Method and system for detecting close attachment of switch rail of turnout equipment based on monitoring image

Abstract

The invention provides a switch rail close-fitting detection method and system for turnout equipment based on a monitoring image, and belongs to the technical field of image detection.The method specifically comprises the steps that setting processing is conducted on a detection label according to a set processing proportion, and according to the matching condition of detection data of the detection label and an identification result of a close-fitting detection model, the identification result of the close-fitting detection model is obtained; and determining an identification processing strategy for adding the monitoring images in different target areas into a training library, and carrying out updating training processing on the close contact detection model based on the training library to obtain an updating model so as to update the change condition of the risk detection target area of the model. According to the method and the device, the adjustment scheme of the setting mode of the detection label closely attached to the detection model is determined according to different target areas and the recognition processing strategies added to the training library in the different target areas, so that the updating processing efficiency of the model is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image detection, and in particular relates to a method and system for detecting the close contact of a switch rail of a switch device based on monitoring images. Background Art

[0002] As one of the three major outdoor components of signaling equipment (turnout switch, signal, and track circuit), turnouts are used to change the direction of train travel and can significantly improve track capacity. However, their complex structure and high technical requirements make them a top priority for railway safety maintenance. The switch rail is a crucial component of the switch machine, and its quality determines the overall quality of the turnout. However, the switch rail is relatively fragile and is exposed to harsh outdoor environments, constantly bearing the heavy impact of train changes. This can easily lead to deformation, wear, and poor adhesion to the stock rail. Switch rail failures account for over 45% of turnout defects, affecting train stability in the turnout area at the very least, and causing derailment at the worst. Therefore, the prompt detection and correction of switch rail failures is a crucial component of turnout maintenance.

[0003] Existing methods, both domestically and internationally, mostly employ structured light / laser approaches. While structured light has proven its application in other fields, it is susceptible to ambient light, limiting its outdoor use. Accuracy also decreases significantly with increasing detection distance. Furthermore, the lifespan of the activator is short, preventing long-term continuous detection. The calibration process is complex when replacing the laser, significantly increasing costs. Some high-precision laser detectors, such as LiDAR, can cost tens of thousands of dollars.

[0004] To address the above technical issues, the invention patent application CN202510158580.5, "High-precision intelligent detection and detection method of switch point rail structure status based on binocular vision," provides an image-based detection and analysis strategy for the switch point rail structure status, which improves the accuracy of the recognition and processing of the close-fitting state. However, the above technical solution has the following technical problems: In the measurement of the close fitting seam of the point rail, the close fitting detection model often has low detection accuracy during the initial recognition processing due to the small amount of training data. Therefore, how to adjust the setting method of the detection labels in different detection areas according to the recognition results of the close fitting detection model after the update processing, that is, how to adjust the labels used to measure the distance between the point rail and the base rail, so as to ensure the training processing efficiency and accuracy of the close fitting detection model during the update training processing, has become a technical problem that needs to be solved urgently.

[0005] In order to solve the above technical problems, the present application provides a method and system for detecting the close contact of the switch equipment's point rail based on monitoring images. Summary of the Invention

[0006] To achieve the purpose of the present invention, the present invention adopts the following technical solutions: Specifically, the application provides a method for detecting the close contact of a switch rail based on monitoring images, which specifically includes: S1 constructs a close fit detection model for the point rail, determines a matching situation with close fit detection results in different target areas based on a recognition result of the close fit detection model, and determines a risk detection target area in the target area using the matching situation; S2 determines the setting processing ratio of the detection tags in the target area based on the deviation of the close detection results of different risk detection target areas; S3 performs setting processing of the detection tags according to the set processing ratio, and determines the recognition processing strategy for adding monitoring images in different target areas to the training library based on the matching between the detection data of the detection tags and the recognition results of the close detection model; S4 performs update training processing on the close fitting detection model based on the training library to obtain an updated model, so as to determine an adjustment plan for setting the detection label of the close fitting detection model based on changes in the risk detection target area of ​​the updated model and the recognition processing strategies in different target areas added to the training library.

