Multi-dimensional coupling algorithm-based high-precision measurement method and system for parameters of track overhead line system

Abstract

The invention relates to a rail contact network parameter high-precision measurement method and system based on a multi-dimensional coupling algorithm, and solves the problems that manual measurement precision is insufficient, continuous collection is difficult, operation is hindered in a severe environment, and safety risks exist in workers. Calculating an initial rail width, generating a rail width reference through temperature compensation, and synchronously combining inertial measurement data to establish a space reference coordinate system; re-calibrating the parameters, comparing with the attitude angle, updating the reference if the parameters exceed the threshold value, and otherwise, compensating and filtering the sensing data; then, a measured point of the contact network is tracked, a holder is adjusted for alignment, and finally, a three-dimensional coordinate, a guide height and a pull-out value are calculated and output by fusing multi-source data. The method has the advantages that high-precision and automatic measurement of the geometric parameters of the overhead line system is achieved, the safety risk and the efficiency bottleneck of manual measurement are remarkably reduced, and reliable data support is provided for intelligent operation and maintenance of railways.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of rail transit monitoring, in particular to a rail catenary parameter high-precision measurement method and system based on multi-dimensional coupling algorithm. BACKGROUND

[0002] As the core power supply component of electrified railway, the rail catenary determines the stability of train power supply and the safety of train operation directly through the precision of its geometric parameters such as height and pull-out value. Therefore, these parameters need to be measured frequently and accurately to meet the real-time and reliability requirements of railway operation and maintenance on equipment state monitoring. With the development of electrified railway towards high speed and intelligence, the traditional measurement method has been difficult to adapt to the high-precision detection requirements under complex working conditions, and more efficient multi-dimensional collaborative measurement technology is urgently needed.

[0003] Currently, the measurement of rail catenary parameters mainly relies on manual measurement technology, in which the staff carries laser range finders, total stations and other equipment to operate on site, and records the data by manually aligning the catenary measurement points.

[0004] For the related technologies in the above, the inventors find that there are the following defects: manual alignment is easily affected by visual errors and operation experience, leading to measurement deviation beyond the high-precision operation and maintenance requirements, and limited by human efficiency, only sampling detection mode (such as selecting a few points every few hundred meters) can be used, which cannot realize continuous data collection on the whole line, is easy to miss local parameter abnormalities, and bad environments such as rain, snow, backlight and dust will directly hinder manual operation, leading to interruption or sharp decline in detection efficiency, at the same time, the staff working on the track and elevated area also faces the safety risk of train collision, which is difficult to guarantee the safety of operation and the continuity of detection. SUMMARY

[0005] In order to realize high-precision and automatic measurement of catenary geometric parameters, significantly reduce the safety risk and efficiency bottleneck of manual measurement, and provide reliable data support for intelligent operation and maintenance of railway, the present application provides a rail catenary parameter high-precision measurement method and system based on multi-dimensional coupling algorithm.

[0006] In the first aspect, the present application provides a rail catenary parameter high-precision measurement method based on multi-dimensional coupling algorithm, which adopts the following technical solution:

[0007] A rail catenary parameter high-precision measurement method based on multi-dimensional coupling algorithm is executed by a measuring device, which includes a machine vision module, an inertial measurement unit, a laser ranging module and a pan-tilt, and the method includes the following steps:

[0008] The rail image is collected by the machine vision module, the rail head contour feature points are recognized and quality grading is performed; the initial rail width value is calculated based on the high-quality feature point set fitting contour, the rail width reference parameter is generated by combining the rail thermal expansion coefficient and the real-time rail temperature data compensation; the initial attitude data of the measuring device is collected by synchronously starting the inertial measurement unit, and the space reference coordinate system is established with the rail width reference as the scale reference and the initial attitude as the azimuth reference;

[0009] The gimbal is controlled to rotate according to the preset angle sequence, the laser ranging module is linked to collect the distance data corresponding to the angle, the angle-distance mapping table is generated by the space-time correlation algorithm, and the calibration parameter table is formed which is bound to the rail width reference and calibrated in the space reference coordinate system;

[0010] In the space reference coordinate system, the attitude angle of the measuring device relative to the track is calculated, and the difference value is compared with the absolute attitude angle measured by the inertial measurement unit; if the difference value exceeds the preset threshold, the rail width reference is refitted based on the rail width reference and the high-quality feature point historical data, and the space reference coordinate system and the calibration parameter table are updated; otherwise, the environmental temperature and vibration data are collected, the preset temperature compensation coefficient and vibration denoising algorithm are called to compensate and filter the output data of the laser ranging module and the inertial measurement unit online, and the compensated laser data and the filtered inertial data are obtained;

[0011] The measured points of the catenary are recognized and tracked by the machine vision module, the trajectory is calculated by combining the motion prediction algorithm and the space reference coordinate system, the gimbal is adjusted to make the laser measurement beam continuously aim at the measured points, and the visual tracking data is output;

[0012] The three-dimensional coordinates, the gradient and the pull-out value parameters of the measured points of the catenary are output by fusing the compensated laser data, the filtered inertial data and the visual tracking data, and calculating by the multi-source fusion algorithm.

[0013] By adopting the above technical scheme, the machine vision, inertial measurement, laser ranging and gimbal multi-module are cooperated, the environmental interference is reduced by combining the temperature and vibration compensation and attitude calibration; the space reference and the calibration parameter are dynamically updated to ensure the attitude accuracy, and the laser continuously aims at the measured points of the catenary; finally, the three-dimensional coordinates, the gradient and the pull-out value of the measured points are accurately output by multi-source data fusion calculation, the high-precision measurement of the track catenary parameters is effectively realized, and the data reliability and the measurement stability are improved.

[0014] In the second aspect, the application provides a multi-dimensional coupling algorithm track catenary parameter high-precision measurement system, which adopts the following technical scheme:

[0015] The application discloses a multi-dimension coupling algorithm track contact network parameter high-precision measurement system, which comprises a memory, a processor and a program stored in the memory and capable of running on the processor. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a flowchart of a multi-dimension coupling algorithm track contact network parameter high-precision measurement method.

[0017] Figure 2 is a flowchart of another embodiment of the application, which collects a steel rail image through a machine vision module, identifies rail head contour feature points and performs quality grading. DETAILED DESCRIPTION

[0018] The application will be further described in detail below with reference to the drawings.

[0019] REFERENCE Figure 1 A multi-dimension coupling algorithm track contact network parameter high-precision measurement method disclosed by the application is executed by a measuring device, the device comprises a machine vision module, an inertial measurement unit, a laser ranging module and a pan-tilt, and the method comprises the following steps.

[0020] In step S100, a steel rail image is collected through the machine vision module, rail head contour feature points are identified and quality grading is performed; an initial rail width value is calculated based on a high-quality feature point set fitting contour, a rail width reference parameter is generated by combining a steel rail thermal expansion coefficient and real-time rail temperature data compensation; an initial attitude data of the measuring device is collected by synchronously starting the inertial measurement unit, and a space reference coordinate system taking the rail width reference as a scale reference and taking the initial attitude as an azimuth reference is established.

[0021] The machine vision module is a device for acquiring and analyzing image information by using optical equipment and image processing technology, and is used for identifying and measuring the shape, size and other features of an object. The rail head contour feature point is a geometric feature point of a steel rail head edge, and is used for determining the shape and size of the steel rail. The quality grading is classification of the collected feature points according to indexes such as definition and integrity, and distinguishes high-quality and low-quality feature points. The rail width reference parameter is a parameter corrected by taking the actual width of the steel rail as a reference and combining environmental factors (such as temperature), and is used for reference in subsequent measurement. The inertial measurement unit (IMU) is a sensor for measuring three-axis attitude angles and acceleration of an object, and is used for determining the space attitude of the measuring device. The space reference coordinate system is a coordinate system established by taking the rail width reference as a scale reference and taking the initial attitude as an azimuth reference, and is used for unifying the space position and attitude of the measuring device.

[0022] The necessary processes are described as follows:

[0023] 1. Image Acquisition and Preprocessing: After the machine vision module collects the rail image, it performs grayscale, filtering, and other preprocessing operations to improve image quality and facilitate subsequent feature point extraction.

[0024] 2. Feature Point Extraction and Quality Grading: The edge detection algorithm (such as Canny algorithm) is used to extract the feature points of the rail head profile from the preprocessed image. According to the edge intensity, continuity, and other indicators of the feature points, the pre-set grading rules are used to grade the quality of the feature points, and high-quality feature points are selected for subsequent calculations.

[0025] 3. Rail Width Reference Parameter Generation: Based on the selected high-quality feature point set, the least squares method is used to fit the rail head profile, and the initial rail width value is calculated. Combined with the thermal expansion coefficient of the rail (such as α = 1.2 × 10⁻ 5 / ℃) and real-time rail temperature data (obtained through temperature sensors), the initial rail width value is corrected through the temperature compensation formula (such as ΔL = α × L0 × ΔT, where ΔL is the rail width change, L0 is the initial rail width, and ΔT is the rail temperature change), and the rail width reference parameter is generated.

[0026] 4. Initial Attitude Data Acquisition and Coordinate System Establishment: The inertial measurement unit collects the initial attitude data of the measurement device, including three-axis attitude angles (such as pitch angle, roll angle, and yaw angle) and acceleration information. With the rail width reference parameter as the scale reference and the initial attitude data as the orientation reference, the spatial reference coordinate system is established through the coordinate transformation algorithm (such as Euler angle transformation), providing a unified reference framework for subsequent measurements.

[0027] Step S200: Control the gimbal to rotate according to the pre-set angle sequence, and link the laser ranging module to collect the corresponding angle-distance data. Through the space-time correlation algorithm, the angle-distance mapping table is generated, forming the calibration parameter table that is bound to the rail width reference and calibrated in the spatial reference coordinate system.

[0028] Wherein, gimbal: a mechanical device that can accurately control the rotation angle, usually equipped with a stepper motor and a motor driver, capable of rotating according to the pre-set angle sequence. Laser ranging module: a device that measures distance using laser technology, calculates distance by emitting laser pulses and measuring the round-trip time of reflected light. Space-time correlation algorithm: an algorithm that correlates and calibrates data collected at different time points with corresponding spatial positions, ensuring data accuracy and consistency. Angle-distance mapping table: a table that records the distance data measured by the laser ranging module at different rotation angles, used for subsequent spatial positioning and calibration. Calibration parameter table: a table that records the calibrated parameters, which are bound to the rail width reference and calibrated in the spatial reference coordinate system, used for subsequent high-precision measurements.

[0029] The necessary process is as follows:

[0030] In step S200, first control the controllable gimbal to rotate according to the preset angle sequence (such as rotating every 10 degrees from 0 degrees to 360 degrees). At each preset angle, start the laser ranging module to collect distance data of the target point (such as a specific point on the catenary). The laser ranging module calculates the distance by emitting a laser pulse and measuring the round-trip time of the reflected light, ensuring that accurate distance data can be obtained at each angle position. For example, at 0 degrees, the laser ranging module measures a distance of 1000 mm; at 10 degrees, the measured distance is 1005 mm. These data are recorded in real time to provide a basis for subsequent calibration.

