A method and system for detecting track risks based on big data analysis

Abstract

The present invention discloses a method and a system for detecting track risks based on big data analysis, which relates to the technical field of track safety. The present invention uses the YOLO model and the SVM classifier to detect the track, extracts multi-scale features of the track image, and can quickly and effectively detect the deformation and damage positions in the track; based on the detection results, the RF model and the BP neural network are respectively used to further analyze the fault points of the track, and the fault points in the track are accurately determined by combining the outputs of each model, reducing the misjudgment and omission of fault points, providing effective support for the subsequent detection path planning; according to the fault points, a detection path is formulated by combining the greedy algorithm and the ant algorithm, which can effectively reduce the detection time, improve the efficiency of the staff, improve the safety of train operation, and assist train operation management.

Description

Technical Field

[0001] The present invention relates to the technical field of track safety, and particularly to a method and system for detecting track risks based on big data analysis. Background Art

[0002] During the operation of the track, it is necessary to check whether the track is deformed, damaged, etc. With the development of technology, track detection technology is also constantly improving to improve the safety of train operation. Currently, track detection methods mainly include manual detection and machine learning algorithm detection. Although the manual detection method is intuitive and reliable, it is inefficient and labor-intensive. While the machine learning algorithm has improved the detection efficiency to a certain extent, there are still problems with insufficient detection accuracy, especially in complex environments, and the accuracy and robustness of the algorithm still need to be improved. Summary of the Invention

[0003] The purpose of the present invention is to effectively reduce the detection time, improve the efficiency of staff, and improve the safety of train operation. For this reason, a method and system for detecting track risks based on big data analysis are provided.

[0004] To achieve the above-mentioned invention purpose, the embodiments of the present invention provide the following technical solutions:

[0005] A method for detecting track risks based on big data analysis, which includes:

[0006] Determine the track to be detected and its track image data, and obtain the position information of each station on the track to be detected; based on the position information of each station, segment the track to be detected to obtain the segmented track;

[0007] Respectively input the track image data corresponding to the segmented track into the track detection model, and output the detection result corresponding to the segmented track;

[0008] The track detection model includes a cascaded YOLO model and an SVM classifier;

[0009] According to the detection result, obtain the environmental data of the segmented track, and input it into the track damage prediction model to output the corresponding damage prediction result;

[0010] The track damage prediction model adopts an RF model;

[0011] According to the detection result, obtain the path signal corresponding to the segmented track, and input it into the track fault judgment model to output the corresponding fault judgment result;

[0012] The track fault judgment model adopts a BP neural network model;

[0013] According to the detection result, the damage prediction result and the fault judgment result, perform path planning on the track to be investigated to obtain the corresponding investigation path;

[0014] Dispatch investigators to investigate the track to be investigated according to the investigation path;

[0015] The training process of the track detection model includes:

[0016] S11. Obtain the training track image data and its first label; the first label includes damage, deformation and normal;

[0017] S12. Select the training track image data corresponding to any one of the segmented tracks, and input it together with the corresponding first label into the YOLO model to obtain the initial training detection result;

[0018] S13. Input the initial training detection result into the SVM classifier to obtain the training detection result; among them, the training detection result is divided into to be investigated and normal; the to be investigated includes the training deformation detection result and the training damage detection result; the normal is the training normal detection result;

[0019] S14. Calculate the corresponding loss function based on the training detection result;

[0020] S15. Repeat S12 to S14 until the loss functions corresponding to all the segmented tracks are obtained;

[0021] S16. Calculate the final loss function based on all the loss functions; based on the final loss function, adjust the weight parameters of the track detection model to complete the training of the track detection model;

[0022] The training process of the track damage prediction model includes:

[0023] S21. Obtain the historical environmental parameters corresponding to the training deformation detection result and its second label; the historical environmental parameters include historical temperature data and historical humidity data; the second label is damage and normal;

[0024] S22. Input the training deformation detection result, the historical environmental parameters and its second label into the RF model, and output the corresponding damage prediction result;

[0025] S23. Calculate the mean square error loss function based on the damage prediction result;

