed, the monitoring samples with repeated storage are removed, and the monitoring samples with format errors are modified or removed;

[0053] The historical monitoring data is processed by dimension reduction using the principal component analysis method;

[0054] In step S7, the preset number = k x M, wherein k = 0.15-0.3, and M = the total number of monitoring samples in the initial training database; the preset ratio is 0.9-0.98.

[0055] Compared with the existing technology, the present application has the following characteristics:

[0056] (1) The method can not only determine whether the pantograph and catenary fault exists, but also further analyze the fine-grained fault type, has high recognition speed and high accuracy, provides a reliable reference for intelligent operation and maintenance of the train, ensures safe operation of the train, and effectively saves manpower and material resources.

[0057] (2) Normally, the temperature values of each monitoring point on the pantograph-catenary system should be within the corresponding temperature range. In particular, the position where the pantograph contacts the catenary is in a high-temperature state, with a maximum temperature of 200°C. However, when the pantograph-catenary system is abnormal, the temperature of some local monitoring points will be abnormal. For example, when the pantograph is abnormally worn, the temperature of the monitoring points near the abnormally worn position of the slide plate will be higher than the normal temperature range. When there is a crack in the middle section of the carbon slide plate, the temperature of the monitoring points in the middle section of the carbon slide plate will be higher than the normal temperature range. When there is an abnormal drop in the carbon slide plate, the temperature in the drop area of the carbon slide plate fluctuates greatly. When there is arcing in the carbon slide plate, the temperature in the contact area between the pantograph slide plate and the high-voltage power grid is abnormally high. When there is a pantograph raising failure, the temperature distribution in the pantograph raising area does not match that in the normal pantograph raising area. When there is an electrical fault, the temperature of the monitoring points near the fault lacks a progressive distribution trend. When there is a missing pantograph horn, the normal temperature distribution in the pantograph horn area is missing. When the pantograph horn is bent, although there is temperature in the monitoring points near the horn, the temperature distribution trend is offset. When there is abnormal wear of the catenary, the temperature of the monitoring points near the catenary is abnormal. When the contact wire is broken, the temperature of the monitoring points near the broken area is missing, and the overall temperature distribution of the contact wire is discontinuous. When the pantograph centerline deviation is too large, some of the monitoring points on the pantograph are missing. When the pantograph is missing as a whole, the temperature of most of the monitoring points is missing.

[0058] The method trains a fault classification model based on temperature information of monitoring points and evaluates the fault type using temperature information. Since the working position under working condition is in a high-temperature state, the temperature of each monitoring point is less affected by the external environment. In order to further reduce the influence of the external environment temperature, the method also normalizes the temperature data. Compared with the pantograph-catenary fault analysis method based on image information, the method is not affected by clouds, rain, and fog, and can work all day. Moreover, the detected fault categories are more diverse: not only can it judge the structural faults of the pantograph (such as slide plate wear, horn bending, etc.), but also can diagnose and analyze internal faults of the pantograph-catenary system (such as abnormal pantograph raising, slide plate arcing, slide plate cracking, catenary wear, etc.).

[0059] (3) In the model training process, the historical monitoring samples are clustered. Only a small number of monitoring samples in each cluster sample set need to be labeled, and there is no need to identify the fault category of each monitoring sample, which improves the processing efficiency and accuracy of the monitoring samples.

[0060] The density-based clustering algorithm is more suitable for aggregating large data. Compared with other clustering algorithms, it has better clustering effect and higher precision.

[0061] (4) The initial fault classification model is optimized, which can not only identify common fault categories, but also accurately detect rare faults, making the identification types more comprehensive. BRIEF DESCRIPTION OF DRAWINGS

[0062] Figure 1 Structure block diagram for fault classification model training in the specific embodiment;

[0063] Figure 2 Process schematic for clustering and labeling historical monitoring data in the specific embodiment. DETAILED DESCRIPTION

[0064] The technical solutions of the present application are described in detail below in combination with the drawings and the specific embodiments.

[0065] For ease of understanding, the following is an exposition in the prior art regarding machine learning classification models and density-based clustering methods:

[0066] The machine learning classification model structure used in the present application is the model structure provided in the prior art, which is suitable for KNN, SVM, logistic regression, decision tree, random forest, naive Bayes, etc. Classification models. The model training process uses existing training methods.