[0031] Subsequently, the space-time correlation algorithm is applied to correlate and calibrate the distance data at different angles. This algorithm considers both the time (measurement time) and space (gimbal angle) dimensions, and through a specific space-time correlation function f(θ,d,t), it converts the distance data at each angle into calibration parameters. Specifically, assuming that the distance measured at angle θ is d, and the measurement time is t, the algorithm will calculate the calibrated distance value based on these parameters. For example, at an angle of 0 degrees, the measured distance is 1000 mm, and the measurement time is t0, then the calibration parameter is f(0,1000,t0). At an angle of 10 degrees, the measured distance is 1005 mm, and the measurement time is t1, then the calibration parameter is f(10,1005,t1). These calibration parameters are then recorded to form a calibration parameter table.

[0032] Finally, the generated calibration parameter table is bound with the track width reference parameters and calibrated in the spatial reference coordinate system. Through coordinate transformation algorithms (such as Euler angle transformation), the data in the calibration parameter table is converted to the spatial reference coordinate system, ensuring consistency with the track width reference.

[0033] Step S300, in the spatial reference coordinate system, the difference value between the attitude angle of the measuring device relative to the track and the absolute attitude angle measured by the inertial measurement unit is calculated; if the difference value exceeds the preset threshold, the track width reference is refitted based on the track width reference and the historical data of high-quality feature points, and the spatial reference coordinate system and the calibration parameter table are updated; otherwise, the environmental temperature and vibration data are collected, and the preset temperature compensation coefficient and vibration denoising algorithm are called to compensate and filter the output data of the laser ranging module and the inertial measurement unit online, to obtain compensated laser data and filtered inertial data.

[0034] where the attitude angles: describe the orientation of an object in space, typically including pitch, roll, and yaw. These angles represent the rotation of the object around three mutually perpendicular axes, respectively. High-quality feature point history data: the data of high-quality rail head profile feature points identified and recorded in step S100, used for subsequent rail width reference fitting and calibration. Pre-set temperature compensation coefficient: a coefficient for correcting the measuring device according to temperature changes, used to compensate for the influence of temperature changes on the measurement results. Vibration noise reduction algorithm: an algorithm for reducing the interference of vibration on measurement data, improving measurement accuracy.

[0035] The process is described as follows:

[0036] 1. In step S300, first calculate the attitude angle of the measuring device relative to the track in the spatial reference coordinate system. This process uses the rail head profile feature points identified by the machine vision module to calculate the relative attitude angle of the measuring device relative to the track. The specific calculation formula is as follows:

[0037] Pitch: Calculate the angle between the rail head profile center line and the horizontal line. The specific formula is as follows: .

[0038] Roll: Calculate the angle between the rail head profile edge line and the vertical line. The specific formula is as follows: .

[0039] where, represents the offset of the rail head profile center line in the horizontal direction, i.e. the distance between the rail head profile center line and the horizontal line in the horizontal direction; represents the offset of the rail head profile center line in the vertical direction, i.e. the distance between the rail head profile center line and the horizontal line in the vertical direction, represents the offset of the rail head profile edge line in the vertical direction, i.e. the distance between the rail head profile edge line and the vertical line in the vertical direction.

[0040] Yaw: Assuming that the motion direction of the measuring device is consistent with the direction of the track, the yaw angle can be approximately 0, or measured by other auxiliary sensors (such as a magnetometer). The specific formula is as follows: .

[0041] 2. Collect the absolute attitude angle of the inertial measurement unit (IMU):

[0042] IMU initialization: Start the pre-set inertial measurement unit (IMU) and perform initialization calibration to ensure the accuracy of its measurement data.

[0043] Data collection: IMU determines the absolute attitude angle of the measuring device in space by measuring acceleration and angular velocity. The specific steps are as follows: the absolute attitude angle output by the IMU includes pitch angle, roll angle and yaw angle. These data can be collected by the sensors of the IMU and transmitted to the processing unit.

[0044] For example, the data format output by the IMU may be as follows:

[0045] Absolute pitch angle = ; Absolute roll angle = ; Absolute yaw angle = .

[0046] 3. Calculate the attitude angle difference value:

[0047] Difference value calculation: compare the attitude angle of the measuring device relative to the track with the absolute attitude angle collected by the IMU, and calculate the attitude angle difference value. The specific steps are as follows:

[0048] Pitch angle difference value: .

[0049] Roll angle difference value: .

[0050] Yaw angle difference value: .

[0051] If any of these difference values exceeds the preset threshold value, it indicates that the attitude of the measuring device may have changed significantly and needs to be recalibrated. At this time, based on the track width reference and the high-quality feature points recorded in the historical data, the track width reference is refitted, and the spatial reference coordinate system and the calibration parameter table are updated to ensure the accuracy of the measurement. The specific steps are as follows:

[0052] Screening high-quality feature points: screen feature points with a confidence level higher than 90% and repeated in the last 10 measurements from historical data to ensure the stability and reliability of the data. Refitting the track head contour: using these high-quality feature points, refitting the track head contour to calculate new track width reference parameters.

[0053] Updating the spatial reference coordinate system: according to the new track width reference parameters, re-establishing the spatial reference coordinate system to ensure the accuracy and consistency of the coordinate system.

[0054] Updating the calibration parameter table: according to the new spatial reference coordinate system, recalculating the parameters in the calibration parameter table to ensure the accuracy and reliability of the calibration parameters.

[0055] On the contrary, if the difference value is within the preset threshold range, it indicates that the posture of the measuring device is relatively stable and does not need to be recalibrated. At this time, the system will collect environmental temperature and vibration data, which will affect the accuracy of the measurement. In order to compensate for the influence of these environmental factors, the system calls the preset temperature compensation coefficient and vibration noise reduction algorithm. The specific steps are as follows:

[0056] 1. Collect environmental temperature and vibration data: Use temperature sensors and vibration sensors to collect real-time environmental temperature and vibration data.

[0057] 2. Temperature compensation: Compensate for the output data of the laser ranging module and the inertial measurement unit according to the temperature change. The specific formula is as follows: ; Where, is the compensated distance, is the original measured distance, is the temperature compensation coefficient, is the temperature change.

[0058] 3. Vibration noise reduction: Use vibration noise reduction algorithms to filter the output data of the inertial measurement unit. For example, use a Kalman filter.

[0059] Step S400, identify and track the measured point of the catenary through the machine vision module, calculate the trajectory in combination with the motion prediction algorithm and the spatial reference coordinate system, adjust the gimbal to make the laser measurement beam continuously aim at the measured point, and output the visual tracking data.

[0060] Where, motion prediction algorithm: an algorithm used to predict the future position of the measured point, helping to adjust the gimbal posture. Spatial reference coordinate system: refer to the disclosure of step S100, used to unify the coordinate system of the spatial position and posture of the measuring device. Laser measurement beam: a laser beam used to measure distance, which needs to continuously aim at the measured point. Visual tracking data: the position data of the measured point output by the machine vision module, used for subsequent processing.

[0061] The specific process can refer to steps S410 to S450.

[0062] Step S500, fuse the compensated laser data, filtered inertial data and visual tracking data, and calculate the three-dimensional coordinates, height and pull-out value parameters of the measured point of the catenary through the multi-source fusion algorithm.

[0063] Multi-source fusion algorithm: an algorithm used to integrate and process data from multiple sensors to improve the accuracy and reliability of measurement results.

[0064] The above specific process can refer to steps S510 to S540.

[0065] Refer to Figure 2, the rail image is collected by the machine vision module, the rail head contour feature points are recognized, and quality grading is performed, including:

[0066] In step S110, the light conditions are adaptively adjusted by a multi-light source dynamic light supplementing device in the machine vision module, a preset denoising algorithm and a distortion correction matrix are used to pre-process the rail image, and a high-quality rail head region image is obtained.

[0067] The multi-light source dynamic light supplementing device is a hardware system composed of multiple groups of adjustable light LED arrays (such as 2 groups of side light sources on both sides of the rail head and 1 group of top light sources) and a light sensor, which can adjust the brightness and angle of each light source in real time according to the environmental light intensity and the rail surface reflection characteristics. The denoising algorithm is a digital image processing method for eliminating random noise (such as sensor thermal noise and environmental interference noise) in the image, which commonly includes Gaussian filtering, median filtering, non-local mean denoising, etc. The distortion correction matrix is a mathematical matrix describing the optical distortion (radial distortion, tangential distortion) of the camera lens, which is obtained by camera calibration and used to restore the distorted image to the real geometric shape. The high-quality rail head region image refers to an image with clear rail head contour edges, moderate contrast, no obvious noise and geometric distortion, and complete coverage of the rail head region (the proportion of the rail head region is not less than 60% of the image).

[0068] The specific process is described as follows:

[0069] 1. Multi-light source dynamic light supplementing adjustment:

[0070] When the device starts, the light sensor (sampling frequency 10 Hz) collects the environmental light intensity (such as 500 lux-10000 lux) and the average brightness value of the rail image (real-time feedback through the machine vision module) in real time.

[0071] Based on the preset brightness threshold (such as target average brightness 80-120 gray value), the control system adjusts the power of each LED light source through PWM (pulse width modulation) technology:

[0072] If the environmental light is too strong (such as direct sunlight at noon, the image is overexposed), reduce the power of the top light source (such as from 100% to 30%), adjust the angle of the side light source (tilt 45°→60°) to reduce the specular reflection;

[0073] If the environmental light is too weak (such as in a tunnel, the image is underexposed), increase the power of the top light source (such as from 30% to 80%), and simultaneously turn on the side light source to ensure that the gray difference between the rail head edge and the background is ≥30.

[0074] 2. Pre-set denoising algorithm processing:

[0075] The collected original image (RGB format, 1920×1080 pixels) is first processed by grayscale processing (retaining brightness information, simplifying calculation).

[0076] For different noise type combination denoising algorithm:

[0077] Gaussian filter (convolution kernel 5x5, σ=1.2) is used to eliminate the Gaussian noise introduced by the sensor (manifested as uniformly distributed small white dots);

[0078] Median filter (window 3x3) is used to eliminate the salt and pepper noise produced by environmental interference (manifested as isolated black and white dots);

[0079] For high noise scenes (such as rainy and snowy weather), superimpose non-local mean denoising (similar window 7x7, weight coefficient 0.1), further smooth the noise by matching the pixel values of similar areas in the image, while preserving the edge details of the rail head.

[0080] 3. Application of distortion correction matrix:

[0081] The distortion correction matrix is obtained in advance by Zhang Zhengyou calibration method: use a checkerboard calibration board (6x9 squares, side length 20mm), take 20 calibration images at different angles within the working distance range of the measuring device (such as 500mm~800mm), calculate the camera intrinsic parameters (focal length f_x=1200pixels, f_y=1200pixels, principal point coordinates u0=960, v0=540) and distortion coefficients (radial distortion k1=-0.03, k2=0.01, tangential distortion p1=0.002, p2=-0.001) by least squares method, and form the distortion correction matrix.

[0082] For the denoised image, apply the correction matrix through the undistort function of OpenCV to eliminate the "bending" of the rail head edge caused by lens distortion (such as the edge straight line becomes an arc), and restore the real geometric shape of the rail head.

