A method for detecting and predicting stuckness of contact network compensation device

Abstract

The present invention relates to the technical field of detection and prediction of railway overhead line compensation devices, and to a method for detecting and predicting the jamming of overhead line compensation devices, comprising: collecting image data of overhead line compensation devices; constructing a target detection model based on a deep learning algorithm, wherein the target detection model performs component recognition and pole number recognition of the overhead line compensation device, associates the recognition results, and establishes a normalized reference coordinate system; obtains the operation data of the overhead line compensation device based on the normalized reference coordinate system, inputs the operation data of the overhead line compensation device into a displacement trend prediction model of the overhead line compensation device for calculation, obtains the displacement trend prediction result of the overhead line compensation device, verifies the prediction result, and outputs the jamming risk analysis result of the overhead line compensation device. The present invention realizes intelligent monitoring and analysis of the operation status of the overhead line compensation device by establishing an operation data collection system of a normalized reference coordinate system.

Description

Technical Field

[0001] The invention relates to the technical field of railway overhead line compensation device detection and prediction, and in particular to a method for detecting and predicting stuck of an overhead line compensation device. Background Art

[0002] The contact network compensation device is a key device to ensure the safe operation of the contact network of electrified railways. Its main function is to compensate for the expansion and contraction of the contact wire caused by temperature changes and maintain the constant tension of the contact network wire. With the rapid development of high-speed railways, the operating speed and quality requirements of the contact network are constantly improving, which puts higher requirements on the reliability and maintenance efficiency of the compensation device. At present, the detection of traditional contact network compensation devices mainly relies on manual inspection methods. The inspection personnel need to check the operating status of the compensation device on site regularly. Not only is the work intensity high, but the inspection efficiency is low. Especially under severe weather conditions, the inspection work is difficult to carry out normally, which easily leads to blind spots in the inspection. At the same time, the measurement accuracy of the manual inspection method is seriously dependent on the personal experience and technical level of the inspection personnel. The inspection results are highly subjective, and it is difficult to ensure the accuracy of the data, which makes it impossible to effectively implement the preventive maintenance and fault warning of the equipment, increasing the safety risk of equipment operation. Summary of the invention

[0003] The purpose of the present invention is to provide a method for detecting and predicting the jamming of a contact network compensation device, which not only realizes the automatic identification and precise positioning of the contact network compensation device through a deep learning algorithm, but also establishes an operation data acquisition system based on a normalized reference coordinate system, thereby realizing the intelligent monitoring and analysis of the operation status of the compensation device. By deploying a contact network safety inspection device, combined with a target detection model and a displacement trend prediction model, the detection efficiency and accuracy of the contact network compensation device are effectively improved, and all-weather, high-precision automatic monitoring is realized, which provides reliable data support for the preventive maintenance of the contact network system and significantly improves the intelligence level of contact network operation and maintenance.

[0004] The present invention is achieved through the following technical solutions:

[0005] A method for detecting and predicting stuckness of a contact network compensation device, the method comprising the steps of:

[0006] Collect image data of the overhead line compensation device based on the overhead line safety inspection device;

[0007] Preprocess and calibrate the image data of the overhead line compensation device;

[0008] Building a target detection model based on a deep learning algorithm, the target detection model performs component recognition of the overhead line compensation device and pole number recognition of the overhead line compensation device, associates the pole number recognition result with the component recognition result, and establishes a normalized reference coordinate system;

[0009] The operating data of the overhead contact network compensation device is obtained based on the normalized reference coordinate system, and the operating data of the overhead contact network compensation device is input into the preset overhead contact network compensation device displacement trend prediction model for calculation to obtain the displacement trend prediction result of the overhead contact network compensation device, and the displacement trend prediction result of the overhead contact network compensation device is verified in combination with the preset threshold judgment logic, and the risk analysis result of the overhead contact network compensation device is output.

[0010] Optionally, the image data of the contact network compensation device is preprocessed and calibrated, which includes: motion detection and redundant data removal, image screening, image enhancement and denoising.

[0011] Optionally, the motion detection and redundant data removal are specifically: performing motion detection analysis on the image data of the contact network compensation device, and detecting the displacement changes in the image data of the contact network compensation device through an inter-frame difference algorithm, thereby completing the motion detection and redundant data removal of the image data of the contact network compensation device.

[0012] Optionally, the image screening is specifically:

[0013] Converting the image data of the overhead line compensation device into a grayscale image;

[0014] Applying a Laplace filter to the grayscale image, calculating the second-order gradient change of the catenary compensation device image data, and generating a Laplace transformed image;

[0015] Calculate the variance of the image after Laplace transformation and determine whether the Laplace variance is less than the set variance threshold;

[0016] If yes, it is determined that the image data of the overhead line compensation device is in a fuzzy state;

[0017] If not, it is determined that the image data of the overhead line compensation device is in a non-fuzzy state;

[0018] Output the fuzzy detection result of the overhead line compensation device image data.