[0007] The beneficial effects of the present invention are: According to the matching between the detection data of the detection label and the recognition results of the close fitting detection model, the recognition processing strategy for adding the monitoring images in different target areas to the training library is determined, thereby taking into account the deviation between the detection data of the detection label and the recognition results of the model, and then realizing an accurate evaluation of the recognition processing reliability of the monitoring images in different target areas from the perspective of the deviation, ensuring that more monitoring images in target areas with higher deviations can be added to the training library, thereby improving the training processing efficiency of the close fitting detection model.

[0008] Based on the changes in the risk detection target areas of the updated model and the recognition processing strategies added to the training library in different target areas, the adjustment plan for the setting method of the detection labels of the close-fitting detection model is determined. The difference in the reliability of the update processing of the updated model due to the difference in the number of newly added risk detection target areas is taken into account, and the difference in the amount of data in the data update processing in the training library due to the difference in recognition processing strategies is also taken into account. Thus, the adjustment plan for the setting method of the detection labels of the close-fitting detection model is determined from the two perspectives of the reliability of the model update processing and the change in the amount of data in the training library, thereby further improving the efficiency of the training processing of the close-fitting detection model.

[0009] A further technical solution is that the close fitting detection model of the point rail is constructed using one or more of an end-to-end railway turnout point rail close fitting seam scale detection model based on monocular vision and VisionTransformer technology, an image recognition model based on convolutional neural network, and an image recognition model based on RNN.

[0010] A further technical solution is that the target area is a road area where the switch equipment's point rail close contact detection is required.

[0011] A further technical solution is that the matching condition of the close fitting detection result is determined according to the deviation between the close fitting detection result and the actual measurement result.

[0012] A further technical solution is that a method for determining a risk detection target area in the target area is: Determine the deviation between the close fit detection result and the actual measurement result in the target area using the matching condition; determining a number of detection deviations in the target area based on the deviation amount; The number of detection deviations is used to determine whether the target area is a risk detection target area.

[0013] A further technical solution is that the method for determining the adjustment scheme of the setting mode of the detection tags of the close detection model is: Based on the change in the risk detection target area of ​​the updated model, newly added risk monitoring target areas of the updated model after the update compared to before the update are determined, and the newly added target areas are used as the newly added target areas; based on the change in the risk detection target area of ​​the updated model, target areas that no longer belong to the risk monitoring target areas of the updated model after the update compared to before the update are determined, and the changed target areas are used; Determining recognition processing types in different target areas using recognition processing strategies added to the training library in different target areas; Based on the recognition processing type, the number of newly added target areas and the number of changed target areas, an adjustment scheme for setting the detection tags of the close detection model is determined.

[0014] In a second aspect, the present invention provides a computer system comprising: a memory and a processor in communication connection, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the above-mentioned method for detecting the close fit of a point rail of a switch equipment based on monitoring images when running the computer program.

[0015] Other features and advantages will be described in the following description, and in part will become apparent from the description, or understood by practicing the invention. The purpose and other advantages of the invention are realized and obtained by the structures particularly pointed out in the description and the drawings.

[0016] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The above and other features and advantages of the present invention will become more apparent by describing in detail exemplary embodiments thereof with reference to the accompanying drawings.

[0018] Figure 1 The present invention is a flow chart of a method for detecting the close contact of a switch rail device based on monitoring images; Figure 2 is a flow chart of a method for determining a target area for risk detection in a target area; Figure 3 is a flow chart of a method for determining a set processing ratio of a detection tag in a target area; Figure 4 It is a flow chart of a method for determining a recognition processing strategy for adding monitoring images within a target area to a training library. DETAILED DESCRIPTION

[0019] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be embodied in many forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete and will fully convey the concepts of the example embodiments to those skilled in the art. Like reference numerals in the figures represent like or similar structures, and thus their detailed description will be omitted.

