Intelligent track defect detection and fault identification system and method

Abstract

The invention discloses an intelligent track defect detection and fault identification system and method. The system comprises a real-time image acquisition module and a real-time Tiny calculation chip module, the real-time Tiny calculation chip module comprises an image processing unit, an image anomaly detection and judgment unit and a fault type identification unit; the real-time image acquisition module acquires a track circuit image in a train travel path in real time; the image processing unit extracts a track circuit physical feature map in real time according to the track circuit image; the image anomaly detection and judgment unit is used for carrying out track defect detection according to the extracted track circuit physical feature map and outputting a detection result; and the fault type identification unit completes track fault identification according to the track defect detection result and outputs an identification result. Through the system, a worker can be supported to quickly and effectively arrive at a fault site to carry out error correction and maintenance, so that the maintenance efficiency of the track circuit is improved, and the maintenance workload of the track circuit is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of track physical defect detection, and in particular to a system and method for intelligent track defect detection and fault identification. Background Art

[0002] In recent years, rail transit technology will continue to innovate and develop, such as high-speed magnetic levitation technology, intelligent rail transit system, etc. The application of these new technologies will further improve the speed, safety and comfort of rail transit, and provide people with better travel services. However, the existing track physical defect detection process requires the cooperation of multiple types of work and multiple departments, which takes a long time, has poor timeliness, and is prone to safety hazards. In addition, workers have limited maintenance time, mostly patrolling the line at night during driving intervals, and the work is hard and dangerous.

[0003] At present, the prior art proposes a track inspection robot and a combined track inspection robot, which perform inspections by entering the robot from an intermediate station of a subway. The inspection robot is suitable for inspections of intermediate running sections of a subway.

[0004] However, since the work is done during the non-working hours of the subway vehicles, the track circuit cannot be inspected in real time; and considering the moving speed of the robot, it is impossible to quickly patrol the track circuit for inspection, which limits the work efficiency; in addition, the current inspection robot cannot perform fault detection at a long distance, nor can it provide feedback on the fault coordinates, resulting in the staff being unable to quickly and effectively reach the fault location for error correction and repair, which in turn leads to low track circuit maintenance efficiency and a large track circuit maintenance workload. Summary of the invention

[0005] In order to overcome the problem that the prior art cannot support the staff to quickly and effectively reach the fault location for error correction and repair, which leads to low track circuit maintenance efficiency and large track circuit maintenance workload, the present invention proposes an intelligent track defect detection and fault identification system and method, which can support the staff to quickly and effectively reach the fault location for error correction and repair, thereby improving the track circuit maintenance efficiency and reducing the track circuit maintenance workload.

[0006] To achieve the purpose of the present invention, the present invention adopts the following technical solutions:

[0007] A system for intelligent rail defect detection and fault identification, the system comprising: a real-time image acquisition module and a real-time Tiny computing chip module, the real-time image acquisition module and the real-time Tiny computing chip module are wirelessly connected; the real-time Tiny computing chip module comprises an image processing unit, an image anomaly detection and judgment unit and a fault type identification unit;

[0008] The real-time image acquisition module acquires the track circuit image in the train path in real time, and transmits the acquired track circuit image to the real-time Tiny computing chip module;

[0009] The image processing unit extracts a physical characteristic diagram of the track circuit in real time based on the track circuit image; the image abnormality detection and judgment unit performs track defect detection based on the extracted physical characteristic diagram of the track circuit and outputs the detection result; the fault type identification unit completes track fault identification based on the track defect detection result and outputs the identification result.

[0010] In the above technical scheme, the track circuit image in the train path is collected in real time by the real-time image acquisition module, and the physical feature map of the track circuit is extracted in real time by the image processing unit in the real-time Tiny computing chip module. The image anomaly detection and judgment unit performs track defect detection according to the extracted physical feature map of the track circuit, and the fault type identification unit completes track fault identification according to the track defect detection result, so as to realize the intelligent track circuit defect detection and fault identification system. The system can conveniently realize track circuit defect detection, which is conducive to reducing the manufacturing cost of related equipment, and is also conducive to the intelligence of related equipment, and can support the staff to quickly and effectively reach the fault site for error correction and maintenance, thereby improving the efficiency of track circuit maintenance and reducing the workload of track circuit maintenance.