[0067] The density-based clustering method involved in the present application is DBSCAN (Density-Based Spatial Clustering of Applications with Noise). This algorithm divides areas with sufficient density into clusters and discovers clusters of arbitrary shape in a spatial database with noise. It defines a cluster as the maximum set of density-connected points. The two algorithm parameters of the DBSCAN algorithm are neighborhood radius Eps and minimum point number MinPoints; when the number of points within the neighborhood radius Eps is greater than the minimum point number MinPoints, it is dense.

[0068] The specific scheme is as follows:

[0069] A train pantograph fault detection method based on machine learning collects temperature data of each monitoring point under the use state of the train pantograph, inputs it into the trained fault classification model for detection, and the model outputs whether the train pantograph has a fault and the type of the fault;

[0070] As shown in Figure 1 The fault classification model is trained using the following steps:

[0071] S1, obtaining historical monitoring data of a train pantograph of the same model as the train pantograph to be detected, the historical monitoring data containing multiple groups of monitoring samples under the use state of the pantograph;

[0072] Each group of monitoring samples contains the position coordinates of each monitoring point on the train pantograph and the corresponding temperature data of each point.

[0073] S2, cluster the historical monitoring data to divide it into a fault data set and a non-fault data set;

[0074] S3, re-cluster the fault data set to divide it into a plurality of cluster sample data sets, and label each cluster sample data set to mark the fault type of each cluster sample data set;

[0075] S4, the non-fault data set and each labeled fault type data set are summarized to form an initial training database, which is used to train the fault classification model to obtain an initial fault classification model;

[0076] S5, collect monitoring samples under the use state of the train pantograph;

[0077] The fault classification model is used to identify the newly collected monitoring samples, and the identification result is output, which is: no fault or fault type;

[0078] S6, check whether the identification result is correct:

[0079] If correct, the identification number is increased by one, and the correct number is increased by one;

[0080] If not correct, the identification number is increased by one; at the same time, the category of the current monitoring sample is corrected, and it is stored in the optimized training database;

[0081] S7, determine whether the identification number is less than the preset number:

[0082] If less than, directly jump to step S5;

[0083] If not less than, calculate the proportion of the correct number in the identification number:

[0084] If the proportion is lower than the preset proportion, the current fault classification model is trained again using the optimized training database to obtain optimized model parameters, which are used to update the fault classification model, and the identification number and the correct number are cleared, and step S5 is executed;

[0085] If the proportion is not lower than the preset proportion, save the current fault classification model, and the model training is completed.

[0086] Specifically, as shown in Figure 2 In order to quickly label the monitoring samples, in step S2, the historical monitoring data is clustered to divide it into a fault data set and a non-fault data set in the following manner:

[0087] In specific implementation, sampling is used to reduce data processing amount, specifically: 5% to 20% of the monitoring samples in the historical monitoring data are selected as a sampling sample set, and the remaining monitoring samples are recorded as a data set A;

[0088] The sample set is clustered to obtain a plurality of clustered sample sets, and the clustered sample set with the largest number of samples is recorded as a fault-free data set, and the other sets are recorded as fault data sets;

[0089] A single monitoring sample from the data set A is extracted as a to-be-processed sample, and the sum of the Euclidean distances between the to-be-processed sample and each monitoring sample in the fault-free data set is calculated. If the sum is less than a threshold value, the to-be-processed sample is stored in the fault-free data set, otherwise it is stored in the fault data set.

[0090] Each monitoring sample in the data set A is traversed, and each monitoring sample is stored in the fault-free data set and the fault data set by using the same judgment method.

[0091] In step S3, the fault data set is clustered again to divide it into a plurality of clustered sample data sets, and each clustered sample data set is labeled. The fault type of each clustered sample data set is labeled as follows:

[0092] The fault data set is clustered to divide it into a plurality of clustered sample data sets, and each clustered sample data set is processed as follows:

[0093] 5% to 20% of the monitoring samples in the clustered sample data set are sampled, and the fault types of the sampled monitoring samples are labeled.