[0019] Optionally, the image enhancement and denoising are specifically:

[0020] Performing three-layer wavelet decomposition on the image data of the contact network compensation device, decomposing the image data of the contact network compensation device into a first frequency component and a second frequency component, wherein the frequency of the second frequency component is greater than the frequency of the first frequency component;

[0021] The second frequency component is subjected to hard threshold processing, and the second frequency component after the hard threshold processing and the first frequency component are subjected to inverse wavelet transform, so as to reconstruct and output the denoised image data of the contact network compensation device.

[0022] Optionally, the target detection model is constructed based on the deep learning algorithm, which specifically includes: constructing a component target detection model based on the Yolov11 algorithm, and constructing a pole number target detection model based on the DBNET algorithm and the SVTR algorithm. The component target detection model and the pole number target detection model are both trained through historical image data of contact network compensation device components with labeled information.

[0023] Optionally, the rod number recognition result is associated with the component recognition result, and a normalized reference coordinate system is established, which is specifically:

[0024] Establish a spatial coordinate system, set the specific area of ​​the oblique arm insulator component in the overhead contact network compensation device as the origin, set the short side of the weight component in the overhead contact network compensation device as the X-axis direction, set the long side of the weight component as the Y-axis direction, and use the projection of the oblique arm insulator component on the Y-axis as the measurement unit to calculate the position coordinates of the weight component relative to the normalized reference coordinate system;

[0025] The solved position coordinates are transformed and scaled, and the position coordinates of the weight component in each image are normalized to form a normalized reference coordinate system. The historical detection data of the weight component is continuously acquired based on the normalized reference coordinate system, and the historical data sequence of the weight component is formed by combining the corresponding time series data and environmental data.

[0026] Optionally, the preset contact network compensation device displacement trend prediction model, which is specifically a LightGBM model, is trained through a historical data sequence of the weight component.

[0027] Optionally, the displacement trend prediction model of the contact network compensation device specifically takes maximizing the predicted stability score as the optimization goal, and its calculation formula is:

[0028]

[0029] Among them, PSS represents the prediction stability score, Indicates the degree of consistency between the predicted trend and the actual trend at time t. Indicates the stability of displacement prediction at time t, It represents the ability of the model to capture dynamic changes at time t. represents the fluctuation range of the forecast results at time t, is the impact weight of historical data, represents the cumulative error of historical forecasts, and T represents the length of the time series.

[0030] The technical solution of the present invention has at least the following advantages and beneficial effects:

[0031] The present invention not only realizes the automatic identification and precise positioning of the contact network compensation device through the deep learning algorithm, but also establishes an operation data collection system based on the normalized reference coordinate system, realizes the intelligent monitoring and analysis of the operation status of the compensation device, and effectively improves the detection efficiency and accuracy of the contact network compensation device by deploying the contact network safety inspection device and combining the target detection model and the displacement trend prediction model, realizes all-weather, high-precision automatic monitoring, provides reliable data support for the preventive maintenance of the contact network system, and significantly improves the intelligence level of contact network operation and maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A schematic flow chart of a method for detecting and predicting a stuck catenary compensation device provided by the present invention;

[0033] Figure 2 A schematic structural diagram of the contact network compensation device provided by the present invention. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be described clearly and completely below in conjunction with the drawings in the present invention. Obviously, what is described is only a part of the present invention, not all of it. Generally, the components of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0035] like Figure 1 and Figure 2 As shown, the present invention provides one embodiment: a method for detecting and predicting stuckness of a contact network compensation device, the method comprising the steps of:

[0036] Collect image data of the overhead line compensation device based on the overhead line safety inspection device;

[0037] Preprocess and calibrate the image data of the overhead line compensation device;

[0038] Building a target detection model based on a deep learning algorithm, the target detection model performs component recognition of the overhead line compensation device and pole number recognition of the overhead line compensation device, associates the pole number recognition result with the component recognition result, and establishes a normalized reference coordinate system;

[0039] The operating data of the overhead contact network compensation device is obtained based on the normalized reference coordinate system, and the operating data of the overhead contact network compensation device is input into the preset overhead contact network compensation device displacement trend prediction model for calculation to obtain the displacement trend prediction result of the overhead contact network compensation device, and the displacement trend prediction result of the overhead contact network compensation device is verified in combination with the preset threshold judgment logic, and the risk analysis result of the overhead contact network compensation device is output.