[0011] Preferably, the real-time image acquisition module includes a high-definition image imaging component and a synchronization device component;

[0012] The high-definition image imaging component includes several groups of high-speed area array cameras and matching fill-light LED lights, which are used to collect real-time images of the track circuit;

[0013] The synchronization device component is used to feed back the mileage information and position information of the train running on the track circuit in real time, and feed back the position information of the physical characteristic diagram of the track circuit.

[0014] In the above technical solution, several groups of high-speed area array cameras and matching fill-light LED lamps in the high-definition image imaging component can meet the needs of the train to capture images of the track circuit at a relative speed and perform real-time online detection of physical defects of the track. The synchronization device component can provide real-time feedback on the mileage information and position information of the train traveling on the track circuit, and feedback on the position information of the physical characteristic map of the track circuit.

[0015] Preferably, the image processing unit includes a fully connected layer and a fully convolutional layer;

[0016] The image processing unit divides the track circuit physical feature map into a candidate frame feature map and a common feature map, and keeps the sizes of the candidate frame feature map and the common feature map consistent;

[0017] The fully convolutional layer is used to extract the physical features of the track circuit of the candidate frame feature map;

[0018] The fully connected layer is used to extract the position coordinates and category information in the common feature map and integrate the features proposed by the convolutional layer;

[0019] Among them, the position coordinates represent the position coordinates of the feature in the image; the category information represents the actual attribute classification of the feature, such as track, sleeper, etc.; the segmentation information represents the feature extraction process of the image.

[0020] In the above technical solution, the fully connected layer and the convolutional layer can enhance the effect of the collected pictures and eliminate noise, vibration or other influences in the collection process.

[0021] Preferably, the image anomaly detection and judgment unit compares the local patch of the track circuit physical feature map with the image feature vector stored in the database, and calculates the Euclidean distance between the feature vector of the track circuit physical feature map and the closest feature vector in the database, which is expressed as:

[0022]

[0023] The smaller the calculated result is, the closer the detected track physical features are to the image feature vector stored in the database in space, and to a certain extent, the closer the features are. Determine whether the calculated Euclidean distance exceeds the pre-set threshold;

[0024] If it exceeds, the track circuit physical characteristic diagram currently undergoing track defect detection is abnormal, and the image abnormality detection and judgment unit transmits the calculated Euclidean distance and the abnormal result to the fault type identification unit;

[0025] If it does not exceed, it means that the physical characteristic diagram of the track circuit currently undergoing track defect detection is normal;

[0026] Where n represents the number of dimensions, x = x1, x2, ..., x n and y = y1, y2, ..., y n They represent the coordinates of two points in the feature space, and d(x,y) represents the Euclidean distance between the track circuit patch and the feature vector in the database.

[0027] Preferably, the fault type identification unit selects k fault samples for initialization as initial cluster centers according to the received Euclidean distance and abnormality results, and the expression is:

[0028] a=a1,a2,...a k ;

[0029] The fault type identification unit assigns the track circuit physical characteristic diagram with abnormality to the cluster center closest to it. At this time, the track circuit physical characteristic diagram can be regarded as a spatial data point. The cluster center and the assigned data point are regarded as a cluster;

[0030] When the fault type identification unit assigns a new track circuit physical characteristic diagram with abnormalities to any cluster, the cluster center of the current cluster is recalculated, and the expression is:

[0031]

[0032] When the number of faults present in the abnormal track circuit physical characteristic diagram identified by the fault type identification unit is greater than a selected cluster number threshold, the fault type identification unit outputs a fault identification result;

[0033] Among them, a represents the initial selected value of the initial cluster center, k represents the selected number of clusters, and c i Indicates the class that is closest to sample i among the k classes, a j Represents the cluster center of the cluster.

[0034] In the above technical scheme, by calculating the Euclidean distance between the feature vector of the track circuit physical characteristic diagram and the closest feature vector in the database, it is possible to quickly and effectively determine whether the current track circuit physical characteristic diagram has an abnormality, and then use the cluster center method to determine the fault type and number of faults in the abnormal track circuit physical characteristic diagram, so as to realize the intelligent track circuit defect detection and fault identification system. The system can conveniently realize track circuit defect detection, which is conducive to reducing the manufacturing cost of related equipment, and is also conducive to the intelligence of related equipment, and can support staff to quickly and effectively reach the fault site for error correction and maintenance, thereby improving the efficiency of track circuit maintenance and reducing the workload of track circuit maintenance.