[0020] The terms "a", "an", "the", and "said" are used to indicate the presence of one or more elements / components / etc.; the terms "including" and "having" are used to express an open-ended inclusive meaning and mean that additional elements / components / etc. may be present in addition to the listed elements / components / etc.

[0021] Example 1 To solve the above problems, according to one aspect of the present invention, Figure 1 As shown, a method for detecting the close contact of a switch rail device based on monitoring images is provided, which specifically includes: S1 constructs a close fit detection model for the point rail, determines a matching situation with close fit detection results in different target areas based on a recognition result of the close fit detection model, and determines a risk detection target area in the target area using the matching situation; Furthermore, the point rail fit detection model is constructed using one or more of an end-to-end railway turnout point rail fit seam scale detection model based on monocular vision and Vision Transformer technology, an image recognition model based on convolutional neural network, and an image recognition model based on RNN.

[0022] Specifically, the target area is a road area where the switch equipment's point rail close contact detection is required.

[0023] It should be noted that the matching condition of the close fitting detection result is determined according to the deviation between the close fitting detection result and the actual measurement result.

[0024] Specifically, such as Figure 2 As shown, the method for determining the risk detection target area in the target area is: Determine the deviation between the close fit detection result and the actual measurement result in the target area using the matching condition; determining a number of detection deviations in the target area based on the deviation amount; The number of detection deviations is used to determine whether the target area is a risk detection target area.

[0025] It can be understood that when the number of detection deviations in the target area does not meet the requirement, that is, is greater than a preset threshold, the target area is determined to be a risk detection target area.

[0026] It should also be noted that the number of detection deviations is the number of detections in which the deviation amount is greater than a preset deviation threshold.

[0027] Optionally, a method for determining a risk detection target area in the target area is: Determine the deviation between the close fit detection result and the actual measurement result in the target area using the matching condition; Determine an average value of the deviation of the target area under different detection times based on the deviation; An average value of the deviations is used to determine whether the target area is a risk detection target area.

[0028] It should be noted that, when the average value of the deviation is greater than a preset deviation threshold, the target area is determined to be a risk detection target area.

[0029] S2 determines the setting processing ratio of the detection tags in the target area based on the deviation of the close detection results of different risk detection target areas; Specifically, such as Figure 3 As shown, the method for determining the setting processing ratio of the detection tag in the target area is: Determining the number of detection deviations for different risk detection target areas based on the deviation of the close fit detection results for the risk detection target area; Determining the proportion of the risk detection target area in the target area by using the composition data of the risk detection target area in the target area; Based on the number of detection deviations and the proportion of the constituent quantities, the setting processing ratio of the detection tags in the target area is determined.

[0030] It is understandable that, based on the number of detection deviations and the proportion of the number of components, determining the setting processing ratio of the detection tags in the target area specifically includes: Determining a deviation detection ratio according to a ratio of the number of detection deviations in the risk detection target area to the risk detection target area; Determine the detection deviation value by taking the average value of the deviation detection proportion and the average value of the constituent quantity proportion; A setting processing ratio of the detection tag in the target area is determined based on the detection deviation value.

[0031] It can be understood that the detection tag is a tag used to detect the spacing distance between the base rail and the point rail, and is specifically constructed using an RFID tag.

[0032] It can be understood that when the detection deviation value is greater than the preset monitoring deviation threshold, the preset setting ratio is used to determine the setting processing ratio of the detection tag in the target area; when the detection deviation value is greater than the second preset monitoring deviation threshold, the second preset setting ratio is used to determine the setting processing ratio of the detection tag in the target area. In a possible embodiment, when it is greater than 0.4, the preset setting ratio, i.e., 0.6, is used to determine the setting processing ratio of the detection tag in the target area; when it is greater than 0.2, i.e., above 0.2, the preset setting ratio, i.e., 0.4, is used to determine the setting processing ratio of the detection tag in the target area; in other cases, the setting processing of the detection tag is performed in the risk detection target area.