[0083] 4. After the pre-processing of the rail head region image, a simple threshold segmentation (based on the gray difference between the rail head and the ballast, such as rail head gray value 60~150, ballast <40) is used to preliminarily crop the image, retain the ROI (region of interest) containing only the rail head, and reduce the calculation amount of subsequent model inference (ROI region size is compressed to 800x400 pixels).

[0084] The final output high-quality rail head region image needs to meet the following requirements: contour edge gradient value ≥20 (clear and distinguishable), noise pixel ratio <0.5%, geometric distortion error <0.1 pixel.

[0085] Step S120, real-time inference of the pre-processed image through the pre-deployed improved U-Net deep learning model, identify the key feature points of the rail head contour through the attention mechanism and depth separable convolution structure, and output the confidence scores of each feature point simultaneously.

[0086] The improved U-Net deep learning model is an improvement upon the classic U-Net architecture, enhanced by introducing an attention mechanism and a depthwise separable convolutional structure to improve its ability to recognize feature points on the railhead contour. The attention mechanism is a technique used to enhance the model's focus on key regions, improving the accuracy and efficiency of feature extraction. The depthwise separable convolutional structure is an optimized convolutional structure that separates standard convolution operations into depthwise convolution and pointwise convolution, reducing computational cost and the number of parameters, thus improving model efficiency. Key feature points are points of significant importance on the railhead contour, such as edge points and inflection points, which are crucial for subsequent calculation of rail width reference parameters.

[0087] The acquisition method and process are described below:

[0088] 1. Data Preparation: Collect a large amount of rail image data, including images under different lighting conditions, angles, and environments. Annotate these images, marking key feature points of the rail head contour, and generate labeled data for training.

[0089] 2. Model Improvement:

[0090] Introducing an attention mechanism: An attention mechanism is introduced into the key layers of the U-Net model to enhance the model's attention to the key regions of the railhead contour.

[0091] By employing channel attention and spatial attention mechanisms, the accuracy and efficiency of feature extraction can be improved.

[0092] The channel attention mechanism extracts channel features through global average pooling and global max pooling, and then generates channel weights through a shared fully connected layer and activation function to enhance the features of important channels.

[0093] The spatial attention mechanism extracts spatial features through global average pooling and global max pooling, and then generates spatial weights through shared convolutional layers and activation functions to enhance the features of important regions.

[0094] Depthwise separable convolutional structure: Using a depthwise separable convolutional structure in the U-Net model optimizes convolution operations and reduces computational cost and the number of parameters.

[0095] Depthwise separable convolution decomposes standard convolution into depthwise convolution and pointwise convolution, performing convolution operations on each input channel separately, and then merging the results through pointwise convolution.

[0096] 3. Model training:

[0097] Data Preprocessing: The collected rail images are preprocessed, including normalization, cropping, flipping, etc., to enhance the generalization ability of the model.

[0098] Training Process: The improved U-Net model is trained using labeled data.

[0099] Select an appropriate loss function, such as Cross-Entropy Loss or Dice Loss, to measure the difference between the model output and the true label.

[0100] Use optimization algorithms such as Adam or SGD to adjust the parameters of the model to minimize the loss function.

[0101] During training, use data augmentation techniques (such as random cropping, rotation, flipping, etc.) to increase the diversity of training data and prevent overfitting.

[0102] Verification and Adjustment: Evaluate the performance of the model on the validation set, and adjust the hyperparameters of the model (such as learning rate, batch size, etc.) according to the validation results to improve the accuracy and generalization ability of the model.

[0103] 4、Real-time Inference:

[0104] Deploy the trained improved U-Net model to actual applications and perform real-time inference on input rail images.

[0105] The model extracts image features through convolution layers, pooling layers and upsampling layers, and generates segmentation results to identify key feature points of the rail head contour.

[0106] 5、Post-processing: Post-process the segmentation results output by the model to extract the precise location of key feature points. Use threshold segmentation and morphological operations (such as dilation and erosion) to remove noise and extract key feature points. For example, use binary threshold segmentation to convert the segmentation result into a binary image, then remove noise through morphological operations to extract key feature points.

[0107] 6、Output Key Feature Points: Output the location information of the identified key feature points.

[0108] Step S130, calculate the local gradient intensity variance of the feature points by the Sobel operator, and calculate the Euclidean distance by comparing with the preset rail head standard contour.

[0109] Sobel operator: a gradient calculation operator used for image edge detection, which can calculate the gradient intensity and direction of each pixel point in the image. Local gradient intensity variance: a statistical measure of the change in gradient intensity around a feature point, used to assess the stability and reliability of the feature point. Euclidean distance: used to measure the straight-line distance between two points in multi-dimensional space, here used to compare the similarity of feature points with the preset railhead standard contour. Pre-set railhead standard contour: a pre-defined standard shape of the railhead contour, used for comparison with the actual detected feature points.

[0110] The process is described as follows:

[0111] 1. Calculate the local gradient intensity variance: use the Sobel operator to calculate the gradient of each feature point and its surrounding area, and get the gradient intensity of each pixel point. Calculate the gradient intensity variance of the surrounding area of the feature point to evaluate the gradient change of the area. The smaller the gradient intensity variance, the more stable the gradient change around the feature point, and the higher the quality of the feature point.

[0112] 2. Calculate the Euclidean distance: compare the position of each feature point with the pre-set railhead standard contour, and calculate the Euclidean distance between them. The smaller the Euclidean distance, the higher the matching degree of the feature point with the standard contour, and the higher the quality of the feature point.

[0113] Step S140, divide the quality level based on the multi-index collaborative judgment rule: divide the feature points with confidence score not less than the preset threshold, gradient intensity variance not more than the preset threshold, and Euclidean distance not more than the preset threshold into high quality level, and the rest into low quality level, to form a high and low quality feature point label set.

[0114] In step S140, the multi-index collaborative judgment rule means to evaluate the quality of feature points by combining multiple key indicators. These indicators include confidence score, gradient intensity variance and Euclidean distance. Each indicator has its specific threshold, and only when a feature point meets all these threshold conditions at the same time, it will be divided into high quality level. This comprehensive evaluation method ensures that the quality evaluation of feature points is comprehensive and reliable, avoiding the misjudgment that may be caused by single indicator.

[0115] In step S140, the system classifies the quality level of the feature points by comprehensively evaluating their confidence scores, gradient intensity variances, and Euclidean distances from the standard profile. First, the system checks whether the confidence score of each feature point reaches a pre-set high confidence threshold, such as 0.9, to ensure that the model's identification of the feature point is highly reliable. Next, the system evaluates whether the gradient intensity variance of the feature point is lower than a pre-set low variance threshold, such as 0.05, to ensure that the gradient changes around the feature point are stable, thereby improving the stability of the feature point. Finally, the system calculates the Euclidean distance between the feature point and the pre-set railhead standard profile, determining whether it is less than a pre-set distance threshold, such as 10 pixels, to ensure that the feature point has a high degree of matching with the standard profile. Only feature points that meet all three conditions are classified as high-quality, and the rest are marked as low-quality, ultimately forming a labeled set of high and low-quality feature points.

[0116] Based on the high-quality feature point set, the profile is fitted to calculate the initial rail width value, and combined with the steel rail thermal expansion coefficient and real-time rail temperature data to compensate for the generation of rail width reference parameters, including:

[0117] Step S1A0, extract the high-quality feature point set from the high and low-quality feature point label set, call the pre-calibrated camera intrinsic matrix and extrinsic matrix, and convert the feature point pixel coordinates to three-dimensional coordinates in the physical coordinate system.

[0118] Wherein, the camera intrinsic matrix: a matrix describing the internal parameters of the camera, including focal length, optical center coordinates, etc., used to convert pixel coordinates to coordinates in the camera coordinate system. The camera extrinsic matrix: a matrix describing the position and attitude of the camera in the world coordinate system, used to convert coordinates in the camera coordinate system to coordinates in the world coordinate system. Physical coordinate system: the actual three-dimensional coordinate system, used to represent the position of the feature point in the actual space.

[0119] The specific steps are as follows:

[0120] 1. Extract the high-quality feature point set: filter out all the feature points marked as high-quality from the high and low-quality feature point label set to form the high-quality feature point set. These feature points have high confidence, stable gradient intensity, and high matching degree with the standard profile in the image, so they are more suitable for subsequent calculations.

[0121] 2. Call the camera intrinsic matrix and extrinsic matrix: use the pre-calibrated camera intrinsic matrix and extrinsic matrix. The camera intrinsic matrix includes focal length, optical center coordinates, etc., used to convert pixel coordinates to coordinates in the camera coordinate system. The extrinsic matrix describes the position and attitude of the camera in the world coordinate system, used to convert coordinates in the camera coordinate system to coordinates in the world coordinate system.

[0122] 3. Coordinate conversion: For each high-quality feature point, convert its pixel coordinates to camera coordinates using the camera intrinsic matrix. Then, convert the camera coordinates to physical coordinates using the extrinsic matrix. This process ensures that the feature points' coordinates are accurately mapped from the image plane to the actual three-dimensional space.

[0123] Step S1B0, based on the converted three-dimensional coordinate data, fit the rail head profile, and calculate the initial rail width value by the spatial position difference of the symmetrical feature points on both sides of the rail head.

[0124] Where, rail head profile: the geometric shape of the rail head, usually inverted trapezoidal or rectangular, used to calculate the rail width. Initial rail width value: the rail width value calculated based on the spatial position difference of the symmetrical feature points on both sides of the rail head, not temperature compensated.

[0125] The specific process is as follows:

[0126] 1. Fit the rail head profile: use the converted three-dimensional coordinate data to fit the profile of the rail head by mathematical fitting methods (such as polynomial fitting or spline fitting). In the fitting process, select the symmetrical feature points on both sides of the rail head to ensure that the fitted profile accurately reflects the actual shape of the rail head.

[0127] 2. Calculate the initial rail width value:

[0128] From the fitted rail head profile, select the symmetrical feature points on both sides of the rail head, and calculate the spatial position difference of these feature points in the horizontal direction. The specific method is as follows:

[0129] For each pair of symmetrical feature points, calculate their distance in the horizontal direction.

[0130] Take the average of multiple pairs of symmetrical feature points to get the initial rail width value.

[0131] Step S1C0, call the preset thermal expansion coefficient of this type of rail, and combine the real-time collected rail head temperature data to correct the initial rail width value.

[0132] Where, thermal expansion coefficient of rail: the coefficient of expansion or contraction of rail material due to temperature change, usually expressed as length change rate per degree Celsius. Real-time rail temperature data: real-time collected rail surface temperature data by temperature sensor. Temperature compensation correction: adjust the initial rail width value according to the thermal expansion coefficient of rail and real-time rail temperature data to eliminate the influence of temperature change on rail width measurement.

[0133] In step S1C0, the preset thermal expansion coefficient of this type of rail is called, and combined with the real-time collected rail head temperature data, the initial rail width value is temperature compensated and corrected. This process ensures the accuracy of rail width measurement and eliminates the influence of environmental temperature change on the measurement result.

[0134] The specific steps are as follows:

[0135] 1. Call the rail thermal expansion coefficient: according to the material and specification of the rail, call the preset rail thermal expansion coefficient.