[0035] Preferably, the fault type identification unit uses a preset PatchCore model to perform fault judgment on the track circuit physical characteristic diagram with abnormalities, and the process model training execution process includes:

[0036] The PatchCore model uses distribution drift to divide the image into multiple local patches, and uses the residual neural network and the local domain formula to extract abnormal features. The expression is:

[0037]

[0038] At the same time, local abnormal features are extracted, and the expression is:

[0039]

[0040] The extracted abnormal features and local abnormal features are compared with the track circuit physical feature map in the original database to identify the fault type and fault quantity of the track circuit physical feature map with abnormalities;

[0041] in, It represents a patch of abnormal feature (patch) with a size of p*p at the position (h, w) on the feature map. h represents the horizontal coordinate of the feature map position, w represents the vertical coordinate of the feature map position, a represents the horizontal coordinate of the abnormal feature, b represents the vertical coordinate of the abnormal feature, and p represents the length of the normal pixel, usually p=3. represents the feature map tensor, f agg Represents the aggregation function of the neighborhood feature vector.

[0042] In the above technical scheme, the preset PatchCore model can quickly and efficiently extract the abnormal features and local abnormal features of the abnormal track circuit physical characteristic diagram during the process of identifying the fault type of the abnormal track circuit physical characteristic diagram, and then compare the features with the track circuit physical characteristic diagram in the original database to identify the fault type and number of the abnormal track circuit physical characteristic diagram, thereby realizing efficient feature screening.

[0043] Preferably, the system further comprises a result output module, which is convenient for wireless communication connection with the real-time image acquisition module and the real-time Tiny computing chip module, and the result output module comprises a vehicle-mounted matching display screen and a dispatch center matching display screen;

[0044] The vehicle-mounted display screen and the dispatch center display screen are both used to visualize the fault type of the track circuit physical characteristic diagram output by the fault type identification unit.

[0045] Preferably, the system further comprises a communication module, and the communication module is used to establish a wireless communication connection for each module in the system.

[0046] Preferably, the system further comprises a control chip, and the control chip is used to control the real-time image acquisition module, the real-time Tiny computing chip module, the result output module and the communication module to complete corresponding tasks.

[0047] In the above technical solution, the on-board matching display screen and the dispatch center matching display screen can enable the staff to intuitively view the physical defects in the track circuit, so as to support the staff to quickly and effectively reach the fault site for error correction and repair, thereby improving the efficiency of track circuit maintenance and reducing the workload of track circuit maintenance; the communication module can provide a safe and reliable communication environment for the connection between each module in the system to maintain the safety and stability of the system; the control chip can control each module to perform corresponding work to ensure that the system can work normally and stably.

[0048] A method for intelligent rail defect detection and fault identification, the method comprising the following steps:

[0049] Real-time acquisition of track circuit images along the train path;

[0050] According to the track circuit image, the physical characteristic map of the track circuit is extracted in real time;

[0051] Perform track defect detection based on the extracted track circuit physical feature map and output the detection results;

[0052] According to the track defect detection results, track fault identification is completed and the identification results are output.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] The present invention proposes a system and method for intelligent track defect detection and fault identification. The real-time image acquisition module is used to collect track circuit images in the train path in real time, and the image processing unit in the real-time Tiny computing chip module is used to extract the track circuit physical feature map in real time. The image abnormality detection and judgment unit performs track defect detection according to the extracted track circuit physical feature map, and the fault type identification unit completes track fault identification according to the track defect detection result, so as to realize the system of intelligent track circuit defect detection and fault identification. The system can conveniently realize track circuit defect detection, which is beneficial to reducing the manufacturing cost of related equipment, and is also beneficial to the intelligence of related equipment, and can support the staff to quickly and effectively reach the fault site for error correction and maintenance, thereby improving the track circuit maintenance efficiency and reducing the track circuit maintenance workload. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 A system structure block diagram of intelligent track defect detection and fault identification provided in an embodiment of the present application;

[0056] Figure 2 A flowchart of the overall execution steps of intelligent track circuit detection and fault identification provided in the embodiment of the present application;

[0057] Figure 3 A diagram showing the execution steps of a real-time image acquisition module provided in an embodiment of the present application;

[0058] Figure 4 A flowchart of the execution steps of the real-time Tiny computing chip module provided in the embodiment of the present application;

[0059] Figure 5 A flowchart of the PatchCore model training provided in the embodiment of the present application;

[0060] Figure 6A flowchart of the steps of a method for intelligent track defect detection and fault identification provided in an embodiment of the present application. DETAILED DESCRIPTION

[0061] In order to facilitate the understanding of the present invention, the present invention will be described more fully below with reference to the relevant drawings. Preferred embodiments of the present invention are provided in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present invention more thorough and comprehensive.