[0026] S24. Based on the mean square error loss function, adjust the parameters of the track damage prediction model to complete the training of the track damage prediction model;

[0027] The training process of the track fault judgment model includes:

[0028] S31. Obtain the training path signal data and its third label; wherein, the training path signal data is divided into the training path signal corresponding to the training deformation detection result, the training path signal corresponding to the training damage detection result, and the training path signal corresponding to the training normal detection result; the third label includes the line faults and non-line faults of the track, that is, signal fault points and signal non-fault points;

[0029] S32. Input the training path signal data and its third label into the BP neural network model, and output the corresponding training fault judgment result;

[0030] S33. Calculate the corresponding cross-entropy loss function based on the training fault judgment result;

[0031] S34. Based on the cross-entropy loss function, use the backpropagation algorithm and the gradient descent algorithm to adjust the weight parameters of the BP neural network model to complete the training of the track fault judgment model; the training fault judgment result includes training signal fault points and training signal non-fault points;

[0032] Performing path planning on the track to be investigated according to the detection result, the risk probability data, and the judgment result to obtain the corresponding investigation path, including:

[0033] Set the damage threshold, and divide the damage prediction result into a high-risk damage prediction result and a low-risk damage prediction result based on the damage threshold; regard the track with the detection result to be investigated, the damage prediction result as a high-risk damage prediction result, and the judgment result as a signal fault point as the first-priority node; regard the track with the detection result to be investigated, the damage prediction result as a high-risk damage prediction result, and the judgment result as a non-signal fault point, and the track with the detection result to be investigated, the damage prediction result as a low-risk damage prediction result, and the judgment result as a signal fault point as the second-priority node; regard the track with the detection result to be investigated, the damage prediction result as a low-risk damage prediction result, and the judgment result as a non-signal fault point as the third-priority node; obtain the position information of the first to third priority nodes;

[0034] Calculate the distance between the first-priority node and the station; and based on the distance, determine the station with the shortest distance from each first-priority node, and obtain the position relationship diagram between the first-priority node and the station;

[0035] Based on the position relationship diagram, the second-priority nodes, the third-priority nodes, and the position information of each site, use the greedy-ant colony algorithm to perform path planning on the track to be investigated, and obtain the investigation path.

[0036] Further, based on the position information of each site, segment the track to be investigated to obtain the segmented track, including: according to the position information of the site, divide the track between two adjacent sites on the track to be investigated into an independent paragraph to obtain the segmented track.

[0037] Further, the process of obtaining the track image data is as follows:

[0038] Use a drone to take side shots and overhead shots of the track to be investigated respectively, to obtain two corresponding side video data and one overhead video data; extract the side video data and the overhead video data according to the preset number of frames to obtain the corresponding side image data and overhead image data;

[0039] Perform data cleaning, image enhancement, size conversion, and normalization processing on each of the side image data and the overhead image data to obtain the corresponding processed side image data and processed overhead image data; splice the processed overhead image data with the corresponding two processed side image data to obtain the track image data.

[0040] Further, the path planning by the greedy-ant colony algorithm includes:

[0041] Initialize the path, and use M sites as the starting nodes respectively, and use the first-priority node as the neighbor node;

[0042] Based on the position relationship diagram, allocate N investigation personnel at each site to the neighbor node with the shortest distance;

[0043] Update the path, and add the selected node to the path; update the position information of all investigation personnel;

[0044] Use the second-priority node as the neighbor node, and based on the updated position information of the investigation personnel, the position information of the second-priority node, and the path selection principle, use the ant colony algorithm to select the next node; update the path, and add the selected node to the path; update the position information of all investigation personnel;

[0045] Repeat the above process until all nodes are selected, and output to obtain the investigation path;

[0046] Among them, the principle for path selection is as follows: select the node with the shortest distance; if multiple inspection personnel have the same distance to the same node, the inspection personnel with fewer passed nodes select this node, and the inspection personnel with more passed nodes select the node with the second shortest distance.

[0047] A detection system for track risks based on big data analysis, which includes an information acquisition and processing module, a track detection module, a track damage prediction model module, a track fault judgment model module, and a path planning module; among them:

[0048] The information acquisition and processing module is used to obtain the track to be inspected, its track image data, environmental data, path signals, and obtain the position information of each station on the track to be inspected; and based on the position information of each station, segment the track to be inspected to obtain the segmented track.