[0136] 2. Collect real-time rail temperature data: use the temperature sensor installed near the rail head to collect real-time rail head temperature data. Assume that the current collected rail head temperature is T current , and the reference temperature (usually the standard installation temperature) is T ref .

[0137] 3. Calculate the temperature change: calculate the difference between the current temperature and the reference temperature:

[0138] ;

[0139] 4. Perform temperature compensation correction:

[0140] According to the rail thermal expansion coefficient and the temperature change, calculate the compensation amount of the rail width: ;

[0141] Where α is the rail thermal expansion coefficient, is the initial rail width value, is the compensation amount of the rail width.

[0142] Correct the initial rail width value: .

[0143] Step S1D0, use a filtering algorithm to eliminate the measurement fluctuations of the compensated rail width value, and generate a rail width reference parameter.

[0144] The specific steps are as follows:

[0145] 1. Select a filtering algorithm: according to the characteristics and requirements of the measurement data, select a suitable filtering algorithm. For example, for scenes that require fast response, sliding average filtering can be selected; for scenes that require high-precision smoothing, Gaussian filtering or Kalman filtering can be selected.

[0146] 2. Apply the filtering algorithm: input the temperature-compensated rail width value into the selected filtering algorithm for smoothing. For example, when using sliding average filtering, calculate the average value of the data within a certain window to reduce measurement fluctuations. The specific formula is as follows:

[0147] ;

[0148] Where, is the compensated rail width value measured for the i-th time, and N is the window size.

[0149] 3. Generate track width reference parameter: output the filtered track width value as the final track width reference parameter.

[0150] Identify and track the measured points of the overhead contact line through the machine vision module, calculate the trajectory in combination with the motion prediction algorithm and the spatial reference coordinate system, and adjust the gimbal to continuously align the laser measurement beam with the measured points, including:

[0151] Step S410: Real-time processing of the overhead contact line image collected by the machine vision module using a target detection algorithm to identify the measured points of the overhead contact line and output their pixel coordinates in the image coordinate system; simultaneously suppressing environmental interference through an image enhancement algorithm to ensure that the recognition confidence of the measured points is not lower than the preset confidence threshold; calling the pre-set camera intrinsic matrix and extrinsic matrix to convert the pixel coordinates to three-dimensional coordinates in the spatial reference coordinate system.

[0152] Target detection algorithm: a computer vision technique used to identify specific objects in an image and determine their location and size. Image coordinate system: a system of coordinates for the position of pixels in an image, usually with the top-left corner of the image as the origin. Pixel coordinates: the position of a pixel in an image, usually represented as (x, y). Image enhancement algorithm: an algorithm used to improve image quality, making it more suitable for subsequent processing. Confidence threshold: a preset value used to determine the reliability of the target detection result, only the detection result with a confidence higher than the threshold will be accepted.

[0153] The acquisition method and process are described as follows:

[0154] 1. Collect overhead contact line images: use the machine vision module to collect images of the overhead contact line in real time. Ensure that the resolution and frame rate of the images meet the requirements of subsequent processing.

[0155] 2. Apply target detection algorithm: apply a target detection algorithm such as YOLO, SSD, or FasterR-CNN to the collected images to identify the measured points of the overhead contact line. The target detection algorithm will locate the measured points in the image and output their pixel coordinates in the image coordinate system.

[0156] 3. Image enhancement processing: to improve the accuracy of the measured point recognition, apply an image enhancement algorithm to suppress environmental interference such as light changes, shadows, reflections, etc. Image enhancement algorithms may include histogram equalization, contrast enhancement, noise reduction, etc.

[0157] 4. Ensure confidence: check the confidence value output by the target detection algorithm to ensure that the recognition confidence of the measured points is not lower than the preset confidence threshold. If the confidence is lower than the threshold, the image may need to be re-collected or the parameters of the image enhancement algorithm may need to be adjusted.

[0158] 5. Output pixel coordinates: output the pixel coordinates of the identified measured points.

[0159] 6. Call camera intrinsic and extrinsic matrices: Use the pre-calibrated camera intrinsic and extrinsic matrices. The camera intrinsic matrix includes parameters such as focal length, optical center coordinates, etc., which are used to convert pixel coordinates to coordinates in the camera coordinate system. The extrinsic matrix describes the position and attitude of the camera in the world coordinate system, which is used to convert coordinates in the camera coordinate system to coordinates in the world coordinate system.

[0160] 7. Coordinate conversion: For each identified measurement point, use the camera intrinsic matrix to convert its pixel coordinates to coordinates in the camera coordinate system. Then, use the extrinsic matrix to convert the coordinates in the camera coordinate system to three-dimensional coordinates in the physical coordinate system.

[0161] Step S420, based on the sliding window method, analyze the three-dimensional coordinate sequence of the measurement point in the spatial reference coordinate system of the previous frame, and extract the motion state parameters including instantaneous speed, acceleration and motion direction angle.

[0162] Where, sliding window method: a time series analysis method, by sliding a fixed size window on the data sequence to analyze the local characteristics of the data. Instantaneous speed: the speed at a certain time or a certain position, which can be calculated by the ratio of position change and time change. Acceleration: the rate of change of speed, indicating the speed of change of speed over time. Motion direction angle: the angle between the direction of motion of the object and the reference direction (such as the horizontal direction). Motion state parameters: parameters describing the motion state of the object, including speed, acceleration and direction angle, etc.

[0163] The acquisition method and process are described as follows:

[0164] 1. Collect the three-dimensional coordinate sequence of the measurement point: From the continuous image frames, collect the three-dimensional coordinates of the measurement point in each frame to form a time sequence.

[0165] 2. Apply sliding window method: Apply sliding window method on the three-dimensional coordinate sequence of the measurement point to analyze the data in the window. The window size can be set according to actual needs, usually choose a window size that can contain enough data points for reliable analysis.

[0166] 3. Calculate instantaneous speed: In the sliding window, calculate the three-dimensional coordinate change of the measurement point between two consecutive frames. According to the coordinate change and the time interval between frames, calculate the instantaneous speed.

[0167] The formula is: Where, v is the instantaneous speed, is the three-dimensional coordinate change, is the time interval.

[0168] 4. Calculate acceleration: Within the sliding window, calculate the change in instantaneous velocity. From the change in instantaneous velocity and the time interval, calculate the acceleration. The formula is: where a is the acceleration and Δv is the change in velocity.

[0169] 5. Calculate direction angle of motion: From the change in three-dimensional coordinates of the measured point in consecutive frames, calculate its direction of motion. Use the arctangent function to calculate the direction angle. The formula is: where θ is the direction angle and Δy and Δx are the pixel coordinate changes in the y-axis and x-axis respectively.

[0170] 6. Output motion state parameters: Output the calculated motion state parameters such as instantaneous velocity, acceleration, and direction angle of motion, to provide data support for subsequent motion prediction.

[0171] Step S430, adaptively match the motion prediction model according to the fluctuation characteristics of the motion state parameters, and output the predicted three-dimensional coordinates of the measured point at the future time in the spatial reference coordinate system using the matched motion prediction model.

[0172] where motion prediction model: a mathematical model used to predict the future position of an object, based on the historical motion data of the object. Fluctuation characteristics: refer to the characteristics of the change of the motion state parameters (such as velocity, acceleration) of the object over time, used to select the most suitable motion prediction model. Predicted three-dimensional coordinates: the expected position of the object at a future time calculated by the motion prediction model, in the three-dimensional coordinates in the spatial reference coordinate system.

[0173] The necessary process is as follows: 1. Input historical motion data: input the historical motion data (such as position, velocity, acceleration) of the measured point into the selected motion prediction model. These data are usually the motion state parameter sequences obtained from the sliding window method. 2. Model prediction: use the selected motion prediction model to predict the position of the measured point at the future time according to the historical motion data. The model output is the predicted three-dimensional coordinates in the spatial reference coordinate system. 3. Output predicted three-dimensional coordinates: output the predicted three-dimensional coordinates, which represent the expected position of the measured point at a future time. These predicted coordinates will be used for subsequent gimbal adjustment and measurement beam alignment.

[0174] The specific process of adaptively matching the motion prediction model according to the fluctuation characteristics of the motion state parameters can refer to steps S431 to S433, which are not repeated here.

[0175] Step S440, convert the predicted three-dimensional coordinates into predicted pixel coordinates in the image coordinate system, and call the gimbal kinematics model to convert into real-time adjustment amounts of the gimbal pitch angle and azimuth angle.

[0176] Wherein, the camera intrinsic matrix: a matrix describing the internal parameters of the camera, including focal length, optical center coordinates, etc., used to convert pixel coordinates to coordinates in the camera coordinate system. Gimbal kinematics model: a model describing the relationship between gimbal (controllable gimbal) motion and gimbal angle change, used to convert pixel coordinates to physical adjustment amount of the gimbal. Pitch angle: the angle of the gimbal rotating around the horizontal axis, used to adjust the vertical direction of the measuring light beam. Azimuth angle: the angle of the gimbal rotating around the vertical axis, used to adjust the horizontal direction of the measuring light beam.

[0177] The acquisition method and process are described as follows:

[0178] 1. Call the camera intrinsic matrix: use the camera intrinsic matrix obtained by pre-calibration to convert the predicted three-dimensional coordinates to coordinates in the camera coordinate system. The camera intrinsic matrix usually includes focal length, optical center coordinates and other parameters, which are used to convert three-dimensional coordinates to normalized camera coordinates.

[0179] 2. Calculate the predicted pixel coordinates in the image coordinate system: calculate the pixel coordinates in the image coordinate system from the coordinates in the camera coordinate system. This process is completed by applying the camera intrinsic matrix in reverse, ensuring that the predicted three-dimensional coordinates can be accurately mapped to the image plane.

[0180] 3. Apply the gimbal kinematics model: according to the gimbal kinematics model, convert the pixel coordinates in the image coordinate system to the physical adjustment amount of the gimbal. The gimbal kinematics model describes the relationship between the gimbal rotation angle and the image coordinates, which is used to calculate the angle that the gimbal needs to rotate.

[0181] 4. Calculate the pitch angle and azimuth angle adjustment amount: according to the pixel coordinates in the image coordinate system, calculate the pitch angle and azimuth angle that the gimbal needs to adjust. The specific method is as follows: if the pixel coordinates indicate that the measured point is above or below the image center, the pitch angle of the gimbal needs to be adjusted. If the pixel coordinates indicate that the measured point is on the left or right side of the image center, the azimuth angle of the gimbal needs to be adjusted.

[0182] 5. Generate adjustment instructions: generate gimbal adjustment instructions according to the calculated pitch angle and azimuth angle adjustment amount. The adjustment instructions include the angle and direction that the gimbal needs to rotate, which are used to accurately align the measured point.

[0183] 6. Output adjustment instructions: output the generated gimbal adjustment instructions to drive the servo motor of the gimbal to adjust. These instructions ensure that the gimbal can quickly respond and accurately align the measured point.

[0184] Step S450, generate control instructions based on the adjustment amount, drive the servo motor of the gimbal to dynamically execute the adjustment, and combine the real-time position deviation feedback in the space reference coordinate system to make the laser measuring light beam continuously align the measured point.