[0062] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items.

[0063] Embodiment 1:

[0064] This embodiment provides an intelligent rail defect detection and fault identification system. Figure 1 and Figure 2 The system comprises: a real-time image acquisition module and a real-time Tiny computing chip module, wherein the real-time image acquisition module is wirelessly connected to the real-time Tiny computing chip module; the real-time Tiny computing chip module comprises an image processing unit, an image anomaly detection and judgment unit, and a fault type identification unit;

[0065] The real-time image acquisition module acquires the track circuit image in the train path in real time, and transmits the acquired track circuit image to the real-time Tiny computing chip module;

[0066] The image processing unit extracts a physical characteristic diagram of the track circuit in real time based on the track circuit image; the image abnormality detection and judgment unit performs track defect detection based on the extracted physical characteristic diagram of the track circuit and outputs the detection result; the fault type identification unit completes track fault identification based on the track defect detection result and outputs the identification result.

[0067] As a preferred embodiment, the real-time image acquisition module includes a high-definition image imaging component and a synchronization device component;

[0068] The high-definition image imaging component includes several groups of high-speed area array cameras and matching fill-light LED lights, which are used to collect real-time images of the track circuit;

[0069] The synchronization device component is used to feed back the mileage information and position information of the train running on the track circuit in real time, and feed back the position information of the physical characteristic diagram of the track circuit.

[0070] Specifically, the high-definition image imaging component is composed of 8 groups of high-speed array cameras and supporting fill-light LED lights. Among them, 4 groups of high-speed array cameras shoot the rail area from two sides, 2 groups of high-speed array cameras shoot the roadbed area in the middle of the rail, and the other 2 groups of high-speed array cameras are used for slope adjustment. The imaging range of the high-speed array camera covers every area and component of the track, including the track plate (or sleeper and roadbed) area, the side and top surface of the rail, with a maximum range of up to 3000mm. The resolution of the high-speed array camera is 2048*2048; the exposure time is 0.02~1000ms; the maximum frame rate is 223fps. In the subway field, for the commonly used track with a sleeper spacing of 0.6m, most urban rail trains have a speed of 80km / h. If the train runs at a speed of 100km / h, the time it takes to pass a sleeper spacing is 0.6 / (100 / 3.6)=21.6ms. According to the frame rate of the high-speed array camera 223, the camera can take 4.8 pictures in 21.6ms, which can meet the needs of the urban rail industry. In the railway field, for a common track with a sleeper spacing of 0.6m, the time it takes for a train to pass through a sleeper spacing when running at a speed of 350km / h is 0.6 / (350 / 3.6)=6.17ms. According to the frame rate of the high-speed array camera 223, the camera can take 1.37 pictures in 6.17ms, which can meet the real-time online detection of track physical defects at this train speed.

[0071] In the synchronization device assembly: an electronic tag is installed at every 2 km interval on the main track. The tag stores the track mileage information. The tags are installed separately for up and down lines and in and out of the depot. At the same time, a high-speed tag reader is installed under the rail vehicle. When the train passes the electronic tag at a certain speed, its card number will be identified, and the card number will be compared with the information in the database to obtain the mileage information corresponding to the card number to realize the feedback position function.

[0072] It can be understood that the several groups of high-speed area array cameras and supporting fill-light LED lights in the high-definition image imaging component can meet the needs of the train to capture images of the track circuit at a relative speed and perform real-time online detection of physical defects in the track. The synchronization device component can provide real-time feedback on the mileage information and position information of the train traveling on the track circuit, and feedback on the position information of the physical characteristic diagram of the track circuit.

[0073] In this embodiment, the track circuit image in the train path is collected in real time by the real-time image acquisition module, and the physical feature diagram of the track circuit is extracted in real time by the image processing unit in the real-time Tiny computing chip module. The image anomaly detection and judgment unit performs track defect detection according to the extracted physical feature diagram of the track circuit, and the fault type identification unit completes track fault identification according to the track defect detection result, so as to realize the intelligent track circuit defect detection and fault identification system. The system can conveniently realize track circuit defect detection, which is conducive to reducing the manufacturing cost of related equipment, and is also conducive to the intelligence of related equipment, and can support the staff to quickly and effectively reach the fault site for error correction and maintenance, thereby improving the efficiency of track circuit maintenance and reducing the workload of track circuit maintenance.