[0049] The track detection module is used to detect the track image data corresponding to the segmented track by using the YOLO model and the SVM classifier respectively to obtain the detection result.

[0050] The track damage prediction model module is used to use the RF model to perform damage prediction on the detection result and the environmental data to obtain the damage prediction result.

[0051] The track fault judgment model module is used to judge the path signal corresponding to the segmented track by using the BP neural network model based on the detection result and the damage prediction result to obtain the fault judgment result.

[0052] The path planning module is used to perform path planning on the track to be inspected by using the greedy-ant colony algorithm based on the detection result, the damage prediction result, and the fault judgment result to obtain the inspection path.

[0053] The present invention has the following beneficial effects:

[0054] The present invention uses the YOLO model and the SVM classifier to detect the track, extracts the multi-scale features of the track image, and can quickly and effectively detect the deformation and damage positions in the track; based on the detection result, uses the RF model and the BP neural network respectively to further analyze the fault points of the track, combines the outputs of each model to accurately determine the fault points in the track, reduces the misjudgment and omission of fault points, provides effective support for the subsequent inspection path planning; according to the fault points, combines the greedy algorithm and the ant algorithm to formulate the inspection path, which can effectively reduce the inspection time, improve the efficiency of the staff, improve the safety of train operation, and assist train operation management. Description of the Drawings

[0055] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for the embodiments will be briefly introduced below. It should be understood that the following accompanying drawings only show some embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related accompanying drawings can be obtained based on these drawings without creative efforts.

[0056] Figure 1 It is the invention flow chart in the embodiment of the present invention;

[0057] Figure 2 It is the schematic diagram of the troubleshooting path in the embodiment of the present invention;

[0058] Figure 3 It is the system module block diagram in the embodiment of the present invention. Detailed implementation manners

[0059] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Usually, the components of the embodiments of the present invention described and shown in the accompanying drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed present invention, but only represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0060] Please refer to Figure 1 , a troubleshooting method for track risks based on big data analysis provided in this embodiment, which includes:

[0061] Determine the track to be troubleshot and its track image data, and obtain the position information of each station on the track to be troubleshot; based on the position information of each station, segment the track to be troubleshot to obtain the segmented track.

[0062] Segmenting the track to be troubleshot based on the position information of each station to obtain the segmented track includes: dividing the track between two adjacent stations on the track to be troubleshot into an independent segment according to the position information of the stations to obtain the segmented track. Segmenting the track to be troubleshot can reduce the operation burden and operation time of the subsequent track detection model, thereby improving the calculation speed of the present invention.

[0063] The process of obtaining the track image data is:

[0064] The drone is used to take side shots and overhead shots of the track to be inspected respectively, obtaining two corresponding side video data and one overhead video data; the side video data and the overhead video data are extracted according to a preset number of frames to obtain corresponding side image data and overhead image data;

[0065] Data cleaning, image enhancement, size conversion and normalization processing are performed on each of the side image data and the overhead image data to obtain corresponding processed side image data and processed overhead image data; the processed overhead image data is stitched with the corresponding two processed side image data to obtain the track image data.

[0066] The track image data corresponding to the segmented track is respectively input into the track detection model, and the detection results corresponding to the segmented track are output.

[0067] The track detection model includes a cascaded YOLO model and an SVM classifier; the YOLO model uses the YOLOv4 neural network model to detect damages such as deformations, cracks and fissures in the track image; the SVM classifier is used to classify the initial detection results output by YOLOv4 into deformation detection results and damage detection results.

[0068] The training process of the track detection model includes:

[0069] S11. Obtain training track image data and its first label; the first label includes damage, deformation and normal;

[0070] S12. Select the training track image data corresponding to any one of the segmented tracks, and input it together with the corresponding first label into the YOLO model to obtain an initial training detection result;

[0071] S13. Input the initial training detection result into the SVM classifier to obtain a training detection result; among them, the training detection result is divided into to be inspected and normal; the to be inspected includes training deformation detection result, training damage detection result; the normal is training normal detection result; the training damage detection result is a crack or fissure.