[0040] In this embodiment, the existing overhead line safety inspection device, hereinafter referred to as 2C device, is used to collect images of the overhead line compensation device at a certain period, and the 2C device is ensured to be stable during the collection process to obtain clear images. At the same time, the collection time is recorded to provide data support for subsequent analysis.

[0041] Specifically, the image data of the overhead line compensation device is preprocessed and calibrated, which is in sequence: motion detection and redundant data removal, image screening, image enhancement and denoising.

[0042] The motion detection and redundant data removal are specifically as follows: for the collected original video data, firstly, motion detection analysis is performed, and the moving objects in the video are detected by using an inter-frame difference algorithm. By analyzing the displacement changes between video frames, the redundant data generated by parking, low-speed driving, etc. are effectively removed. This processing step reduces the number of video frames, reduces the amount of data calculation, and further ensures the real-time and effectiveness of subsequent image processing.

[0043] The image screening is specifically as follows:

[0044] Convert the image data of the overhead line compensation device into a grayscale image to reduce the computational complexity;

[0045] Applying a Laplace filter to the grayscale image, calculating the second-order gradient change of the catenary compensation device image data, and generating a Laplace transformed image;

[0046] Calculate the variance of the image after Laplace transformation and determine whether the Laplace variance is less than the set variance threshold;

[0047] If yes, it is determined that the image data of the overhead line compensation device is in a fuzzy state;

[0048] If not, it is determined that the image data of the overhead line compensation device is in a non-fuzzy state;

[0049] Output the fuzzy detection results of the overhead line compensation device image data for image screening.

[0050] The image enhancement and denoising are specifically as follows:

[0051] Decompose the image into low-frequency and high-frequency components;

[0052] The high-frequency components are processed by hard thresholding to set the high-frequency coefficients less than the threshold to zero, and the high-frequency components are processed by hard thresholding. The high-frequency components processed by thresholding and the original low-frequency components are processed by inverse wavelet transform to reconstruct the denoised image.

[0053] The number of wavelet decomposition layers and threshold size parameters are selected according to image characteristics and noise; the denoised image is output to enhance the clarity and visibility of key components.

[0054] In the specific application of this embodiment, the target detection model is constructed based on the deep learning algorithm, which specifically includes: constructing a component target detection model based on the Yolov11 algorithm, and constructing a pole number target detection model based on the DBNET algorithm and the SVTR algorithm. The component target detection model and the pole number target detection model are both trained through historical image data of contact network compensation device components with labeled information.

[0055] During the implementation, regarding the component target detection model: a target detection model based on Yolov11 deep learning is constructed, and a large number of labeled images of overhead line compensation device components are used for training. The labeled information includes the category, position and shape characteristics of components such as weights and insulators. The preprocessed image is input into the trained model, and the model outputs the detection results of each component in the image, including accurate bounding box coordinates, component category confidence and other information. Regarding the pole number target detection model: a deep learning character detection and recognition model based on DBNET and SVTR is constructed to complete the detection of the pole number plate position and pole number plate content to obtain the overhead line support positioning information.

[0056] In specific implementation, this embodiment determines the precise geographic location corresponding to the image based on the pole number plate information in the image, combined with the pre-established railway contact network pole number and geographic location coordinate basic information. At the same time, the collected geographic coordinate information is used for double verification and calibration to ensure the accuracy of geographic location positioning; the positioning information is associated with the detected component information to establish a geographic location-component information mapping relationship, providing a spatial dimension reference for subsequent data analysis.

[0057] In a further implementation of this embodiment, the rod number recognition result is associated with the component recognition result, and a normalized reference coordinate system is established, which is specifically:

[0058] Establish a spatial coordinate system, set the specific area of ​​the oblique arm insulator component in the overhead contact network compensation device as the origin, set the short side of the weight component in the overhead contact network compensation device as the X-axis direction, set the long side of the weight component as the Y-axis direction, and use the projection of the oblique arm insulator component on the Y-axis as the measurement unit to calculate the position coordinates of the weight component relative to the normalized reference coordinate system;

[0059] The solved position coordinates are transformed and scaled, and the position coordinates of the weight component in each image are normalized to form a normalized reference coordinate system. The historical detection data of the weight component is continuously acquired based on the normalized reference coordinate system, and the historical data sequence of the weight component is formed by combining the corresponding time series data and environmental data.