[0185] The specific process can refer to steps 1 to 5, which will not be repeated here.

[0186] Synchronization suppresses environmental interference through image enhancement algorithms, ensuring that the recognition confidence of the measured point is not lower than the pre-set confidence threshold, including:

[0187] Step S411, real-time collection of environmental data by environmental sensors, identification of interference types, including backlight, dust and rain and snow, and quantification of interference intensity.

[0188] Among them, the environmental sensor: a device used to detect the surrounding environmental conditions (such as light, dust concentration, humidity, etc.). Interference type: refers to environmental factors that may affect image quality and machine vision system performance, such as backlight, dust and rain and snow, etc. Interference intensity: the intensity or degree of influence of the quantified interference factor.

[0189] The acquisition method and process are described as follows:

[0190] 1. Real-time collection of environmental data: use environmental sensors (such as light sensors, dust sensors, humidity sensors, etc.) to monitor the current environmental conditions in real time. Sensor data can include light intensity, dust concentration, humidity level, temperature, etc.

[0191] 2. Identify interference types: identify possible interference types that may affect image quality based on sensor data.

[0192] For example, high light intensity may indicate backlight conditions, high dust concentration may indicate dust interference, and specific readings from the humidity sensor may indicate rain and snow weather.

[0193] 3. Quantify interference intensity: quantify the identified interference types to determine their intensity levels. For example, backlight interference can be classified as "mild", "moderate" and "severe" according to the readings of light intensity. Similarly, dust concentration and humidity level can also be converted into corresponding interference intensity levels.

[0194] 4. Output interference information: output the identified interference types and quantified interference intensity.

[0195] Step S412, based on the interference type and intensity, adaptively select and apply image enhancement algorithms.

[0196] Among them, the image enhancement algorithm: an algorithm used to improve image quality, making the image more suitable for subsequent processing, such as improving contrast, enhancing edges, etc. Adaptive selection: automatically select the most appropriate image enhancement algorithm based on real-time collection of environmental data and interference types.

[0197] The acquisition method and process are described as follows:

[0198] 1. Analyze the type and intensity of interference: Use the identified interference type and quantified interference intensity from step S411 to analyze the specific impact of the current environment on image quality.

[0199] 2. Adaptively select image enhancement algorithms: Based on the interference type and intensity, select the most appropriate algorithm from a pre-set image enhancement algorithm library. For example, for backlight interference, histogram equalization or CLAHE may be selected; for dust interference, denoising algorithms such as median filtering or bilateral filtering may be selected.

[0200] 3. Apply image enhancement algorithms: Apply the selected image enhancement algorithm to the original image to improve image quality. For example, if histogram equalization is selected, the contrast of the image will be improved by stretching the gray scale distribution; if a denoising algorithm is selected, the noise in the image will be reduced through filtering operations.

[0201] 4. Adjust algorithm parameters: Dynamically adjust the parameters of the image enhancement algorithm according to the specific situation of the interference to achieve the best enhancement effect. For example, for histogram equalization, the equalization range or intensity may need to be adjusted; for denoising algorithms, the filter window size or intensity may need to be adjusted.

[0202] 5. Output the enhanced image: Output the image after image enhancement processing for subsequent measured point recognition and tracking.

[0203] Step S413, calculate the objective quality indicators of the enhanced image, including contrast, sharpness and noise level; output the recognition confidence score based on the quality indicators through the pre-trained support vector machine model.

[0204] Objective quality indicators: A series of quantitative measures used to evaluate the quality of image enhancement, including contrast, sharpness and noise level, etc. Support vector machine (SVM) model: A supervised learning algorithm used for classification and regression analysis, which is used here to predict the recognition confidence score based on image quality indicators.

[0205] The acquisition method and process are described as follows:

[0206] 1. Calculate the contrast: Contrast is a measure of the difference between light and dark areas in an image, which can usually be evaluated by calculating the width or standard deviation of the image histogram. The higher the contrast, the greater the difference between light and dark areas in the image, and the clearer the details.

[0207] 2. Calculate the sharpness: Sharpness reflects the clarity of edges in an image, which can be calculated by edge detection algorithms (such as Sobel operator) or using sharpness metrics (such as Laplacian operator). The higher the sharpness, the clearer the edges in the image.

[0208] 3. Calculate noise level: Noise level reflects random brightness or color changes in the image. It can be evaluated by calculating the variance of the image or using specific noise evaluation algorithms. The lower the noise level, the better the image quality.

[0209] 4. Evaluate confidence score using pre-trained SVM model: The calculated objective quality indicators (contrast, sharpness, and noise level) are input into a pre-trained SVM model. The SVM model outputs a confidence score based on these features, which reflects the impact of image quality on subsequent recognition tasks.

[0210] 5. Output confidence score: The confidence score output by the SVM model is output for subsequent confidence threshold comparison.

[0211] Step S414, compare the confidence score with the pre-set confidence threshold in real time.

[0212] Step S415, if the confidence score ≥ the pre-set confidence threshold, output the enhanced image to the machine vision module for subsequent recognition.

[0213] The process is described as follows:

[0214] 1. Real-time comparison of confidence score: Compare the confidence score calculated in step S613 with the pre-set confidence threshold in real time.

[0215] 2. Determine if the confidence meets the requirements: If the confidence score is greater than or equal to the pre-set confidence threshold, it is considered that the enhanced image quality meets the requirements and can be used for subsequent recognition tasks.

[0216] 3. Output enhanced image: When the confidence score meets the requirements, output the enhanced image to the machine vision module. This step ensures that only high-quality image data is used for subsequent processing and analysis, thereby improving the recognition accuracy and reliability of the entire system.

[0217] 4. Prepare for subsequent recognition: After the machine vision module receives the enhanced image, it prepares to perform subsequent target detection and feature point recognition tasks.

[0218] Step S416, if the confidence score < the pre-set confidence threshold, dynamically adjust the enhancement algorithm parameters and re-execute the enhancement and evaluation, with a maximum of 3 cycles.

[0219] Wherein, the enhancement algorithm parameters: adjustable settings in the image enhancement algorithm, such as filter size, contrast enhancement intensity, etc. These parameters affect the effectiveness of the algorithm. Re-evaluation: After adjusting the enhancement algorithm parameters, the quality of the enhanced image is evaluated again and the confidence score is calculated.

[0220] The acquisition method and process are described as follows:

[0221] 1. Determine confidence score is insufficient: If the confidence score in step S414 is lower than a pre-set confidence threshold, it indicates that the current image enhancement effect does not meet the requirements.

[0222] 2. Dynamically adjust enhancement algorithm parameters: Based on the current image quality and confidence score, dynamically adjust the parameters of the image enhancement algorithm. For example, if the noise level is high, the strength of the denoising algorithm can be increased; if the contrast is insufficient, the amplitude of contrast adjustment can be enhanced.

[0223] 3. Re-execute image enhancement: Re-enhance the image using the adjusted parameters.

[0224] 4. Recalculate objective quality indicators: Recalculate the objective quality indicators such as contrast, sharpness and noise level for the re-enhanced image.

[0225] 5. Re-evaluate confidence score through SVM model: Use the updated objective quality indicators to re-calculate the recognition confidence score through the pre-trained support vector machine model.

[0226] 6. Loop count control: Set an upper limit on the number of loops (e.g. 3 times) to avoid infinite loops. If the confidence score reaches the pre-set threshold before reaching the upper limit, stop the loop.

[0227] Adaptive matching of motion prediction model according to fluctuation characteristics of motion state parameters includes:

[0228] Step S431, based on the sliding window method, analyze the motion state parameter sequence of the previous frame number, and calculate the acceleration variance.

[0229] Wherein, the sliding window method: a method for analyzing time series data, by sliding a fixed size window on the data sequence to analyze the local characteristics of the data. Motion state parameters: including speed, acceleration, etc., parameters describing the motion state of the object. Acceleration variance: the dispersion degree of acceleration data, used to measure the stability of the object motion.

[0230] The acquisition method and process are described as follows: 1. Collect motion state parameter sequence: collect the motion state parameters of the measured point from consecutive image frames to form a time series. 2. Apply sliding window method: apply sliding window method on the motion state parameter sequence to analyze the data within the window. The window size can be set according to actual needs, usually choose a window size that can contain enough data points for reliable analysis. 3. Calculate acceleration: in the sliding window, calculate the speed change of the measured point between two consecutive frames, then calculate the acceleration according to the speed change and the time interval between frames. The formula is: ; where a is the acceleration, Δv is the change in velocity, and Δt is the time interval.4. Calculate the acceleration variance: Within the sliding window, calculate the variance of all acceleration data. The variance formula is: where, is the variance of acceleration, N is the number of data points within the window, and a i is the i-th acceleration data point, and μ is the mean of the acceleration data.5. Output the acceleration variance: Output the calculated acceleration variance to provide data support for subsequent motion prediction models.

[0231] Step S432, if the acceleration variance is lower than the preset stationary threshold, match the Kalman filter algorithm for recursive prediction.

[0232] where, Kalman filter algorithm: A high-efficiency recursive filter that can estimate the state of a dynamic system from a series of noisy measurements. It is widely used in signal processing and time series analysis, especially in predicting and estimating dynamic systems in noisy environments. Recursive prediction: Using previous time step data to predict the state of the next time step, it is a time series-based prediction method. Stationary threshold: A pre-set threshold value used to determine whether the acceleration variance is within a small range, thereby determining whether to use the Kalman filter algorithm for prediction.

[0233] The acquisition method and process are described as follows: 1. Determine the acceleration variance: Check whether the acceleration variance calculated in step S431 is lower than the preset stationary threshold. The preset stationary threshold is set according to system requirements and experimental data to determine whether the system motion is relatively stable. 2. Match the Kalman filter algorithm: If the acceleration variance is lower than the stationary threshold, it indicates that the motion of the measured point is relatively stable, and the Kalman filter algorithm is selected for motion prediction. The Kalman filter algorithm is suitable for processing linear systems with Gaussian noise and can provide optimal estimation of system state. 3. Initialize the Kalman filter: Set the initial state of the Kalman filter, including the initial estimate and the initial error covariance matrix. The initial estimate can be based on previous measurement data, and the error covariance matrix reflects the uncertainty of the initial estimate. 4. Recursive prediction: Use the Kalman filter algorithm for recursive prediction to predict the state of the measured point at future time. Kalman filtering includes two main steps: prediction and update. In the prediction step, the next state is predicted based on the system model; in the update step, the prediction is corrected using new measurement data to obtain a more accurate state estimate. 5. Output the prediction result: Output the future state of the measured point predicted by the Kalman filter algorithm.

[0234] Step S433, if the acceleration variance exceeds the preset shaking threshold, match the long-short term memory network time series prediction model.

[0235] Among them, long short-term memory network (LSTM): a special type of recurrent neural network (RNN) that can learn long-term dependencies. LSTM is suitable for processing and predicting important events with long intervals and delays in time series. Time series prediction model: a model that predicts future data points based on historical data, commonly used to process data sequences arranged in chronological order. Shake threshold: a pre-set threshold value to determine whether the acceleration variance exceeds a certain value, thereby determining whether to use the LSTM model for prediction.