[0072] S14. Calculate the corresponding loss function based on the training detection result ;

[0073] S15. Repeat S12 to S14 until the loss functions corresponding to all the segmented tracks are obtained;

[0074] S16. Calculate the final loss function based on all the above loss functions; adjust the weight parameters of the track detection model based on the final loss function to complete the training of the track detection model.

[0075] Final loss function The calculation formula is:

[0076]

[0077]

[0078] Where represents the total number of segmented tracks, represents the summation function, represents the maximum value function, represents the weight coefficient, represents the generalized intersection ratio between the predicted bounding box and the ground truth bounding box of the YOLO model for the th time, represents the output of the sigmoid function in the SVM classifier for the th time, represents the label corresponding to the initial detection result for the

[0079] Using the YOLO model to extract the image features of the track to be investigated and output the initial detection result, and then using the SVM classifier to classify the initial detection result. This multi-stage detection can improve the accuracy of identifying damage and deformation; the YOLO model captures the multi-scale features of the track image, and the SVM classifier further classifies the deformation damage based on the multi-scale features of the track image, increasing the classification granularity, which helps to improve the workload of the staff, reduce the complexity of the subsequent path planning, and assist the path planning to provide a more accurate investigation path.

[0080] According to the detection result, obtain the environmental data of the segmented track and input it into the track damage prediction model to output the corresponding damage prediction result.

[0081] The track damage prediction model adopts the RF model, which is used to predict the damage of the deformation detection result output by the SVM classifier and its corresponding environmental parameters; combined with temperature and humidity, using the powerful feature processing ability of the RF model and its ability to handle non-linear relationships, it can well capture the relationship between temperature, humidity, deformation result and the occurrence of cracks in the track, so as to accurately predict the probability of track damage.

[0082] The training process of the track damage prediction model includes:

[0083] S21. Obtain the historical environmental parameters corresponding to the training deformation detection result and its second label; the historical environmental parameters include historical temperature data and historical humidity data; the second label is damage and normal;

[0084] S22. Input the training deformation detection result, the historical environmental parameters and the second label into the RF model, and output the corresponding damage prediction result;

[0085] S23. Calculate the mean square error loss function based on the damage prediction result;

[0086] S24. Based on the mean square error loss function, adjust the parameters of the track damage prediction model to complete the training of the track damage prediction model.

[0087] Use the RF model to capture the relationship between temperature, humidity, deformation detection result and track damage, and predict track damage based on this relationship. Combining environmental parameters and deformation results provides strong support for track maintenance and damage prevention.

[0088] According to the detection result, obtain the path signal corresponding to the segmented track and input it into the track fault judgment model, and output the corresponding fault judgment result.

[0089] The track fault judgment model adopts a BP neural network model, which is used to detect faults in the track to be checked with path signals, and detect whether there is a red light band phenomenon (red light band fault), that is, the signal fault point, in the track to be checked.

[0090] The training process of the track fault judgment model includes:

[0091] S31. Obtain the training path signal data and its third label; among them, the training path signal data is divided into the training path signal corresponding to the training deformation detection result, the training path signal corresponding to the training damage detection result and the training path signal corresponding to the training normal detection result; the third label includes the line fault and line non-fault corresponding to the track, that is, the signal fault point and signal non-fault point;

[0092] S32. Input the training path signal data and the third label into the BP neural network model, and output the corresponding training fault judgment result;

[0093] S33. Calculate the corresponding cross-entropy loss function based on the training fault judgment result;

[0094] S34. Based on the cross - entropy loss function, use the backpropagation algorithm and the gradient descent algorithm to adjust the weight parameters of the BP neural network model, and complete the training of the track fault judgment model; the training fault judgment results include training signal fault points and training signal non - fault points.

[0095] Taking advantage of the powerful non - linear mapping ability, self - learning ability, generalization ability, etc. of the BP neural network model, it can achieve high - precision diagnosis of track red light band faults, ensure the normal operation of the track, assist the platform in accurately obtaining train information on the track, and maintain communication between the platform and the train.