[0033] In addition, it can be understood that the above-mentioned setting processing ratio is to determine the target area for setting the detection label according to the preset setting ratio in the target area excluding the risk detection target area. Specifically, if the number of target areas excluding the risk detection target area is 10, then according to the setting ratio of 0.4, the number of target areas that need to be set for detection labels is 4. Regardless of the detection deviation value, the setting processing of the detection label is required in the risk detection target area.

[0034] Optionally, the method for determining the setting processing ratio of the detection tag in the target area is: S21 determines the proportion of the risk detection target area in the target area by using the composition data of the risk detection target area in the target area; Specifically, in the above steps, if the number of risk detection target areas is large, that is, the proportion of the risk detection target areas in the target area is high, that is, it is greater than the preset proportion threshold, therefore, at this time it means that the detection accuracy of the close-fitting detection model is not high, so on this basis, it can be directly determined to adopt the preset setting ratio to determine the setting processing ratio of the detection tag in the target area.

[0035] In addition, it should be noted that if the proportion of the risk detection target area in the target area is not greater than the preset proportion threshold, it is necessary to further determine whether the number of risk detection target areas is too small. In a possible embodiment, when it is less than 3, it means that the detection accuracy of the close-fitting detection model is high, so on this basis, it can be directly determined to set the detection label in the risk detection target area, and other target areas do not need to set the detection label.

[0036] If the number of risk detection target areas is not too small, it is necessary to proceed to the next step to determine the deviation detection ratio.

[0037] S22 determines the number of detection deviations of different target areas based on the deviation of the close-fitting detection results of the target area, and determines the proportion of the number of detection deviations in the number of detections in different risk detection target areas based on the number of detection deviations, and uses the proportion as the deviation detection proportion; Furthermore, in a possible embodiment, if the sum of the number of detection deviations of different target areas in the above steps is large, that is, greater than the threshold, it means that the detection accuracy of the close-fitting detection model is not high, so on this basis, it can be directly determined to adopt the preset setting ratio to determine the setting processing ratio of the detection tag in the target area.

[0038] It can also be understood that if the sum of the number of detection deviations of different target areas is not greater than the threshold, it is necessary to further determine whether there is a risk detection target area where the detection deviation ratio does not meet the requirements. In a possible embodiment, that is, when the detection deviation ratio is greater than 0.7 of the risk detection target area, it means that the detection accuracy of the close-fitting detection model is not high, so on this basis, it can be directly determined to adopt the preset setting ratio to determine the setting processing ratio of the detection tag in the target area.

[0039] If there is no risk detection target area where the detection deviation ratio does not meet the requirements, it is necessary to further determine whether the average value of the deviation detection ratio of different risk detection target areas meets the requirements. It should be noted that when the average value of the deviation detection ratio of different risk detection target areas is greater than 0.5, it means that the detection accuracy of the close-fitting detection model is not high. Therefore, on this basis, it can be directly determined to adopt the preset setting ratio to determine the setting processing ratio of the detection label in the target area. In other cases, proceed to the next step.

[0040] S23 determines the setting processing ratio of the detection tags in the target area based on the deviation detection ratio and the component quantity ratio of different target areas.

[0041] In a possible embodiment, the detection deviation value is determined by the average value of the deviation detection ratio and the average value of the component quantity ratio of different target areas, and the setting processing ratio of the detection tag in the target area is determined based on the detection deviation value.