[0094] If the proportion of the number of monitoring samples with the same fault type in the total number of sampled monitoring samples exceeds a threshold value B, the label of the clustered sample data set is recorded as the same fault type, and all monitoring samples in the data set are labeled as the fault type. The threshold value B is 50% to 90%.

[0095] Otherwise, the current clustered sample data set is a rare fault sample set, and each monitoring sample in the clustered sample data set needs to be judged and labeled with the corresponding fault type.

[0096] For example, if the threshold value B is 50%, if there are more than half of the monitoring samples with the same fault type, the label of the clustered sample data set is recorded as the same fault type, and all monitoring samples in the data set are labeled as the fault type. That is, the label of the clustered sample data set is labeled as the fault type of the majority of monitoring samples. At this time, there may be a small part of the monitoring samples in the clustered sample data set with inaccurate labels, which will be corrected in the subsequent model training and optimization process.

[0097] The above process uses clustering and sampling to reduce data computation and speed up the processing speed of monitoring sample labeling.

[0098] The following description uses overall pantograph-catenary fault detection as an example:

[0099] Temperature sensors can be installed at monitoring points, or thermal or infrared imagers can be mounted on the roof of the train to take pictures of the monitoring points on the train's overhead wires.

[0100] In this embodiment, a thermal imager is fixed on the roof of the train to obtain the test sample;

[0101] In step S1, historical monitoring data is obtained through the following methods:

[0102] Throughout the entire life cycle of the train pantograph-catenary system, temperature data at various monitoring points under the operating conditions of one or more train pantograph-catenary systems are continuously observed at preset sampling intervals, and the data collected at each sampling time point are summarized to form a historical monitoring sample.

[0103] The preset sampling interval ranges from 1 min to 60 min.

[0104] The pantograph-catenary system of the train under continuous observation is the same model as the pantograph-catenary system to be tested;

[0105] The monitoring points are: 10 to 2000 points pre-selected on the pantograph wire mesh, which are evenly distributed on the pantograph wire mesh or distributed in the key monitoring areas of the pantograph wire mesh.

[0106] In this embodiment, the preset sampling interval is 10 minutes, and there are 1000 monitoring points, which are evenly distributed on the pantograph net to detect the entire pantograph. The format of a single monitoring sample is: [x,y,z,t], where x,y,z are the position coordinates of the monitoring point, and t is the temperature data at that position.

[0107] To reduce the influence of external ambient temperature on the temperature at the monitoring points, the temperature data of each monitoring point in the monitoring sample were normalized to obtain the normalized temperature data tnormalized. ij :

[0108]

[0109] Among them, t ij Let t be the temperature data of the j-th monitoring point in the i-th monitoring sample. i最大值 t represents the highest temperature value in the i-th monitoring sample. i最小值 It represents the lowest temperature value in the i-th monitoring sample.

[0110] Preprocessing of historical monitoring data includes:

[0111] Remove monitoring samples with missing data; remove duplicate monitoring samples; modify or remove monitoring samples with formatting errors;

[0112] The historical monitoring data is processed by dimension reduction using principal component analysis, and a single monitoring sample [x, y, z, t] is reduced from 4 dimensions to 1 dimension.

[0113] In this embodiment, the clustering method is a density-based clustering algorithm.

[0114] In step S2, the parameters are set as follows: neighborhood radius 0.2-0.8, minimum number of points 0.05-0.1; in this embodiment, the neighborhood radius is 0.5 and the minimum number of points is 0.08.

[0115] In step S3, the parameters are set as follows: neighborhood radius 0.05-0.1, minimum number of points 0.005-0.02; in this embodiment, the neighborhood radius is 0.06 and the minimum number of points is 0.01.

[0116] In step S4, before training the fault classification model, the monitoring samples of the fault types known by humans but not collected are also stored in the initial training database. For example, special faults such as contact line fracture and pantograph whole loss. Such faults rarely occur but have obvious pantograph-catenary temperature distribution characteristics. Specifically, the monitoring samples are constructed according to the following method:

[0117] First, according to the general fault judgment standard in the industry, the monitoring points of the special fault are marked, and then the temperature values corresponding to the positions of these monitoring points are set as the temperature values appearing under the fault, and the temperatures of other positions of the pantograph-catenary are set as the temperatures under no fault.