[0096] According to the detection result, the damage prediction result and the fault judgment result, perform path planning on the track to be investigated, and obtain the corresponding investigation path.

[0097] According to the detection result, the risk probability data and the judgment result, perform path planning on the track to be investigated, and obtain the corresponding investigation path, including:

[0098] Set a damage threshold, and based on the damage threshold, divide the damage prediction results into high - risk damage prediction results and low - risk damage prediction results; regard the tracks corresponding to the detection result of to be investigated, the high - risk damage prediction result and the judgment result of signal fault points as the first - priority nodes; regard the tracks corresponding to the detection result of to be investigated, the high - risk damage prediction result and the judgment result of non - signal fault points, and the tracks corresponding to the detection result of to be investigated, the low - risk damage prediction result and the judgment result of signal fault points as the second - priority nodes; regard the tracks corresponding to the detection result of to be investigated, the low - risk damage prediction result and the judgment result of non - signal fault points as the third - priority nodes; obtain the position information of the first to third - priority nodes; each piece of position information is DEM data.

[0099] In one embodiment, as Figure 2 shown, determine a certain area of the track as the track to be investigated, and obtain 3 stations on the track to be investigated and the corresponding position information. Obtain the corresponding detection results, risk probability data and judgment results through each model; regard the tracks corresponding to the detection results and judgment results as nodes; count the number of nodes as 19, and randomly label the nodes; according to the detection results, risk probability data and judgment results, regard nodes 1 to 4 as the first - priority nodes, regard nodes 5 to 11 as the second - priority nodes, and regard nodes 12 to 19 as the third - priority nodes.

[0100] Calculate the distance between the first-priority nodes and the site; and based on the distance, determine the site with the shortest distance from each first-priority node, and obtain the position relationship diagram between the first-priority nodes and the site;

[0101] Based on the position relationship diagram, the second-priority nodes, the third-priority nodes, and the position information of each site, use the greedy-ant colony algorithm to perform path planning on the to-be-investigated track to obtain the investigation path.

[0102] Path planning using the greedy-ant colony algorithm includes:

[0103] Initialize the path, and use M sites as the starting nodes respectively, and use the first-priority nodes as neighbor nodes;

[0104] Based on the position relationship diagram, assign N investigation personnel at each site to the neighbor node with the shortest distance;

[0105] Update the path, and add the selected nodes to the path; update the position information of all investigation personnel;

[0106] Use the second-priority nodes as neighbor nodes. Based on the updated position information of the investigation personnel, the position information of the second-priority nodes, and the path selection principle, use the ant colony algorithm to select the next node; update the path, and add the selected nodes to the path; update the position information of all investigation personnel;

[0107] When using the ant colony algorithm to select the next node, set the initial pheromone concentration of the ant colony and parameters, where the parameters include the pheromone importance factor, the heuristic information importance factor, and the pheromone evaporation rate; use the current position information of the investigation personnel as the initial point of the ants, and use the investigation personnel as ants. At this time, the position information of the investigation personnel may all be the positions of the first-priority nodes, or may be partially the positions of the first-priority nodes and partially the positions of the sites; randomly initialize the starting order of the ants, and calculate the probability of each ant selecting the second-priority node according to the ant colony parameters and the initial pheromone concentration; according to the probability, select the second-priority node as the next node of the ant until all the second-priority nodes are selected.

[0108] Repeat the above process until all nodes are selected, and output to obtain the investigation path;

[0109] Among them, the path selection principle is: select the node with the shortest distance; if the distances of multiple investigation personnel to the same node are the same, then the investigation personnel with fewer passed nodes select this node, and the investigation personnel with more passed nodes select the node with the second shortest distance.