[0060] In implementation, this embodiment establishes a normalized reference coordinate system based on the size information of the detected insulator and pole number plate, takes the lower left corner of the oblique wrist arm insulator as the origin, takes the short side of the weight as the X-axis direction, the long side as the coordinate axis Y-axis direction, and takes the oblique wrist arm projection on the Y-axis as the unit of measurement to calculate the position coordinates of the weight relative to the coordinate system; through operations such as coordinate transformation and scaling, the position of the weight in different images is normalized to obtain normalized relative position information; for the detected compensation device image, regional weather information (including temperature, wind direction, wind force) is obtained according to its image acquisition time and geographical location; the relative position information and the corresponding contact network pillar ID, acquisition time, weather information and other metadata are recorded and stored in the database to form a historical data sequence; when the next image acquisition and analysis is performed at the same geographical location, the new normalized relative position information is obtained. It is compared with the information of the corresponding position in the historical data sequence, and the position change amount and change trend are calculated and recorded.

[0061] In this embodiment, the preset displacement trend prediction model of the contact network compensation device, which is specifically a LightGBM model, is trained through the historical data sequence of the weight component.

[0062] Specifically, this embodiment uses historical data for analysis, and for multiple historical detection data of each weight, calculates the speed and acceleration indicators of displacement changing over time, and analyzes the correlation between displacement and factors such as ambient temperature, wind direction, and wind force, and determines the displacement trend range and change law under normal operation. Based on the results of historical data analysis, the LightGBM model is used to build a prediction model to process data with time series characteristics. The key features (displacement, temperature, wind direction, wind force, etc.) in the historical data are used as input, and the future displacement trend is used as output, and the model is trained and optimized using historical data. In model training, the accuracy and generalization ability of the prediction model are improved by tuning the key parameters of LightGBM (learning rate, number of leaf nodes, number of iterations, etc.). In addition, the best parameter combination is selected through cross-validation to reduce the impact of data deviation on the model. In the real-time monitoring stage, the latest 2C detection results are input into the prediction model, and the model outputs the prediction results of the future displacement trend of the weight. According to the pre-set warning threshold (such as the displacement speed exceeds a certain value or the displacement trend changes abnormally), it is judged whether there is a risk of stagnation. If there is a risk, early warning information will be issued in a timely manner to inform the location of the weight, the expected stuck time, the possible cause of the failure (such as judging whether it is a grease problem or a component wear problem based on data characteristics) and preliminary maintenance suggestions (such as adding grease, checking component wear, etc.). At the same time, new test data will be fed back to the model, and the model will be updated and optimized in real time to continuously improve the reliability of the prediction.

[0063] The displacement trend prediction model of the contact network compensation device specifically takes maximizing the predicted stability score as the optimization goal, and its calculation formula is:

[0064]

[0065] Among them, PSS represents the prediction stability score, Indicates the degree of consistency between the predicted trend and the actual trend at time t. Indicates the stability of displacement prediction at time t, It represents the ability of the model to capture dynamic changes at time t. represents the fluctuation range of the forecast results at time t, is the impact weight of historical data, represents the cumulative error of historical forecasts, and T represents the length of the time series.

[0066] in, , is the change in the predicted value, is the change in the actual value, n is the window size, , is the predicted value at time k, is the sampling time interval, , is the predicted value, is the actual value, is the actual average value in the window, is the time weight, m ​​is the size of the dynamic monitoring window, , is the mean of the predicted values ​​within the window, , β is the decay rate parameter, i is the historical moment, t is the current moment, .

[0067] Furthermore, the preset threshold judgment logic also includes: this embodiment extracts information such as pillar number, weight height, weight width, relative key position displacement, collection time, ambient temperature, wind direction, wind force, etc. from the 2C detection results, cleans these data, removes abnormal values ​​(such as obviously wrong displacement data) and noise data (such as wrong data caused by sensor failure or interference); analyzes the change in ambient temperature during collection. If the relative position change is less than the preset threshold, and the ambient temperature fluctuates within a certain range (indicating that the position change is not caused by normal thermal expansion and contraction due to temperature change), it is preliminarily judged that the compensation device has a risk of jamming; further based on the current geographical location information, request the contact network height information at the location in the 6C data center for comprehensive judgment to improve the accuracy of the jam judgment; if the comprehensive judgment result shows that the compensation device is jammed, a jam alarm message is generated, including the jam location, possible jam degree, related image evidence, etc.

[0068] In the application example, the specific process of this embodiment is divided into (1) collecting image data through 2C multiplication; (2) detecting the pole number based on the image data to establish one-pole-one-file positioning information; (3) detecting the position of the weight, and performing relative position calculation and activity offset calculation; (4) completing the jam analysis based on the previous N times of archive information on the same pole; (5) completing the early warning analysis based on the previous N times of archive information on the same pole.