[0118] In this embodiment, the monitoring sample types contained in the initial training database and the optimized training database include: no fault, abnormal pantograph wear, carbon slide plate middle segment crack, carbon slide plate drop, carbon slide plate arc, pantograph lifting failure, electrical fault, goat horn loss, contact net abnormal wear, bow angle bending, pantograph center line deviation too large, contact line fracture, and pantograph whole loss.

[0119] In step S7, the preset number of times = k x M, where k = 0.15-0.3 and M = the total number of monitoring samples in the initial training database; the preset ratio is 0.9-0.98. In this embodiment, k = 0.15 and the preset ratio is 0.95.

[0120] The fault classification model trained by the above method has an identification accuracy of more than 95%. Using this model to process the temperature information of each monitoring point collected during the operation of the pantograph-catenary can effectively output the current pantograph-catenary fault: no fault, abnormal pantograph wear, carbon slide plate middle segment crack, carbon slide plate drop, carbon slide plate arc, pantograph lifting failure, electrical fault, goat horn loss, contact net abnormal wear, bow angle bending, pantograph center line deviation too large, contact line fracture, or pantograph whole loss.

[0121] The trained model has high recognition speed and high accuracy, provides a reliable reference for intelligent operation and maintenance of the train, ensures safe operation of the train, and effectively saves manpower and material resources.

[0122] The foregoing description of specific exemplary embodiments of the application presented herein is not intended as an inflexible or exhaustive description of the application or the scope of the application. Obviously, many changes and modifications can be made to the exemplary embodiments, and equivalents should be resorted to, where appropriate, to the full scope of the application. The exemplary embodiments were chosen and described in order to explain the principles of the application and its practical application, to thereby enable others skilled in the art to understand the application for various exemplary embodiments and with various modifications that are suited to the particular use contemplated. It is intended that the scope of the application be defined by the following claims and their equivalence.

Claims

1. A method for detecting train pantograph fault based on machine learning, comprising the following steps: collecting temperature data of each monitoring point under the use state of the train pantograph, inputting the data into a trained fault classification model to detect whether the train pantograph has a fault and the type of the fault; the fault classification model is trained by the following steps: S1.acquiring historical monitoring data of the same model of the train pantograph to be detected, wherein the historical monitoring data comprises a plurality of monitoring samples under the use state of the train pantograph; each monitoring sample comprises the position coordinates of each monitoring point on the train pantograph and the corresponding temperature data of each point; S2.clustering the historical monitoring data to divide it into a fault data set and a non-fault data set; S3.clustering the fault data set again to divide it into a plurality of cluster sample data sets, and labeling each cluster sample data set to mark the fault type of each cluster sample data set; S4.combining the non-fault data set and each labeled fault type data set to form an initial training database, using the initial training database to train the fault classification model to obtain an initial fault classification model; S5.acquiring monitoring samples under the use state of the train pantograph; using the fault classification model to identify the newly acquired monitoring samples to output an identification result, wherein the identification result is non-fault or a fault type; S6.checking whether the identification result is correct: if yes, increasing the identification times by one and the correct times by one; if no, increasing the identification times by one, correcting the category of the current monitoring sample, and storing it in an optimized training database; S7.determining whether the identification times are less than a preset number: if yes, directly jumping to step S5; if no, calculating the proportion of the correct times in the identification times: if the proportion is lower than a preset proportion, using the optimized training database to train the current fault classification model again to obtain optimized model parameters, using the optimized model parameters to update the fault classification model, clearing the identification times and the correct times, and jumping to step S5; if the proportion is not lower than the preset proportion, saving the current fault classification model, and ending the model training. characterized in that In step S2, the historical monitoring data is clustered to divide it into a fault data set and a non-fault data set in the following manner: selecting 5%-20% of the monitoring samples from the historical monitoring data as a sample set, and the remaining monitoring samples as a data set A; clustering the sample set to obtain a plurality of cluster sample sets, taking the cluster sample set with the largest number of samples as the non-fault data set, and taking the other cluster sample sets as the fault data set; extracting a single monitoring sample from the data set A as a to-be-processed sample, calculating the sum of the Euclidean distances between the to-be-processed sample and each monitoring sample in the non-fault data set, and storing the to-be-processed sample in the non-fault data set if the sum is less than a threshold value, otherwise storing the to-be-processed sample in the fault data set; traversing each monitoring sample in the data set A, and storing each monitoring sample in the non-fault data set and the fault data set by using the same judgment manner. ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 2. The method of claim 1, wherein the method comprises: ​ ​ ​ ​ ​ 3. The method of claim 1, wherein the method further comprises: determining a fault of the pantograph based on the comparison result. In step S3, the faulty data set is clustered again to divide it into a plurality of cluster sample data sets, and each cluster sample data set is labeled. The manner of labeling the fault type of each cluster sample data set is as follows: The faulty data set is clustered to divide it into a plurality of cluster sample data sets, and each cluster sample data set is processed as follows: 5% to 20% of the monitoring samples in the cluster sample data set are sampled, and the fault types of the sampled monitoring samples are labeled; If the proportion of the same fault type monitoring samples in the total number of sampled monitoring samples exceeds the threshold B, the label of the cluster sample data set is marked as the same fault type, and all the monitoring samples in the data set are labeled as the fault type; The threshold B is 50% to 90%; Otherwise, the current cluster sample data set is a rare fault sample set, and each monitoring sample in the cluster sample data set needs to be judged and labeled with the corresponding fault type.