[0110] There are 3 investigators at each site, for a total of 9 investigators; calculate the location relationship diagram between the first-priority nodes and the sites, and use the 3 sites as starting points and the first-priority nodes as neighbor nodes; use the greedy algorithm to select neighbor nodes. Nodes 1 and 2 are closer to Site 3, and Nodes 3 and 4 are closer to Site 1. Therefore, two investigators from both Site 1 and Site 3 are dispatched to Nodes 3, 4, 1, and 2 respectively, and the investigators at Site 2 and the remaining investigators at Site 1 and Site 3 remain stationary. Update the location information of all investigators. Calculate the location distances between all investigators and the second-priority nodes, and use the location information of all investigators as the initial nodes of the ants; select the next node according to the pheromone level and path selection principle, that is: Node 5 is closer to Site 3, and Node 9 is closer to Node 2. Therefore, the remaining investigators at Site 3 are dispatched to Node 5, and the investigators at Node 2 are dispatched to Node 9; Node 8 is closer to Node 3, and Node 10 is closer to Site 1. Therefore, the investigators at Node 3 are dispatched to Node 8, and the remaining investigators at Site 1 are dispatched to Node 10; Nodes 6, 7, and 11 are closer to Site 2. Therefore, the 3 investigators at Site 2 are dispatched to Nodes 6, 7, and 11 respectively. Update the location information of all investigators. Calculate the location distances between all investigators and the third-priority nodes, and use the location information of all investigators as the initial nodes of the ants; select the next node according to the pheromone level and path selection principle, that is: Node 13 is the closest to Node 6, so the investigators at Node 6 are dispatched to Node 13; Node 14 is the closest to Node 1, so the investigators at Node 1 are dispatched to Node 14; Node 16 is the closest to Node 9, so the investigators at Node 9 are dispatched to Node 16; Node 19 is the closest to Node 11, so the investigators at Node 11 are dispatched to Node 19; Node 15 is the closest to Node 10, so the investigators at Node 10 are dispatched to Node 15; Node 4 is the closest to Node 18, so the investigators at Node 4 are dispatched to Node 18; at this time, Node 12 may not be selected, so update the location information of all investigators again and repeat the above process to determine that Node 16 is the closest to Node 12, so the investigators at Node 16 are dispatched to Node 12.

[0111] Combining the advantages of the greedy algorithm and the ant colony algorithm can improve the problem that the greedy algorithm is prone to falling into local optimality during path planning, can more accurately select nodes with the shortest time or distance, improve the rationality of the investigation path and the efficiency of the staff; improve the search efficiency of path planning; dividing the nodes into priorities can reduce the calculation time.

[0112] According to the inspection path, dispatch inspectors to inspect the to-be-inspected track. According to the detection results, the risk probability data, and the judgment results, corresponding inspectors can be allocated and equipped with relevant tools to reduce the load of the inspectors. If other problems are found when the inspectors are inspecting the corresponding track, they can be promptly reported to the cloud, and then further response plans can be formulated.

[0113] As Figure 3 shown, a track risk inspection system based on big data analysis includes an information collection and processing module, a track detection module, a track damage prediction model module, a track fault judgment model module, and a path planning module; where:

[0114] The information collection and processing module is used to obtain the to-be-inspected track and its track image data, environmental data, and path signals, and obtain the position information of each station on the to-be-inspected track; and based on the position information of each station, segment the to-be-inspected track to obtain the segmented track.

[0115] The track detection module is used to use the YOLO model and the SVM classifier to detect the track image data corresponding to the segmented track respectively to obtain the detection results.

[0116] The track damage prediction model module is used to use the RF model to perform damage prediction on the detection results and the environmental data to obtain the damage prediction results.

[0117] The track fault judgment model module is used to use the BP neural network model to judge the path signals corresponding to the segmented track based on the detection results and the damage prediction results to obtain the fault judgment results.

[0118] The path planning module is used to perform path planning on the to-be-inspected track using the greedy-ant colony algorithm based on the detection results, the damage prediction results, and the fault judgment results to obtain the inspection path.

[0119] In the specific implementation process, the damage detection results of the track to be checked are obtained through the track detection model, and the crack size or area in the damage detection results is calculated. A crack threshold is set, and the damage detection results with the crack size or area greater than the crack threshold are extracted. When formulating the inspection path, the damage detection results with the crack size or area greater than the crack threshold can be used as reference factors to formulate the first to third priority nodes. Considering the damage size of the track, the inspection path can be optimized, the inspection personnel can be assigned to the locations that need to be inspected more, the probability of false reduction can be reduced, and the accuracy of damage detection can be improved; a more targeted inspection arrangement can be provided, unnecessary inspections can be reduced, and work efficiency can be improved; a new priority order can be formulated, and larger cracks can be repaired in time, which helps to improve the service life of the track and reduce the potential accident cost.