[0042] S3 performs setting processing of the detection tags according to the set processing ratio, and determines the recognition processing strategy for adding monitoring images in different target areas to the training library based on the matching between the detection data of the detection tags and the recognition results of the close detection model; Specifically, such as Figure 4 As shown, the method for determining the recognition processing strategy for adding the monitoring images in the target area to the training library is: Determine a time period during which the spacing distance between the point rail and the base rail in the target area varies based on a match between the detection data of the detection tags in the target area and the recognition results of the close fitting detection model, and use the time period as the time period during which the spacing distance varies; Determining detection inconsistency moments in different interval distance variation periods based on matching between detection data in different interval distance variation periods and recognition results of the close fit detection model; Based on the distribution of detection inconsistency moments in different interval distance change periods, a recognition processing strategy for adding monitoring images within the target area to a training library is determined.

[0043] It is understandable that as the train approaches, the distance between the point rail and the base rail will inevitably change due to vibration, which is the period of distance change. Therefore, by combining the matching situation when the distance changes, the determination of the differentiated recognition processing strategy for adding to the training library is achieved.

[0044] It can be understood that the detection inconsistency moment is the moment when the deviation between the detection data of the detection tag and the recognition result of the close detection model is not within the preset deviation range.

[0045] The interval distance change moment is determined based on the detection data of the detection tag. Specifically, the moment when the deviation from the average value of the detection data at different moments is greater than a preset detection data deviation threshold is used as the interval distance change moment.

[0046] Specifically, when there is a time period of interval distance change in which the proportion of the number of detection inconsistency moments does not meet the requirements, that is, when there is a time period of interval distance change in which the proportion of the number of detection inconsistency moments is greater than 0.3, the recognition processing strategy for adding the monitoring images in the target area to the training library is to perform recognition processing in different time periods of interval distance change, and add all monitoring images in the time period of interval distance change in which the proportion of the number of detection inconsistency moments is greater than the first proportion threshold to the training library.

[0047] It should also be noted that when there is no interval distance change period in which the proportion of the number of detection inconsistency moments does not meet the requirements, that is, when there is no interval distance change period in which the proportion of the number of detection inconsistency moments is greater than 0.3, the recognition processing strategy for determining whether the monitoring images in the target area are added to the training library is to only perform recognition processing on the interval distance change period in which the number of interval distance change moments is greater than the preset number of moments, and to add all monitoring images in the interval distance change period in which the number of detection inconsistency moments is greater than the second proportion threshold and the proportion of the number of detection moments is greater than the second proportion threshold to the training library, where the first proportion threshold is less than the second proportion threshold.

[0048] It should be noted that the interval distance change period is a period in which there are interval distance change moments and the number of interval moments between different adjacent interval distance change moments is less than a preset interval moment number threshold.

[0049] Optionally, a method for determining a recognition processing strategy for adding monitoring images within the target area to a training library is: S31 determines a time period of variation in the spacing distance between the point rail and the base rail in the target area based on a match between the detection data of the detection tags in the target area and the recognition results of the close fitting detection model, and uses the time period of variation in the spacing distance as the time period of variation in the spacing distance; It should be noted that in the above steps, when the target area belongs to the risk detection target area, identification processing is performed in different interval distance change periods, and all monitoring images in the interval distance change period where there are inconsistent detection moments are added to the training library.

[0050] In addition, it should be noted that when the target area does not belong to the risk detection target area, the number of interval distance change periods in the target area is obtained. When the number of interval distance change periods in the target area is small, that is, the average number of interval distance change periods on different dates is less than 5, then for the reliability of the model's recognition processing, recognition processing is performed in different interval distance change periods in the target area, and all monitoring images in the interval distance change periods where there are inconsistent detection moments are added to the training library.

[0051] It can also be understood that when the number of interval distance change periods in the target area is not small, and the average number of interval distance change periods on different dates is greater than 20, only the interval distance change periods in which the number of interval distance change moments is greater than the preset number of moments are identified and processed, and all monitoring images in the interval distance change periods in which the number of inconsistent detection moments is greater than the second proportion threshold are added to the training library.