[0069] The above is only the preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for detecting and predicting stuckness of a contact network compensation device, characterized in that: The steps of the method include: Collect image data of the overhead line compensation device based on the overhead line safety inspection device; Preprocess and calibrate the image data of the overhead line compensation device; Building a target detection model based on a deep learning algorithm, the target detection model performs component recognition of the overhead line compensation device and pole number recognition of the overhead line compensation device, associates the pole number recognition result with the component recognition result, and establishes a normalized reference coordinate system; Acquire the operation data of the catenary compensation device based on the normalized reference coordinate system, input the operation data of the catenary compensation device into a preset catenary compensation device displacement trend prediction model for calculation, obtain the catenary compensation device displacement trend prediction result, verify the catenary compensation device displacement trend prediction result in combination with a preset threshold judgment logic, and output the catenary compensation device jam risk analysis result; The target detection model is constructed based on the deep learning algorithm, which specifically includes: constructing a component target detection model based on the Yolov11 algorithm, and constructing a pole number target detection model based on the DBNET algorithm and the SVTR algorithm, wherein the component target detection model and the pole number target detection model are trained by using the historical image data of the contact network compensation device components with labeled information; The rod number recognition result is associated with the component recognition result, and a normalized reference coordinate system is established, which is specifically: Establish a spatial coordinate system, set the specific area of ​​the oblique arm insulator component in the overhead contact network compensation device as the origin, set the short side of the weight component in the overhead contact network compensation device as the X-axis direction, set the long side of the weight component as the Y-axis direction, and use the projection of the oblique arm insulator component on the Y-axis as the measurement unit to calculate the position coordinates of the weight component relative to the normalized reference coordinate system; The solved position coordinates are transformed and scaled, and the position coordinates of the weight component in each image are normalized to form a normalized reference coordinate system. The historical detection data of the weight component is continuously acquired based on the normalized reference coordinate system, and the historical data sequence of the weight component is formed by combining the corresponding time series data and environmental data.

2. The method for detecting and predicting stuckness of a contact network compensation device according to claim 1, characterized in that: The image data of the contact network compensation device is preprocessed and calibrated, which is sequentially: motion detection and redundant data removal, image screening, image enhancement and denoising.

3. The method for detecting and predicting the stuck state of the contact network compensation device according to claim 2, characterized in that: The motion detection and redundant data removal are specifically as follows: performing motion detection and analysis on the image data of the contact network compensation device, and detecting the displacement changes in the image data of the contact network compensation device through an inter-frame difference algorithm, thereby completing the motion detection and redundant data removal of the image data of the contact network compensation device.

4. The method for detecting and predicting stuckness of a contact network compensation device according to claim 3, characterized in that: The image screening is specifically as follows: Converting the image data of the overhead line compensation device into a grayscale image; Applying a Laplace filter to the grayscale image, calculating the second-order gradient change of the catenary compensation device image data, and generating a Laplace transformed image; Calculate the variance of the image after Laplace transformation and determine whether the Laplace variance is less than the set variance threshold; If yes, it is determined that the image data of the overhead line compensation device is in a fuzzy state; If not, it is determined that the image data of the overhead line compensation device is in a non-fuzzy state; Output the fuzzy detection result of the overhead line compensation device image data.

5. The method for detecting and predicting stuckness of a contact network compensation device according to claim 4, characterized in that: The image enhancement and denoising are specifically as follows: Performing three-layer wavelet decomposition on the image data of the contact network compensation device, decomposing the image data of the contact network compensation device into a first frequency component and a second frequency component, wherein the frequency of the second frequency component is greater than the frequency of the first frequency component; The second frequency component is subjected to hard threshold processing, and the second frequency component after the hard threshold processing and the first frequency component are subjected to inverse wavelet transform, so as to reconstruct and output the denoised image data of the contact network compensation device.

6. The method for detecting and predicting stuckness of a contact network compensation device according to claim 5, characterized in that: The preset displacement trend prediction model of the contact network compensation device, which is specifically a LightGBM model, is trained through the historical data sequence of the weight component.

7. The method for detecting and predicting stuckness of a contact network compensation device according to claim 6, characterized in that: The displacement trend prediction model of the contact network compensation device specifically takes maximizing the predicted stability score as the optimization goal, and its calculation formula is: Among them, PSS represents the prediction stability score, Indicates the degree of consistency between the predicted trend and the actual trend at time t. Indicates the stability of displacement prediction at time t, It represents the ability of the model to capture dynamic changes at time t. represents the fluctuation range of the forecast results at time t, is the impact weight of historical data, represents the cumulative error of historical forecasts, and T represents the length of the time series.

Images