4. The method according to any one of claims 1-3, wherein the method comprises: The clustering method is a density-based clustering algorithm; in step S2, the parameters are set as follows: neighborhood radius 0.2 to 0.8, minimum point number 0.05 to 0.1; In step S3, the parameters are set as follows: neighborhood radius 0.05 to 0.1, minimum point number 0.005 to 0.

02.

5. The method of claim 1, wherein the method further comprises: determining a fault of the pantograph based on the comparison result. In step S1, the historical monitoring data is obtained by the following method: During the entire life cycle of the train pantograph, the temperature data of each monitoring point under one or more train pantograph use states is continuously observed at a preset sampling interval, and the data collected at each sampling time point is aggregated to form a historical monitoring sample; The preset sampling interval is 1 minute to 60 minutes; The continuously observed train pantograph is of the same type as the pantograph to be detected; The monitoring points are preselected points on the pantograph, and there are 10 to 2000 monitoring points, which are uniformly distributed on the pantograph or distributed in the key monitoring area of the pantograph.

6. The method of claim 1, wherein the method further comprises: determining a fault type of the pantograph based on the determined fault type of the pantograph and the determined fault type of the catenary. The temperature data of each monitoring point under the train pantograph use state is collected by installing a temperature detection sensor at the monitoring point, or by fixing a thermal imager or an infrared imager on the train roof and shooting the monitoring points on the train pantograph from above.

7. The method of claim 1, wherein the method further comprises: determining a fault type of the pantograph based on the determined fault type of the pantograph and the determined fault type of the catenary. The temperature data of each monitoring point in the monitoring sample is normalized to obtain normalized temperature data Wherein, t ij is the temperature data of the jth monitoring point in the ith monitoring sample, t i最大值 is the highest value of the temperature data in the ith monitoring sample, t i最小值 is the lowest value of the temperature data in the ith monitoring sample.

8. The train pantograph-catenary fault detection method based on machine learning as described in claim 1, characterized in that: In step S4, before training the fault classification model, the fault type monitoring samples that are known but not collected are also stored in the initial training database. 9.The train pantograph fault detection method based on machine learning of claim 1, wherein: The monitoring sample types included in the initial training database and the optimized training database include: no fault, abnormal wear of the pantograph, crack in the middle section of the carbon slide plate, carbon slide plate shedding, carbon slide plate arc, pantograph lifting failure, electrical failure, absence of the sheep horn, abnormal wear of the contact net, bending of the pantograph horn, pantograph center line deviation being too large, contact line fracture, and pantograph whole loss.

10. The train pantograph-catenary fault detection method based on machine learning as described in claim 1, characterized in that: In step S1, the historical monitoring data is preprocessed, including: Eliminating monitoring samples with missing data, eliminating repeatedly saved monitoring samples, modifying or eliminating monitoring samples with format errors; The historical monitoring data is processed by principal component analysis for dimension reduction; In step S7, the preset number of times = k x M, wherein k = 0.15-0.3, and M = the total number of monitoring samples in the initial training database; the preset ratio is 0.9-0.98.

Images