[0074] Embodiment 2:

[0075] This embodiment is further described based on the first embodiment, and the details are as follows:

[0076] As a preferred embodiment, the image processing unit includes a fully connected layer and a fully convolutional layer;

[0077] The image processing unit divides the track circuit physical feature map into a candidate frame feature map and a common feature map, and keeps the sizes of the candidate frame feature map and the common feature map consistent;

[0078] The fully convolutional layer is used to extract the physical features of the track circuit of the candidate frame feature map;

[0079] The fully connected layer is used to extract the position coordinates and category information in the common feature map and integrate the features proposed by the full convolution layer;

[0080] Among them, the position coordinates represent the position coordinates of the feature in the image; the category information represents the actual attribute classification of the feature, such as track, sleeper, etc.; the segmentation information represents the feature extraction process of the image.

[0081] Specifically, see Figure 3 and Figure 4 The high-definition image imaging component continuously collects images of the track area that the train passes through in real time; and continuously obtains the track circuit images generated by the track circuit area that the rail vehicle passes through in real time, and inputs them into the image processing unit of the real-time Tiny computing chip module;

[0082] The image processing unit performs feature pyramid processing on the track circuit image to extract the track circuit physical feature map; the image processing unit performs "box selection" processing of the region proposal network (RPN) on the extracted track circuit physical feature map, crops out the part of the feature map corresponding to the candidate box, and sends it to the proposal region alignment layer (Rol Align); the unprocessed feature map of the track circuit physical feature map extracted by the real-time Tiny computing chip module is input into the proposal region alignment layer (Rol Align); further, the image processing unit obtains a combination of the processed candidate box feature map and the unprocessed feature map of the same size;

[0083] Furthermore, the image processing unit sends the processed feature maps of consistent size to the branched fully connected layer (FC) and the branched fully convolutional network (FCN) respectively; the fully connected layer (FC) is used to process the position coordinates and category information in the acquired image; the fully convolutional network (FCN) is used to process the segmentation information (Mask) in the acquired image; this process enhances the effect of the acquired image and eliminates noise or other influences in the acquisition process.

[0084] As a preferred embodiment, see Figure 4 The image anomaly detection and judgment unit compares the local patch (Core) of the track circuit physical feature map with the image feature vector stored in the database M, and calculates the Euclidean distance between the feature vector of the track circuit physical feature map and the closest feature vector in the database, which is expressed as:

[0085]

[0086] Determine whether the calculated Euclidean distance exceeds a preset threshold;

[0087] If it exceeds, the track circuit physical characteristic diagram currently undergoing track defect detection is abnormal, and the image abnormality detection and judgment unit transmits the calculated Euclidean distance and the abnormal result to the fault type identification unit;

[0088] If it does not exceed, it means that the track circuit physical characteristic diagram currently being inspected for track defects is normal and no other processing is performed;

[0089] Where n represents the number of dimensions, x = x1, x2, ..., x n and y = y1, y2, ..., y n They represent the coordinates of two points in the feature space, and d(x,y) represents the Euclidean distance between the track circuit patch and the feature vector in the database.

[0090] Furthermore, the fault type identification unit selects k fault samples for initialization as initial cluster centers according to the received Euclidean distance and abnormal results, and the expression is:

[0091] a=a1,a2,...a k ;

[0092] The fault type identification unit assigns the track circuit physical characteristic diagram with abnormality to the cluster center closest to it, and regards the cluster center and the assigned data points as a cluster;

[0093] When the fault type identification unit assigns a new track circuit physical characteristic diagram with abnormalities to any cluster, the cluster center of the current cluster is recalculated, and the expression is:

[0094]

[0095] When the number of faults present in the abnormal track circuit physical characteristic diagram identified by the fault type identification unit is greater than a selected cluster number threshold, the fault type identification unit outputs a fault identification result;

[0096] Specifically, the fault type identification unit will repeat this process until a certain termination condition is met. The termination condition can be that no (or a minimum number) of objects are reallocated to different clusters, no (or a minimum number) of cluster centers change again, and the error sum of squares is locally minimized;

[0097] Among them, a represents the initial selected value of the initial cluster center, k represents the selected number of clusters, and c i Indicates the class that is closest to sample i among the k classes, a j Represents the cluster center of the cluster.