[0120] The present invention uses the YOLO model and the SVM classifier to detect the track, extracts the multi-scale features of the track image, and can quickly and effectively detect the deformation and damage positions in the track; based on the detection results, the RF model and the BP neural network are respectively used to further analyze the fault points of the track, and the fault points in the track are accurately determined by combining the outputs of each model, reducing the misjudgment and omission of fault points, and providing effective support for the subsequent inspection path planning; according to the fault points, the inspection path is formulated by combining the greedy algorithm and the ant algorithm, which can effectively reduce the inspection time, improve the efficiency of the staff, improve the safety of train operation, and assist train operation management.

[0121] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A method for detecting rail risks based on big data analysis, characterized in that, Including: Determine the orbit to be investigated and its orbit image data, and obtain the location information of each station on the orbit to be investigated; Based on the location information of each station, segment the orbit to be investigated to obtain the segmented orbit; Input the orbit image data corresponding to the segmented orbit into the orbit detection model respectively to obtain the detection results corresponding to the segmented orbit; The detection results include the detection results to be investigated and the normal detection results. The results to be investigated include deformation and damage; Obtain the environmental data of the segmented orbit corresponding to the detection result of deformation and input it into the orbit damage prediction model to obtain the damage prediction result; Obtain the path signal corresponding to the segmented orbit corresponding to the detection result to be investigated and input it into the orbit fault judgment model to obtain the fault judgment result; The fault judgment result includes signal fault points and non-signal fault points. The signal fault point is a red light band fault; Among them, the orbit detection model includes a cascaded YOLO model and an SVM classifier; The orbit damage prediction model uses an RF model; The orbit fault judgment model uses a BP neural network model; According to the detection results, damage prediction results and fault judgment results, conduct path planning for the orbit to be investigated to obtain the investigation path. The corresponding process is: Set the damage threshold, and based on the damage threshold, divide the damage prediction results into high-risk damage prediction results and low-risk damage prediction results; The orbits corresponding to the detection result to be investigated, the high-risk damage prediction result and the judgment result of signal fault points are used as the first-priority nodes; The orbits corresponding to the detection result to be investigated, the high-risk damage prediction result and the judgment result of non-signal fault points, and the orbits corresponding to the detection result to be investigated, the low-risk damage prediction result and the judgment result of signal fault points are used as the second-priority nodes; The orbits corresponding to the detection result to be investigated, the low-risk damage prediction result and the judgment result of non-signal fault points are used as the third-priority nodes; Obtain the location information of the first to third priority nodes; Calculate the distance between the first-priority nodes and the stations, and determine the stations with the shortest distance from each first-priority node to obtain the location relationship diagram of the first-priority nodes and the stations; Based on the location relationship diagram, the second-priority nodes, the third-priority nodes and the location information of each station, use the greedy-ant colony algorithm to conduct path planning for the orbit to be investigated to obtain the investigation path; According to the investigation path, dispatch investigation personnel to investigate the orbit to be investigated.

2. A method for investigating rail risks based on big data analysis according to claim 1, characterized in that, The process of obtaining the orbit image data is: Use the drone to conduct side shooting and top-down shooting on the orbit to be investigated respectively to obtain two corresponding side video data and one top-down video data; Extract the side video data and top-down video data according to the preset number of frames to obtain the corresponding side image data and top-down image data; Data cleaning, image enhancement, size conversion, and normalization are performed on the side image data and the top-down image data to obtain the corresponding processed side image data and processed top-down image data; the processed top-down image data is stitched with the corresponding two processed side image data to obtain the track image data. Based on the location information of each station, the track to be investigated is segmented to obtain the segmented track, including: according to the location information of the station, the track between two adjacent stations on the track to be investigated is divided into an independent paragraph to obtain the segmented track.