[0236] The acquisition method and process are described as follows:

[0237] 1. Determine the acceleration variance: check whether the acceleration variance calculated in step S431 exceeds the preset shake threshold. The shake threshold is used to identify cases where the measured point motion has large fluctuations or instability. 2. Match the LSTM time series prediction model: if the acceleration variance exceeds the shake threshold, it indicates that the motion of the measured point may have unpredictable fluctuations, and the LSTM model is selected for motion prediction at this time. The LSTM model can handle complex nonlinear relationships and long-term dependencies, making it suitable for prediction in this case. 3. Prepare training data: collect historical motion state parameters of the measured point, such as position, velocity, acceleration, etc., as training data. The data needs to be properly preprocessed, such as normalization, splitting into training set and test set, etc. 4. Train the LSTM model: use the prepared training data to train the LSTM model. During training, adjust the model's hyperparameters, such as learning rate, batch size, number of hidden layer units, etc., to achieve optimal performance. 5. Perform time series prediction: use the trained LSTM model to predict the future motion state of the measured point. According to the output of the model, predict the position of the measured point at future time. 6. Output the prediction result: output the future state of the measured point predicted by the LSTM model, which is used for subsequent gimbal adjustment and measurement beam alignment.

[0238] Based on the adjustment amount, generate control instructions to drive the servo motor of the gimbal to dynamically execute adjustments, combined with real-time position deviation feedback in the space reference coordinate system, so that the laser measurement beam continuously aligns with the measured point, including:

[0239] Step 1, based on the adjustment amount of the gimbal pitch angle and azimuth angle, and the real-time position deviation of the measured point in the space reference coordinate system, an incremental PID control algorithm is used to calculate the real-time control amount, and a control instruction containing the target angle parameter is generated.

[0240] Incremental PID control algorithm: An implementation of PID control that calculates the change in control variable based on the error increment from the current and previous control periods, rather than directly calculating the control variable. Control command: An instruction sent to the servo motor, indicating how the motor should adjust the gimbal's attitude. Pitch and yaw angles: The angles that the gimbal needs to adjust, corresponding to vertical and horizontal rotations, respectively. Real-time position deviation: The deviation between the current position of the measured point in the spatial reference coordinate system and the target position.

[0241] The acquisition method and process are described as follows:

[0242] 1. Determine the angle adjustment amount: Based on the pitch and yaw angle adjustment amounts calculated in step, determine the angles that the gimbal needs to adjust.

[0243] 2. Initialize PID parameters: Set the parameters of the PID controller, including the proportional coefficient (P), integral coefficient (I), and derivative coefficient (D). These parameters affect the response speed and stability of the control system.

[0244] 3. Calculate the PID control amount: Use the incremental PID control algorithm to calculate the change in control amount. The algorithm takes into account the current error, the error of the previous period, and the rate of change of the error. The formula for calculating the incremental PID control amount is:

[0245] .

[0246] where, is the change in control amount at the kth moment, is the error increment at the kth moment, Kp, Ki, and Kd are the proportional, integral, and derivative coefficients, respectively, and Δt is the sampling time.

[0247] 4. Generate control commands: Convert the calculated control amount change into specific control commands, which will be sent to the servo motor. The control commands should include information such as the angles that the gimbal needs to adjust and the speed, etc. For example, the control command can be an instruction containing the target pitch and yaw angles, indicating how the servo motor should adjust the attitude of the gimbal.

[0248] 5. Output control commands: Output the generated control commands, ready to be sent to the servo driver. These commands will be used to drive the servo motor of the gimbal to adjust, ensuring that the laser measurement beam can accurately aim at the measured point.

[0249] Step 2, send the control command to the servo driver of the gimbal through the preset servo communication protocol.

[0250] Among them, the servo communication protocol: a standardized communication method, used to control the transmission of instructions for servo motor, to ensure that the control command can be accurately sent to the servo driver. Servo driver: an electronic device, used to receive control instructions and drive servo motor to perform actions according to instructions.

[0251] The specific steps are as follows:

[0252] 1, prepare control instructions: according to the control instructions generated in step 1, these instructions contain the angle and speed information that the gimbal needs to adjust. Ensure that the format of the control instruction meets the requirements of the servo driver.

[0253] 2, select servo communication protocol: use the preset servo communication protocol, such as RS-232, RS-485, CAN bus or Ethernet, etc., to ensure that the control instruction can be reliably transmitted to the servo driver. The selected communication protocol should be selected according to the interface of the servo driver and the actual application environment.

[0254] 3, send control instructions: through the selected communication protocol, send the control instructions from the control system to the servo driver of the gimbal. Ensure the data integrity and accuracy during communication, avoid data loss or error.

[0255] 4, confirm instruction sending: after sending the control instruction, confirm that the servo driver has received the instruction. This can be done by checking the communication status or receiving the feedback signal of the servo driver. If communication errors or instructions are not received correctly, you need to resend the control instruction.

[0256] Step 3, after the servo driver parses the instruction, it drives the servo motor to rotate according to the preset acceleration and deceleration curve.

[0257] Acceleration and deceleration curve: the speed change curve of servo motor during startup and stop, used to smooth the acceleration and deceleration process of the motor, reduce mechanical impact.

[0258] The specific steps are as follows: 1, parse control instructions: after receiving the control instructions, the servo driver parses the instruction content and extracts the target pitch angle and azimuth angle parameters. Ensure that the parsing process is accurate and error-free, so as to correctly execute the subsequent actions. 2, drive servo motor: according to the parsed control instructions, the servo driver drives the servo motor to rotate according to the preset acceleration and deceleration curve. The acceleration and deceleration curve helps to reduce the mechanical impact of the motor during startup and stop, prolongs the service life of the equipment and improves the control accuracy.

[0259] Step 4, real-time acquisition of the actual angle position of the gimbal through the motor encoder, combined with the position deviation data in the space reference coordinate system, and closed loop comparison with the target angle parameter, to calculate the angle error.

[0260] The specific steps are as follows: 1. Real-time acquisition of angular position: use the motor encoder to real-time acquire the actual angular position of the gimbal, including the pitch angle and azimuth angle. The high-precision measurement of the encoder ensures the accuracy and reliability of the gimbal position. 2. Obtain position deviation data: obtain the real-time position deviation data of the measured point from the spatial reference coordinate system. These data reflect the deviation between the current position of the measured point in space and the target position. 3. Closed-loop comparison: compare the actual angular position feedback by the encoder with the target angular parameters, and calculate the angular error. The specific formula is as follows: angular error = target angle - actual angle; 4. Determine whether the error is within the tolerance range: check whether the calculated angular error is within the preset tolerance range. The preset tolerance range is set according to the actual measurement accuracy requirements, for example, the tolerance range of the pitch angle and azimuth angle can be set to ±0.1 degrees. 5. Output angular error: if the angular error is within the preset tolerance range, it means that the gimbal has accurately aligned the measured point, and the angular error can be output and continued to be monitored. If the angular error exceeds the preset tolerance range, further adjustment is needed.

[0261] Step 5, if the angular error exceeds the preset tolerance range, recalculate the PID control amount based on the latest position deviation, generate a secondary adjustment instruction and repeat steps 2 to 4 until the laser measurement beam accurately aligns the measured point, and the position deviation in the spatial reference coordinate system is less than the preset threshold. If the angular error exceeds the preset tolerance range, recalculate the PID control amount based on the latest position deviation. Use the incremental PID control algorithm to calculate the change of the control amount according to the current error, the last period error and the rate of change of the error.

[0262] Fusion of compensated laser data, filtered inertial data and visual tracking data, through multi-source fusion algorithm, output the three-dimensional coordinates of the measured point of the overhead contact system, including:

[0263] Step S510, synchronize the real-time collected environmental temperature and vibration data with the compensated laser data, filtered inertial data and visual tracking data in time stamp, to ensure that the multi-source data is aligned in time dimension.

[0264] The specific steps are as follows: 1. Collect multi-source data: real-time collect environmental temperature data and vibration data. Collect compensated laser ranging data. Collect filtered inertial measurement unit (IMU) data.

[0265] Collecting the visual tracking data output by the machine vision module. 2. Timestamping: Adding precise timestamps to each data ensures that the collection time of the data is accurately recorded. The precision of the timestamp should be high enough to meet the requirements of high-precision measurement. 3. Timestamp alignment: Align the timestamps of all data to ensure that the data of different sensors are aligned in the time dimension. The specific method is as follows: Select a reference timestamp, usually the timestamp of the machine vision module, because its collection frequency is usually higher. Adjust the timestamps of other sensor data to align with the reference timestamp. If the timestamps of the data do not completely match, you can adjust them by interpolation or time correction method. 4. Data synchronization verification: Verify the synchronization of the data after timestamp alignment to ensure that all data are aligned in the time dimension. The synchronization can be verified by calculating the difference between the timestamps to ensure that the difference is within the allowed error range.

[0266] Step S520, the synchronized multi-source data is uniformly converted to the spatial reference coordinate system, the Kalman filter algorithm is used to remove the abnormal values in the data, and the pre-set temperature compensation coefficient and vibration noise reduction algorithm are used for secondary optimization, so that the error of the pre-processed data is controlled within the pre-set data error threshold.

[0267] The necessary process is as follows: 1. Unified conversion to spatial reference coordinate system: The synchronized multi-source data (including compensated laser data, filtered inertial data, and visual tracking data) is uniformly converted to the spatial reference coordinate system. This step ensures that all data is represented in the same coordinate system, facilitating subsequent fusion processing. Using the preset coordinate transformation matrix, each type of data is converted from its original coordinate system to the spatial reference coordinate system. For example, for laser ranging data, use the extrinsic matrix of the laser ranging module for conversion; for inertial data, use the extrinsic matrix of the IMU for conversion; for visual tracking data, use the intrinsic and extrinsic matrices of the camera for conversion. 2. Remove outliers using Kalman filter algorithm: Apply the Kalman filter algorithm to process the converted data to remove outliers. The Kalman filter provides the best estimate of the system state through prediction and update steps, combining prior knowledge and new measurements. Initialize the state vector and covariance matrix of the Kalman filter. The state vector usually includes position, velocity, etc., and the covariance matrix represents the uncertainty of the state estimate. At each time step, use the state transition matrix to predict the next state, then update the state estimate according to the new measurement data. By comparing the predicted value and the measured value, identify and remove outliers. 3. Secondary optimization through temperature compensation coefficient and vibration noise reduction algorithm: After Kalman filtering, the data is again optimized using the preset temperature compensation coefficient and vibration noise reduction algorithm. This step further reduces the impact of environmental factors on the data and improves the accuracy of the data. Use the temperature compensation coefficient to compensate for the data to ensure that the data is not affected by temperature changes. For example, for laser ranging data, adjust the ranging result according to the current environmental temperature and the preset temperature compensation coefficient. Apply the vibration noise reduction algorithm to reduce the interference of vibration on the data. For example, use a low-pass filter or adaptive filter to remove high-frequency vibration noise. 4. Control data error within the preset threshold: Evaluate the data error after preprocessing to ensure that the error is within the preset data error threshold. The preset threshold is set according to the measurement accuracy requirements, for example, the position error can be set to ±1mm, and the velocity error can be set to ±0.01m / s. If the data error exceeds the preset threshold, adjust the parameters of the temperature compensation coefficient and vibration noise reduction algorithm, or reapply the Kalman filter algorithm until the data error meets the requirements.