[0052] S32: determining the detection inconsistency moments in the different interval distance variation periods based on the matching between the detection data in the different interval distance variation periods and the recognition results of the close fitting detection model, and determining the change values ​​in the different interval distance variation periods based on the distribution data of the different detection inconsistency moments in the different interval distance variation periods; It can also be understood that, in a possible embodiment, when there is a time interval distance change period in which the number of detection inconsistency moments does not meet the requirements, that is, when there is a time interval distance change period in which the number of detection inconsistency moments accounts for more than 0.3, then the recognition processing strategy for adding the monitoring images in the target area to the training library is to perform recognition processing in different time interval distance change periods, and add all monitoring images in the time interval distance change period in which the number of detection inconsistency moments accounts for more than the first proportion threshold to the training library.

[0053] It should also be noted that, when there is no interval distance variation period in which the ratio of the number of inconsistent detection moments does not meet the requirement, the variation value is determined according to the ratio of the number of inconsistent detection moments in different interval distance variation periods.

[0054] It can be understood that when there is no interval distance variation period that does not meet the requirement of the variation value, that is, there is no interval distance variation period in which the proportion of the number of detection inconsistent time points is greater than 0.3, at this time, it is necessary to further determine the number of interval distance variation periods in which the detection inconsistent time points exist, and when the number of interval distance variation periods in which the detection inconsistent time points exist is less than the preset period number threshold, at this time, only the interval distance variation period in which the number of interval distance variation time points is greater than the preset time point number is identified and processed, and all the monitoring images in the interval distance variation period in which the proportion of the number of detection inconsistent time points is greater than the second proportion threshold are added to the training library.

[0055] S33 determines the identification processing strategy of the monitoring images in the target area into the training library based on the variation values in different interval distance variation periods, the distribution data of the interval distance variation periods in different dates, and the distribution data of the distance variation periods in which the detection inconsistent time points exist.

[0056] It should be noted that in one possible embodiment, the comprehensive variation value is determined based on the average value of the variation values in different interval distance variation periods and the average value of the proportion of the number of interval distance variation periods in which the detection inconsistent time points exist. Based on the preset period number threshold corresponding to the comprehensive variation value, it can be understood that the larger the comprehensive variation value is, the larger the preset period number threshold is. When the number of interval distance variation periods in different dates is greater than the preset period number threshold, it indicates that the number of interval distance variation periods under the current variation condition is relatively large, and therefore only the interval distance variation period in which the number of interval distance variation time points is greater than the preset time point number is identified and processed, and all the monitoring images in the interval distance variation period in which the proportion of the number of detection inconsistent time points is greater than the second proportion threshold are added to the training library.

[0057] If the number of interval distance variation periods in different dates is not greater than the preset period number threshold, the identification processing strategy of the monitoring images in the target area into the training library is determined to be identification processing in different interval distance variation periods, and all the monitoring images in the interval distance variation period in which the proportion of the number of detection inconsistent time points is greater than the first proportion threshold are added to the training library.

[0058] It can be understood that the monitoring images in the training library are used as training data, and the training data is used to update and train the close detection model to obtain an updated close detection model, which is used as an updated model.

[0059] S4 performs update training processing on the close fitting detection model based on the training library to obtain an updated model, so as to determine an adjustment plan for setting the detection label of the close fitting detection model based on changes in the risk detection target area of ​​the updated model and the recognition processing strategies in different target areas added to the training library.

[0060] Furthermore, the method for determining the adjustment scheme of the setting mode of the detection tags of the close fitting detection model is: Based on the change in the risk detection target area of ​​the updated model, newly added risk monitoring target areas of the updated model after the update compared to before the update are determined, and the newly added target areas are used as the newly added target areas; based on the change in the risk detection target area of ​​the updated model, target areas that no longer belong to the risk monitoring target areas of the updated model after the update compared to before the update are determined, and the changed target areas are used; Determining recognition processing types in different target areas using recognition processing strategies added to the training library in different target areas; Based on the recognition processing type, the number of newly added target areas and the number of changed target areas, an adjustment scheme for setting the detection tags of the close detection model is determined.