[0098] It can be understood that by calculating the Euclidean distance between the feature vector of the track circuit physical characteristic diagram and the closest feature vector in the database, it is possible to quickly and effectively determine whether the current track circuit physical characteristic diagram has an abnormality, and then use the cluster center method to determine the fault type and number of faults in the abnormal track circuit physical characteristic diagram, so as to realize the intelligent track circuit defect detection and fault identification system. The system can conveniently realize track circuit defect detection, which is conducive to reducing the manufacturing cost of related equipment, and is also conducive to the intelligence of related equipment, and can support staff to quickly and effectively reach the fault site for error correction and maintenance, thereby improving the efficiency of track circuit maintenance and reducing the workload of track circuit maintenance.

[0099] As a preferred embodiment, the fault type identification unit uses a preset PatchCore model to identify the fault type of the track circuit physical characteristic diagram with abnormalities, and the process includes:

[0100] The PatchCore model uses distribution drift to divide the image into multiple local patches, and uses the residual neural network and the local domain formula to extract abnormal features. The expression is:

[0101]

[0102] At the same time, local abnormal features are extracted, and the expression is:

[0103]

[0104] The extracted abnormal features and local abnormal features are compared with the track circuit physical feature map in the original database to identify the fault type and fault quantity of the track circuit physical feature map with abnormalities;

[0105] in, It represents a patch of abnormal feature (patch) with a size of p*p at the position (h, w) on the feature map. h represents the horizontal coordinate of the feature map position, w represents the vertical coordinate of the feature map position, a represents the horizontal coordinate of the abnormal feature, b represents the vertical coordinate of the abnormal feature, and p represents the length of the normal pixel, usually p=3. represents the feature map tensor, f agg Represents the aggregation function of the neighborhood feature vector.

[0106] Specifically, see Figure 5 , by selecting the most representative feature points to approximate the original features, and at the same time building a feature vector memory pool, only retaining 1% to 10% of the number of features, to achieve efficient feature screening. The advantage of this is that it can reduce memory usage while still retaining as much information as possible, thereby improving the efficiency and performance of the algorithm.

[0107] Before using the preset PatchCore model to identify the fault type of the physical feature map of the track circuit with abnormalities, the image anomaly detection and judgment unit tests the pre-trained PatchCore model: the test sample's ability to extract patch features, and uses the nearest neighbor method to calculate the Euclidean distance to obtain the anomaly score. This step mainly involves using the extracted patch features to compare with the features in the memory pool to calculate the anomaly score.

[0108] The process of pre-training the PatchCore model includes: analyzing and memorizing about 20,000 abnormal images to avoid unknown abnormal patterns that cannot match the training supervision model, and obtaining feature center blocks. The model generated by pre-training can cope with more than 150,000 images of different object categories.

[0109] The feature map of the extracted abnormal image is modeled: first, the blocks of the image to be detected are evaluated and used with the trained model output. The most representative feature points are selected using KNN Greedy CoreSet sampling, and a feature vector memory pool is constructed. Only 1% to 10% of the features are retained. The maximum distance of any one of these blocks from the core set to its nearest neighbor is taken as the anomaly score, and finally assigned to the abnormal image, thereby achieving efficient feature screening and using it for anomaly detection.

[0110] It can be understood that the preset PatchCore model can quickly and efficiently extract the abnormal features and local abnormal features of the abnormal track circuit physical characteristic diagram during the process of identifying the fault type of the abnormal track circuit physical characteristic diagram, and then compare the features with the track circuit physical characteristic diagram in the original database to identify the fault type and number of the abnormal track circuit physical characteristic diagram, thereby realizing efficient feature screening.

[0111] As a preferred embodiment, see Figure 1 , the system also includes a result output module, which is convenient for wireless communication connection with the real-time image acquisition module and the real-time Tiny computing chip module, and the result output module includes a vehicle-mounted matching display screen and a dispatch center matching display screen;

[0112] The vehicle-mounted display screen and the dispatch center display screen are both used to visualize the fault type of the track circuit physical characteristic diagram output by the fault type identification unit.

[0113] As a preferred embodiment, see Figure 1 The system also includes a communication module, which is used to establish wireless communication connections for various modules in the system.