3. A method for detecting risks of a track based on big data analysis according to claim 1, characterized in that, The training process of the track detection model includes: S11. Obtain the training track image data and its first label; the first label includes track damage, track deformation, and track normal. S12. Select the training track image data corresponding to any one of the segmented tracks, and input it together with the corresponding first label into the YOLO model to obtain the initial training detection result. S13. Input the initial training detection result into the SVM classifier to obtain the training detection result; the training detection result is divided into to be investigated and normal; to be investigated includes the training deformation detection result and the training damage detection result; normal is the training normal detection result. S14. Calculate the corresponding loss function based on the training detection result. S15. Repeat S12 to S14 until the loss functions corresponding to all segmented tracks are obtained. S16. Calculate the final loss function based on all loss functions; based on the final loss function, adjust the weight parameters of the track detection model to complete the training of the track detection model. The training process of the track damage prediction model includes: S21. Obtain the historical environmental parameters corresponding to the training deformation detection result and its second label; the historical environmental parameters include historical temperature data and historical humidity data; the second label is damage and normal. S22. Input the training deformation detection result, the historical environmental parameters, and its second label into the RF model, and output the corresponding damage prediction result. S23. Calculate the mean square error loss function based on the damage prediction result. S24. Based on the mean square error loss function, adjust the parameters of the track damage prediction model to complete the training of the track damage prediction model. The training process of the track fault judgment model includes: S31. Obtain the training path signal data and its third label; the training path signal data is divided into the training path signal corresponding to the training deformation detection result, the training path signal corresponding to the training damage detection result, and the training path signal corresponding to the training normal detection result; the third label includes the line fault and line non-fault corresponding to the track, that is, the signal fault point and the signal non-fault point. S32. Input the training path signal data and its third label into the BP neural network model, and output the training fault judgment result. S33. Calculate the corresponding cross-entropy loss function based on the training fault judgment result. S34. Based on the cross-entropy loss function, use the backpropagation algorithm and the gradient descent algorithm to adjust the weight parameters of the BP neural network model to complete the training of the track fault judgment model; the training fault judgment result includes the training signal fault point and the training signal non-fault point.

4. A method for detecting track risks based on big data analysis as described in claim 1, characterized in that The path planning using the greedy-ant colony algorithm includes: Initialize the path, take M stations as the starting nodes respectively, and take the first-priority node as the neighbor node; Based on the position relationship graph, assign N inspectors at each station to the neighbor node with the shortest distance; Update the path by adding the selected node to the path; update the position information of all inspectors; Take the second-priority node as the neighbor node, and use the ant colony algorithm to select the next node based on the updated position information of the inspectors, the position information of the second-priority node, and the path selection principle; update the path by adding the selected node to the path; update the position information of all inspectors; Repeat the above process until all nodes are selected, and output the inspection path; Among them, the path selection principle is: select the node with the shortest distance; if the distances of multiple inspectors to the same node are the same, the inspector with fewer passed nodes selects this node, and the inspector with more passed nodes selects the node with the second shortest distance.

5. A track risk investigation system based on big data analysis, which is used to implement a track risk investigation method based on big data analysis according to any one of claims 1 to 4, and is characterized in that, It includes an information acquisition and processing module, an orbit detection module, an orbit damage prediction model module, an orbit fault judgment model module, and a path planning module; among them: The information acquisition and processing module is used to obtain the orbit to be inspected, its orbit image data, environmental data, path signals, and obtain the position information of each station on the orbit to be inspected; and based on the position information of each station, segment the orbit to be inspected to obtain the segmented orbit; The orbit detection module is used to detect the orbit image data corresponding to the segmented orbit by using the YOLO model and the SVM classifier respectively to obtain the detection results; The orbit damage prediction model module is used to predict the damage of the detection result of orbit deformation and its corresponding environmental data by using the RF model to obtain the damage prediction results; The orbit fault judgment model module is used to judge the path signals corresponding to the segmented orbit corresponding to the inspection result to be inspected by using the BP neural network model to obtain the fault judgment results; The path planning module is used to plan the path of the orbit to be inspected by using the greedy-ant colony algorithm based on the detection results, damage prediction results, and fault judgment results to obtain the inspection path.

Images