[0061] In a possible embodiment, when there is no newly added target area, there is no need to perform a recognition processing strategy of adding monitoring images in different target areas to the training library and adjust the target area of ​​the detection label.

[0062] When there are new target areas, when the ratio of the number of new target areas to the number of changed target areas is greater than the preset ratio threshold, detection labels are set in different target areas, and the first recognition processing type is used to determine the recognition processing strategy for adding monitoring images to the training library.

[0063] When the ratio of the number of newly added target areas to the number of changed target areas is not greater than the preset ratio threshold, it is determined whether the number of target areas of different recognition processing types is greater than the preset value of the number of target areas. If so, there is no need to implement the recognition processing strategy of adding monitoring images in different target areas to the training library and the adjustment processing of the target areas that need to be detected with labels. If not, the recognition processing types of different target areas are all set to the first recognition processing type.

[0064] Example 2 In a second aspect, the present invention provides a computer system comprising: a memory and a processor in communication connection, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the above-mentioned method for detecting the close fit of a point rail of a switch equipment based on monitoring images when running the computer program.

[0065] It should also be noted that when the number of newly added target areas does not meet the requirements, that is, when it is greater than the preset target area number threshold, the changes in the risk monitoring target areas at this time are more drastic. Therefore, on this basis, detection tags need to be set in different target areas.

[0066] It can also be understood that when the number of newly added target areas meets the requirements, but the difference between the number of newly added target areas and the number of changed target areas is positive and greater than the preset value of the number of areas, it means that the number of newly added risk detection target areas is large. Therefore, on this basis, in order to ensure the reliability of the identification processing, detection labels are set in different target areas.

[0067] Furthermore, when the number of newly added target areas meets the requirements and the difference between the number of newly added target areas and the number of changed target areas is not positive or not greater than the preset value of the area number, the updated change risk value is determined based on the ratio of the number of newly added target areas to the number of changed target areas.

[0068] When the updated change risk value is greater than the preset change risk threshold, the change in the risk monitoring target area is relatively drastic. Therefore, on this basis, detection tags need to be set in different target areas.

[0069] However, if the updated change risk value is not greater than the preset change risk threshold, but is within a certain change risk value range, that is, the updated edge change risk value is within the preset change risk interval, then all target areas that do not belong to the first identification processing type need to be modified to the first identification processing type.

[0070] In addition, if the updated edge change risk value is not within the preset change risk range, there is no need to perform the recognition processing strategy of adding monitoring images in different target areas to the training library and the adjustment processing of the target area that needs to be detected.

[0071] Furthermore, the first recognition processing type is to perform recognition processing in different interval distance change periods, and add all monitoring images in the interval distance change period in which the proportion of the number of detection inconsistency moments is greater than a first proportion threshold to the training library.

[0072] The various embodiments in this specification are described in a progressive manner. Similar portions between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from the other embodiments. In particular, the device, apparatus, and non-volatile computer storage medium embodiments are generally similar to the method embodiments, so their descriptions are relatively simplified. For relevant details, refer to the descriptions of the method embodiments.

[0073] The foregoing description of this specification describes specific embodiments. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in an order different from that described in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order shown or the sequential order to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0074] The foregoing description is merely one or more embodiments of this specification and is not intended to limit this specification. It will be apparent to those skilled in the art that various modifications and variations may be made to one or more embodiments of this specification. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of one or more embodiments of this specification are intended to be within the scope of the claims of this specification.