[0114] As a preferred embodiment, see Figure 1 The system also includes a control chip, which is used to control the real-time image acquisition module, the real-time Tiny computing chip module, the result output module and the communication module to complete corresponding tasks.

[0115] It can be understood that the on-board matching display screen and the dispatch center matching display screen can enable the staff to intuitively view the physical defects in the track circuit, so as to support the staff to quickly and effectively reach the fault site for error correction and repair, thereby improving the efficiency of track circuit maintenance and reducing the workload of track circuit maintenance; the communication module can provide a safe and reliable communication environment for the connection between each module in the system to maintain the safety and stability of the system; the control chip can control each module to perform corresponding work to ensure that the system can work normally and stably.

[0116] In this embodiment, the control chip controls the image processing unit to process the collected track image, extract the track feature map, and controls the image abnormality detection and judgment unit through the control chip to judge whether the abnormality index of the current track feature map is greater than or equal to a predetermined track abnormality index threshold, wherein the current track image abnormality index can be used to assess the difference between the track image and the abnormal sample: that is, if the current abnormality index is greater than or equal to the predetermined track abnormality index threshold, it can be determined that the difference between the target track image and the abnormal track image is as expected, then the control chip controls the image abnormality detection and judgment unit to transmit the image to the fault type recognition module, the control chip controls the fault type recognition module to compare the target image with the fault image in the database library, and then the control chip controls the result output module to output the result recognized by the fault type recognition module. If the current abnormality index is not greater than or equal to the predetermined track abnormality index threshold, it can be determined that the difference between the target track image and the abnormal track image is not as expected, and the control chip does not proceed to the next step; the system that realizes intelligent track circuit defect detection and fault identification can easily realize track circuit defect detection, which is conducive to reducing the manufacturing cost of related equipment, and is also conducive to the intelligence of related equipment, and can support staff to quickly and effectively reach the fault site for error correction and maintenance, thereby improving the efficiency of track circuit maintenance and reducing the workload of track circuit maintenance.

[0117] The storage resources and computing resources required by the track physical defect detection and fault identification system disclosed in the present invention mainly lie in the processing of image processing, defect detection and fault clustering involved in the step of judging the fault type according to the target track image clustering. Compared with the existing technical solutions for track physical defect detection based on deep learning technology, the track physical defect detection and fault identification system disclosed in the present invention has lower configuration requirements for storage resources and computing resources during actual application.

[0118] Embodiment three:

[0119] This embodiment provides a method for intelligent rail defect detection and fault identification. Figure 6 , the method comprises the following steps:

[0120] S1: Real-time acquisition of track circuit images along the train path;

[0121] S2: Extract the physical feature map of the track circuit in real time based on the track circuit image;

[0122] S3: perform track defect detection according to the extracted track circuit physical feature map and output the detection result;

[0123] S4: According to the track defect detection results, complete the track fault identification and output the identification results.

[0124] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. An intelligent track defect detection and fault identification system, characterized in that: The system comprises: a real-time image acquisition module and a real-time Tiny computing chip module, wherein the real-time image acquisition module is wirelessly connected to the real-time Tiny computing chip module; the real-time Tiny computing chip module comprises an image processing unit, an image anomaly detection and judgment unit, and a fault type identification unit; The real-time image acquisition module acquires the track circuit image in the train path in real time, and transmits the acquired track circuit image to the real-time Tiny computing chip module; The image processing unit extracts a physical characteristic diagram of the track circuit in real time based on the track circuit image; the image abnormality detection and judgment unit performs track defect detection based on the extracted physical characteristic diagram of the track circuit and outputs the detection result; the fault type identification unit completes track fault identification based on the track defect detection result and outputs the identification result.

2. The intelligent track defect detection and fault identification system according to claim 1 is characterized in that: The real-time image acquisition module includes a high-definition image imaging component and a synchronization device component; The high-definition image imaging component includes several groups of high-speed area array cameras and matching fill-light LED lights, which are used to collect real-time images of the track circuit; The synchronization device component is used to feed back the mileage information and position information of the train running on the track circuit in real time, and feed back the position information of the physical characteristic diagram of the track circuit.

3. The intelligent track defect detection and fault identification system according to claim 1 is characterized in that: The image processing unit includes a fully connected layer and a fully convolutional layer; The image processing unit divides the track circuit physical feature map into a candidate frame feature map and a common feature map, and keeps the sizes of the candidate frame feature map and the common feature map consistent; The fully connected layer is used to extract the position coordinates and category information in the candidate frame feature map; The fully convolutional layer is used to extract segmentation information in the common feature map; Among them, the position coordinates and category information are represented, and the segmentation information is represented.