[0268] Step S530, based on the sensor accuracy parameters and historical performance data of the machine vision module, inertial measurement unit, and laser ranging module, combined with real-time environmental temperature and vibration data, dynamically assigns weight coefficients to each data source, and uses the weighted least squares method to fuse and solve the preprocessed data, to obtain the optimal three-dimensional coordinate estimate value of the measured point of the catenary, ensuring that the fusion accuracy meets the preset fusion accuracy requirements.

[0269] where the weighted least squares method: a mathematical optimization technique used to find the best function match of data in the presence of errors. Fusion accuracy: the accuracy of the result after data fusion, meeting the preset accuracy requirements.

[0270] The necessary process is described as follows: 1. Evaluate sensor accuracy and performance: based on the accuracy and historical performance data of sensors such as machine vision modules, inertial measurement units (IMU), laser ranging modules, etc. 2. Dynamically allocate weight coefficients: based on the evaluation results and real-time environmental temperature and vibration data, dynamically allocate weight coefficients for each sensor data. The weight coefficient reflects the relative credibility and importance of each sensor data in the fusion process. 3. Apply the weighted least squares method: use the weighted least squares method for data fusion to solve the optimal three-dimensional coordinate estimate. The weighted least squares method considers the weights of different sensor data to minimize the weighted sum of squares of errors. The specific formula is as follows: ; where, is the optimal three-dimensional coordinate estimate, is the measurement data of the i-th sensor, is the measurement matrix, is the measurement error covariance matrix, is the weight coefficient, is the number of sensors. 4. Adjust the weight coefficient: adjust the weight coefficient according to the real-time environmental changes and sensor performance to optimize the fusion accuracy. For example, if a certain sensor performs poorly in a particular environment, its weight coefficient can be reduced. 5. Verify fusion accuracy: check whether the fused data meets the preset fusion accuracy requirements. If not, the weight coefficient may need to be adjusted or the data fusion algorithm improved. 6. Output the fusion result: output the optimal three-dimensional coordinate estimate after fusion for subsequent steps.

[0271] And further in the dynamic allocation of weight coefficients based on sensor accuracy and historical performance data, it also includes a laser ranging dynamic correction mechanism based on train airflow disturbance, which specifically includes:

[0272] Step S531, real-time acquisition of airflow data and disturbance level judgment: through the preset miniature wind speed and direction sensor, real-time acquisition of airflow speed and direction data at the measurement point with a sampling frequency of ≥100Hz.

[0273] Step S532, combine the preset train running speed-airflow disturbance correlation model to divide the current airflow disturbance into weak (airflow speed ≤2m / s), moderate (airflow speed 2-5m / s), and strong (airflow speed >5m / s) levels.

[0274] Step S533, laser ranging dynamic correction model call: for different disturbance levels, weak disturbance uses linear compensation formula correction value = original value x (1+k1 x AV) (k1 is the linear compensation coefficient, AV is the air flow speed change);

[0275] Medium disturbance uses quadratic polynomial compensation model correction value = original value x (1+k2 x AV+k3 x AV²) (k2, k3 are polynomial coefficients);

[0276] Strong disturbance enables LSTM air flow disturbance prediction correction model, inputs real-time air flow data, train position and laser ranging original value, outputs predicted correction value and prediction error ≤10% of laser ranging module accuracy.

[0277] Step S534, correction data verification and weight dynamic adjustment: perform ≤1ms timestamp synchronized spatio-temporal correlation verification on the corrected laser ranging data and inertial measurement unit attitude data, calculate the ranging deviation before and after correction; if the deviation > 15% of the laser ranging module accuracy, adjust the correction model weight coefficient by 0.05 as the step size, iterate correction at most 3 times until the deviation meets the requirements.

[0278] Step S535, data fusion weight optimization: in the weighted least squares fusion, the weight coefficient of the laser ranging data after air flow correction is floated according to the disturbance level, weak disturbance is floated by 5%, medium disturbance is floated by 10%, and strong disturbance is floated by 15%, to ensure that the fusion accuracy meets the preset requirements (such as ±0.5mm).

[0279] Step S540, based on the fused three-dimensional coordinate data, calculate the gauge and pullout value through the geometric model, and output the comprehensive detection results containing three-dimensional coordinates, gauge and pullout value geometric parameters.

[0280] The acquisition method and process are described as follows:

[0281] 1. Obtain fused three-dimensional coordinate data: obtain fused three-dimensional coordinate data from step S530, which represents the position of the measured point of the overhead contact system in space.

[0282] 2. Establish geometric model: establish a mathematical model according to the geometric relationship between the track and the overhead contact system. The model should be able to calculate the gauge and pullout value according to the given three-dimensional coordinates.

[0283] 3. Calculate the gauge: use the geometric model to extract the Z-axis (vertical direction) coordinate value from the three-dimensional coordinates and calculate the gauge. The gauge is the vertical distance from the overhead contact line to the track plane.

[0284] 4、Calculate the pull-out value: Similarly, use the geometric model to extract the coordinate values of the X-axis (horizontal direction) and Y-axis (horizontal direction) from the three-dimensional coordinates, and calculate the pull-out value. The pull-out value is the horizontal distance between the overhead line and the track center line.

[0285] 5、Generate a comprehensive detection result: Integrate the calculated three-dimensional coordinates, height and pull-out value into a comprehensive detection result.

[0286] Based on the fused three-dimensional coordinate data, the height and pull-out value are calculated by a geometric model, including:

[0287] Step S541, based on the fused three-dimensional coordinate data, determine the track center and track surface reference based on the track width reference parameter, calculate the initial height value using a preset vertical projection formula, and calculate the initial pull-out value using a preset horizontal distance formula.

[0288] Calculate the height: Use the preset vertical projection formula to calculate the height. Generally, if the track plane is defined as Z=0, the calculation formula of the height is: H=|Z|; wherein Z is the vertical coordinate value in the fused three-dimensional coordinate data.

[0289] Calculate the pull-out value: Use the preset horizontal distance formula to calculate the pull-out value. Assuming that the track center line is at the position of X=0 and Y=0, the pull-out value can be obtained by calculating the Euclidean distance: ; wherein X and Y are the horizontal coordinate values in the fused three-dimensional coordinate data.

[0290] Step S542, based on the pre-output predicted three-dimensional coordinates of the measured point at the future time, calculate the deviation based on the real-time three-dimensional coordinates in the space reference coordinate system, and dynamically correct the height and pull-out value.

[0291] The acquisition method and process are described as follows:

[0292] 1、Calculate the deviation: Obtain the predicted three-dimensional coordinates of the measured point at the future time ; obtain the real-time three-dimensional coordinates of the measured point at the current time . Calculate the deviation: ; ; .

[0293] 2、Dynamically correct the height: Use the deviation ΔZ to correct the initial height value : , wherein .

[0294] 3、Dynamically correct the pull-out value: Use the deviations ΔX and ΔY to correct the initial pull-out value : . Wherein, .

[0295] 4. Verify the corrected value: Check the corrected guide height. and pull value Is it within a reasonable range? If the corrected value exceeds the preset reasonable range, it may be necessary to re-evaluate the prediction model or adjust the correction algorithm.

[0296] Step S543: Based on real-time ambient temperature and vibration data, apply a preset temperature compensation coefficient to correct the guide height value, and apply a preset vibration noise reduction algorithm to correct the pull-out angle parameter.

[0297] The process is briefly described as follows:

[0298] 1. Apply temperature compensation coefficient: Based on the real-time ambient temperature, the guide height value is corrected using a preset temperature compensation coefficient. The correction formula is as follows:

[0299] Where α is the temperature coefficient and ΔT is the temperature change. : The compensated conduction height value, that is, the corrected conduction height value after taking into account the effect of temperature change; : Corrected guide height value.

[0300] 2. Apply vibration noise reduction algorithm: Use a preset vibration noise reduction algorithm to correct the pull-out value in order to eliminate the influence of vibration.

[0301] Correction formula: ;in, Represents the vibration noise reduction function. : This indicates the pull-out value after vibration noise reduction processing, that is, the pull-out value after eliminating the influence of vibration; This indicates the pull-out value that has been dynamically adjusted.

[0302] Step S544: Compare the current guide height and pull-out value with the preset standard value in real time, and calculate the absolute deviation.

[0303] The process is described as follows: 1. Obtain the current calculation results: Obtain the corrected guide height Hcorrected and pull-out value Lcorrected from step S543. 2. Obtain historical data or standard values: Obtain historical guide height and pull-out value data for comparison, or preset standard values. 3. Calculate the absolute deviation: Calculate the absolute deviation between the current measurement value and the historical data or standard value.

[0304] In step S545, if the deviation exceeds the preset threshold, the exception handling mechanism is triggered.

[0305] Step S546, conversely, the calculated guide height and pull-out value are used as the actual output calculation results.

[0306] The abnormality processing mechanism includes: step S545.1, calling real-time coordinates, predicted coordinates, and environmental data, combining the recognition confidence of the machine vision module and the attitude fluctuation value of the inertial measurement unit, determining the abnormality type through a preset decision tree model, and the abnormality type includes data collection deviation, motion prediction misalignment, and environmental interference overload. The attitude fluctuation value is the amount of device attitude change detected by the inertial measurement unit. The decision tree model is a tree structure model used for classification and decision making, which is used to determine the abnormality type.

[0307] The process is briefly described as follows: 1. Collect data: collect real-time coordinates, predicted coordinates, environmental data, recognition confidence, and attitude fluctuation value. 2. Apply the decision tree model: use the decision tree model to determine the abnormality type according to the collected data. The decision tree model determines the abnormality type through a series of judgment conditions. 3. Output the abnormality type: output the determined abnormality type to provide a basis for subsequent correction.

[0308] Step S545.2, according to the abnormality type, adopt the matching correction strategy and execute it. The correction strategy is the corresponding correction measure taken for different abnormality types. The process is briefly described as follows: 1. Match the correction strategy: according to the abnormality type, select the corresponding correction strategy. 2. Execute the correction: for data collection deviation, recalibrate the sensor or adjust the data collection method. For motion prediction misalignment, retrain or select a more accurate prediction model. For environmental interference overload, apply environmental compensation algorithm to adjust the measurement value: compensated measurement value = measurement value x (1 + compensation coefficient x ΔT). 3. Output the correction result: output the corrected measurement result for use in subsequent steps.

[0309] Step S545.3, after executing the correction strategy, re-time stamp synchronize, coordinate convert, and weighted fusion of the multi-source data, solve the corrected three-dimensional coordinates, and recalculate the height of lead and the pull-out value. The weighted fusion is the combination of data from different sensors, and the fusion calculation is performed according to the weight and accuracy. The re-weighted fusion is the use of weighted least squares method to re-fuse the data.

[0310] Step S545.4, compare the current height of lead and pull-out value calculation results with historical data or standard values in real time, and calculate the absolute deviation.