Claims

1. A method for detecting the close contact of a switch rail based on monitoring images, characterized in that: Specifically include: Constructing a close fit detection model for the point rail, determining matching results with close fit detection results in different target areas based on recognition results of the close fit detection model, and determining a risk detection target area in the target area using the matching results; Based on the deviation of the close-fitting detection results of different risk detection target areas, determine the setting and processing ratio of detection tags in the target area; Performing setting processing of the detection tags according to the set processing ratio, and determining the recognition processing strategy for adding monitoring images in different target areas to the training library based on the matching between the detection data of the detection tags and the recognition results of the close-fitting detection model; Based on the training library, the close fitting detection model is updated and trained to obtain an updated model, so as to determine the adjustment scheme for the setting method of the detection label of the close fitting detection model based on the changes in the risk detection target area of ​​the updated model and the recognition processing strategies in different target areas added to the training library.

2. The method for detecting the close contact of a switch rail based on monitoring images according to claim 1, wherein: The switch rail close fitting detection model is constructed by using one or more of an end-to-end railway switch rail close fitting seam scale detection model based on monocular vision and Vision Transformer technology, an image recognition model based on convolutional neural network, and an image recognition model based on RNN.

3. The method for detecting the close contact of a switch rail based on monitoring images according to claim 1, wherein: The target area is a road area where the switch equipment's point rail close contact detection is required.

4. The method for detecting the close contact of a switch rail based on monitoring images according to claim 1, wherein: The matching condition of the close fitting detection result is determined according to the deviation between the close fitting detection result and the actual measurement result.

5. The method for detecting the close contact of a switch rail based on monitoring images according to claim 1, wherein: The method for determining the risk detection target area in the target area is: Determine the deviation between the close fit detection result and the actual measurement result in the target area using the matching condition; determining a number of detection deviations in the target area based on the deviation amount; The number of detection deviations is used to determine whether the target area is a risk detection target area.

6. The method for detecting the close contact of a switch rail based on monitoring images according to claim 1, wherein: The method for determining the setting processing ratio of the detection tag in the target area is: Determining the number of detection deviations for different risk detection target areas based on the deviation of the close fit detection results for the risk detection target area; Determining the proportion of the risk detection target area in the target area by using the composition data of the risk detection target area in the target area; Based on the number of detection deviations and the proportion of the constituent quantities, the setting processing ratio of the detection tags in the target area is determined.

7. The method for detecting the close contact of a switch rail based on monitoring images according to claim 6, wherein: Based on the number of detection deviations and the proportion of the number of components, determining the setting processing ratio of the detection tags in the target area specifically includes: Determining a deviation detection ratio according to a ratio of the number of detection deviations in the risk detection target area to the risk detection target area; Determine the detection deviation value by taking the average value of the deviation detection proportion and the average value of the constituent quantity proportion; A setting processing ratio of the detection tag in the target area is determined based on the detection deviation value.

8. The method for detecting the close contact of a switch rail based on monitoring images according to claim 1, wherein: The detection tag is a tag used for detecting the spacing distance between the base rail and the point rail, and is specifically constructed using an RFID tag.

9. The method for detecting the close contact of a switch rail based on monitoring images according to claim 1, wherein: The method for determining the adjustment scheme of the setting mode of the detection label of the close detection model is as follows: Based on the change in the risk detection target area of ​​the updated model, newly added risk monitoring target areas of the updated model after the update compared to before the update are determined, and the newly added target areas are used as the newly added target areas; based on the change in the risk detection target area of ​​the updated model, target areas that no longer belong to the risk monitoring target areas of the updated model after the update compared to before the update are determined, and the changed target areas are used; Determining recognition processing types in different target areas using recognition processing strategies added to the training library in different target areas; Based on the recognition processing type, the number of newly added target areas and the number of changed target areas, an adjustment scheme for setting the detection tags of the close detection model is determined.

10. A computer system comprising: A memory and a processor in communication connection, and a computer program stored in the memory and capable of running on the processor, characterized in that when the processor runs the computer program, it executes a method for detecting the close fit of a switch device based on monitoring images as described in any one of claims 1 to 9.

Images