4. The intelligent track defect detection and fault identification system according to claim 3 is characterized in that: The image anomaly detection and judgment unit compares the local patch of the track circuit physical feature map with the image feature vector stored in the database, and calculates the Euclidean distance between the feature vector of the track circuit physical feature map and the closest feature vector in the database, which is expressed as: Determine whether the calculated Euclidean distance exceeds a preset threshold; If it exceeds, the track circuit physical characteristic diagram currently undergoing track defect detection is abnormal, and the image abnormality detection and judgment unit transmits the calculated Euclidean distance and the abnormal result to the fault type identification unit; If it does not exceed, it means that the physical characteristic diagram of the track circuit currently undergoing track defect detection is normal; Where n represents the number of dimensions, x = x1, x2, ..., x n and y = y1, y2, ..., y n They represent the coordinates of two points in the feature space, and d(x,y) represents the Euclidean distance between the track circuit patch and the feature vector in the database.

5. The intelligent track defect detection and fault identification system according to claim 4 is characterized in that: The fault type identification unit selects k fault samples for initialization as the initial clustering center according to the received Euclidean distance and abnormal result, and the expression is: <h2 style=";text-align:left;direction:ltr">a = a1, a2,...a<h2 style=";text-align:left;direction:ltr"> k <h2 style=";text-align:left;direction:ltr"> ; The fault type identification unit assigns the track circuit physical characteristic diagram with abnormality to the cluster center closest to it, and regards the cluster center and the assigned data points as a cluster; When the fault type identification unit assigns a new track circuit physical characteristic diagram with abnormalities to any cluster, the cluster center of the current cluster is recalculated, and the expression is: When the number of faults present in the abnormal track circuit physical characteristic diagram identified by the fault type identification unit is greater than a selected cluster number threshold, the fault type identification unit outputs a fault identification result; Among them, a represents the initial selected value of the initial cluster center, k represents the selected number of clusters, and c i Indicates the class that is closest to sample i among the k classes, a j Represents the cluster center of the cluster.

6. The intelligent track defect detection and fault identification system according to claim 5 is characterized in that: The fault type identification unit uses a preset PatchCore model to identify the fault type of the track circuit physical characteristic diagram with abnormalities, and the process includes: The PatchCore model uses distribution drift to divide the image into multiple local patches, and uses the residual neural network and the local domain formula to extract abnormal features. The expression is: At the same time, local abnormal features are extracted, and the expression is: The extracted abnormal features and local abnormal features are compared with the track circuit physical feature map in the original database to identify the fault type and fault quantity of the track circuit physical feature map with abnormalities; in, It represents a patch of abnormal feature (patch) with a size of p*p at the position (h, w) on the feature map. h represents the horizontal coordinate of the feature map position, w represents the vertical coordinate of the feature map position, a represents the horizontal coordinate of the abnormal feature, b represents the vertical coordinate of the abnormal feature, and p represents the length of the normal pixel, usually p=3. represents the feature map tensor, f agg Represents the aggregation function of the neighborhood feature vector.

7. The intelligent track defect detection and fault identification system according to claim 1 is characterized in that: The system also includes a result output module, which is convenient for wireless communication connection with the real-time image acquisition module and the real-time Tiny computing chip module, and the result output module includes a vehicle-mounted matching display screen and a dispatch center matching display screen; The vehicle-mounted display screen and the dispatch center display screen are both used to visualize the fault type of the track circuit physical characteristic diagram output by the fault type identification unit.

8. The intelligent track defect detection and fault identification system according to claim 7 is characterized in that: The system further comprises a communication module, which is used to establish wireless communication connections for various modules in the system.

9. The intelligent track defect detection and fault identification system according to claim 7, characterized in that: The system also includes a control chip, which is used to control the real-time image acquisition module, the real-time Tiny computing chip module, the result output module and the communication module to complete corresponding tasks.

10. An intelligent method for rail defect detection and fault identification, characterized in that: The method comprises the following steps: Real-time acquisition of track circuit images along the train path; According to the track circuit image, the physical characteristic map of the track circuit is extracted in real time; Perform track defect detection based on the extracted track circuit physical feature map and output the detection results; According to the track defect detection results, track fault identification is completed and the identification results are output.

Images