[0311] Step S545.5, if the absolute deviation is less than or equal to the preset threshold, it is determined that the correction is successful; otherwise, re-execute the correction strategy up to a preset number of times.

[0312] Step S545.6, if the correction is successful, output the corrected height of lead and pull-out value as the final measurement result; otherwise, record the abnormality log including the abnormality type, processing process, and deviation data.

[0313] Exception log: records relevant information when an exception occurs, including exception type, processing process, and measurement results, etc.

[0314] Based on the same inventive concept, the embodiment of the application provides a track overhead line parameter high-precision measurement system of a multi-dimensional coupling algorithm, which comprises a memory and a processor, the memory is stored with a program capable of being run on the processor to realize the method shown in Figures 1 to 2 The embodiment of the present application is not limited to the protection scope of the application.

[0315] The embodiments of the present application are the preferred embodiments of the application, and are not limited to the protection scope of the application, so that: any equivalent changes made according to the structure, shape, principle of the application should be covered in the protection scope of the application.

Claims

1. A high-precision measurement method for track contact network parameters using a multi-dimensional coupling algorithm, characterized in that, Performed by a measuring device, which includes a machine vision module, an inertial measurement unit, a laser ranging module, and a gimbal, the method includes the following steps: The machine vision module acquires rail images, identifies rail head contour feature points and performs quality grading; based on the high-quality feature point set, the contour is fitted to calculate the initial rail width value, and the rail width reference parameters are generated by compensating with the rail thermal expansion coefficient and real-time rail temperature data; the inertial measurement unit is simultaneously started to acquire the initial attitude data of the measuring device, and a spatial reference coordinate system is established with the rail width reference as the scale reference and the initial attitude as the orientation reference. The control unit rotates according to a preset angle sequence, and the laser ranging module collects corresponding angle distance data. An angle-distance mapping table is generated through a spatiotemporal correlation algorithm, forming a calibration parameter table that is bound to the track width reference and calibrated in the spatial reference coordinate system. In the spatial reference coordinate system, the attitude angle of the measuring device relative to the track is calculated and compared with the absolute attitude angle measured by the inertial measurement unit to obtain the difference value. If the difference value exceeds the preset threshold, the track width reference is refitted and the spatial reference coordinate system and calibration parameter table are updated based on the track width reference and high-quality feature point historical data. Otherwise, ambient temperature and vibration data are collected, and preset temperature compensation coefficients and vibration noise reduction algorithms are called to compensate and filter the output data of the laser ranging module and the inertial measurement unit online to obtain compensated laser data and filtered inertial data. The machine vision module identifies and tracks the points to be measured on the contact wire, and the trajectory is calculated by combining the motion prediction algorithm and the spatial reference coordinate system. The gimbal is adjusted to keep the laser measurement beam continuously aligned with the points to be measured, and visual tracking data is output. By integrating compensated laser data, filtered inertial data, and visual tracking data, and using a multi-source fusion algorithm, the three-dimensional coordinates, guide height, and pull-out value parameters of the measured points on the contact wire are output.

2. The high-precision measurement method for track contact network parameters using the multi-dimensional coupling algorithm according to claim 1, characterized in that, The process of acquiring rail images using a machine vision module, identifying rail head contour feature points, and performing quality grading includes: The lighting conditions are adaptively adjusted by the multi-source dynamic lighting device in the machine vision module, and the rail image is preprocessed by a preset denoising algorithm and distortion correction matrix to obtain a high-quality rail head area image. The improved U-Net deep learning model is deployed in advance to perform real-time inference on the preprocessed image. Through attention mechanism and depth-separable convolutional structure, the key feature points of the rail head contour are identified and the confidence score of each feature point is output synchronously. The local gradient intensity variance of the feature points is calculated using the Sobel operator, and the Euclidean distance is calculated by comparing it with the preset standard profile of the rail head. Quality levels are determined based on multi-indicator collaborative judgment rules: feature points with confidence scores not lower than a preset threshold, gradient strength variance not exceeding a preset threshold, and Euclidean distance not exceeding a preset threshold are classified as high-quality points, and the rest are classified as low-quality points, forming a high-quality and low-quality feature point label set.

3. The high-precision measurement method for track contact network parameters using the multi-dimensional coupling algorithm according to claim 2, characterized in that, The initial rail width value is calculated by fitting the contour based on a high-quality feature point set, and the rail width reference parameters are generated by compensating with the rail thermal expansion coefficient and real-time rail temperature data, including: Extract a high-quality feature point set from the high- and low-quality feature point label set, call the pre-calibrated camera intrinsic and extrinsic parameter matrices, and convert the feature point pixel coordinates into three-dimensional coordinates in the physical coordinate system; The rail head profile is fitted based on the transformed three-dimensional coordinate data, and the initial rail width value is calculated by the spatial position difference of symmetrical feature points on both sides of the rail head. The preset rail thermal expansion coefficient is used, combined with real-time collected rail head temperature data, to perform temperature compensation correction on the initial rail width value. A filtering algorithm is used to eliminate measurement fluctuations in the compensated rail width value and generate rail width reference parameters.

4. The high-precision measurement method for track contact network parameters using the multi-dimensional coupling algorithm according to claim 1, characterized in that, The machine vision module identifies and tracks the points being measured on the overhead contact line. Combining motion prediction algorithms with a spatial reference coordinate system to calculate the trajectory, the pan-tilt unit is adjusted to continuously align the laser measurement beam with the points being measured. The target detection algorithm is used to process the catenary images acquired by the machine vision module in real time, identify the measured points of the catenary and output their pixel coordinates in the image coordinate system; at the same time, the image enhancement algorithm is used to suppress environmental interference to ensure that the confidence of the measured point identification is not lower than the preset confidence threshold; and the preset camera intrinsic and extrinsic parameter matrices are called to transform the pixel coordinates to three-dimensional coordinates in the spatial reference coordinate system. The sliding window method is used to analyze the three-dimensional coordinate sequence of the measured point in the spatial reference coordinate system of the preceding frames, and extract motion state parameters including instantaneous velocity, acceleration and motion direction angle. Based on the fluctuation characteristics of motion state parameters, an adaptive matching motion prediction model is used, and the matched motion prediction model is used to output the predicted three-dimensional coordinates of the measured point at future times in the spatial reference coordinate system. The predicted 3D coordinates are inversely calculated into predicted pixel coordinates in the image coordinate system, and the gimbal kinematic model is called to convert them into real-time adjustment values ​​for the gimbal pitch and azimuth angles. Based on the adjustment amount, control commands are generated to drive the servo motor of the gimbal to dynamically perform adjustments. Combined with real-time position deviation feedback in the spatial reference coordinate system, the laser measurement beam is continuously aligned with the measured point.

5. The high-precision measurement method for track contact network parameters using the multi-dimensional coupling algorithm according to claim 4, characterized in that, Simultaneously suppressing environmental interference through image enhancement algorithms to ensure that the confidence level of the identified test point is not lower than a preset confidence threshold includes: Environmental data is collected in real time by environmental sensors to identify the type of interference, including backlight, dust and rain and snow, and to quantify the intensity of the interference. Based on the type and intensity of interference, an image enhancement algorithm is adaptively selected and applied. Calculate objective quality metrics for the enhanced image, including contrast, sharpness, and noise level; output recognition confidence scores based on these quality metrics using a pre-trained support vector machine model. The confidence score is compared with a preset confidence threshold in real time. If the confidence score is greater than or equal to the preset confidence threshold, the enhanced image is output to the machine vision module for subsequent recognition. If the confidence level is less than the preset confidence threshold, the augmentation algorithm parameters are dynamically adjusted, and the augmentation and evaluation are re-executed, with a maximum of 3 iterations.

6. The high-precision measurement method for track contact network parameters using the multi-dimensional coupling algorithm according to claim 4, characterized in that, Adaptive matching motion prediction models based on the fluctuation characteristics of motion state parameters include: The motion state parameter sequence of the preceding frames is analyzed using the sliding window method, and the acceleration variance is calculated. If the acceleration variance is lower than the preset stability threshold, then the Kalman filter algorithm is used for recursive prediction. If the acceleration variance exceeds the preset sway threshold, then a long short-term memory network time-series prediction model is matched.

7. The high-precision measurement method for track contact network parameters using the multi-dimensional coupling algorithm according to claim 4, characterized in that, Based on the adjustment amount generation control command, the servo motor driving the gimbal dynamically performs adjustments. Combined with real-time position deviation feedback in the spatial reference coordinate system, this ensures that the laser measurement beam is continuously aligned with the measured point, including: Step 1: Based on the adjustment of the gimbal's pitch and azimuth angles, and by integrating the real-time position deviation of the measured point in the spatial reference coordinate system, an incremental PID control algorithm is used to calculate the real-time control quantity and generate a control command containing the target angle parameter. Step 2: Send the control commands to the servo driver of the gimbal via a preset servo communication protocol; Step 3: After parsing the instructions, the servo driver drives the servo motor to rotate according to the preset acceleration and deceleration curve; Step 4: The actual angle position of the gimbal is collected in real time by the motor encoder, and the position deviation data under the spatial reference coordinate system is compared with the target angle parameters in a closed loop to calculate the angle error. Step 5: If the angle error exceeds the preset tolerance range, the PID control quantity is recalculated based on the latest position deviation, a secondary adjustment command is generated, and steps 2 to 4 are repeated until the laser measurement beam is precisely aligned with the measured point and the position deviation in the spatial reference coordinate system is less than the preset threshold.

8. The high-precision measurement method for track contact network parameters using the multi-dimensional coupling algorithm according to claim 4, characterized in that, By fusing compensated laser data, filtered inertial data, and visual tracking data, and using a multi-source fusion algorithm, the output parameters include the three-dimensional coordinates, guide height, and pull-out value of the measured points on the contact wire. The real-time collected ambient temperature and vibration data are time-stamped and synchronized with compensated laser data, filtered inertial data, and visual tracking data to ensure that the multi-source data are aligned in the time dimension. The synchronized multi-source data is uniformly converted to a spatial reference coordinate system. The Kalman filter algorithm is used to remove outliers from the data. Through the aforementioned preset temperature compensation coefficient and vibration noise reduction algorithm, the preprocessed data error is controlled within the preset data error threshold. Based on the sensor accuracy parameters and historical performance data of the machine vision module, inertial measurement unit, and laser ranging module, and combined with real-time ambient temperature and vibration data, the weight coefficients of each data source are dynamically allocated. The weighted least squares method is used to fuse and solve the preprocessed data to find the optimal three-dimensional coordinate estimate of the contact wire measured point, ensuring that the fusion accuracy meets the preset fusion accuracy requirements. Based on the fused 3D coordinate data, the preset contact wire geometry model is called to calculate the guide height and pull-out value, and the comprehensive detection result including 3D coordinates, guide height and pull-out value is output.

9. A high-precision measurement system for track contact network parameters using a multi-dimensional coupling algorithm, characterized in that, The method includes a memory, a processor, and a program stored in the memory and executable on the processor, which, when loaded and executed by the processor, implements a high-precision measurement method for track contact network parameters using a multi-dimensional coupling algorithm as described in any one of claims 1 to 8.

Images