Railway wagon cleaning system and method

Abstract

The application discloses a railway freight car cleaning system and method, and relates to the technical field of railway transportation cleaning equipment. The railway freight car cleaning system and method, through a cleaning device and an image acquisition unit installed on a portal frame, combines a controller to execute the following steps: dynamically dividing an uncleaned area and a cleaned area, taking the instantaneous projection of the cleaning device in the image as the benchmark, and realizing automatic boundary tracking with the movement of the frame. The residual detection of the cleaned area is carried out, and when the standard is exceeded, the secondary cleaning is automatically triggered, forming a mechanism of self-detection and triggering secondary cleaning. In addition, the system also fuses the top and side camera data to generate a grid map, provides residual coal thickness and spatial distribution information, and realizes automatic identification and accurate positioning of multiple car types.

Description

Technical Field

[0001] A railway freight car cleaning system and method are disclosed, relating to the field of railway transportation cleaning equipment technology. Background Technology

[0002] In the energy and railway transportation sectors, after unloading from bulk cargo wagons, wet coal containing moisture adheres to the wagon's sides, forming coal deposits or rust. These debris not only affect the vehicle's appearance but also, during operation, can easily cause significant damage to passing high-speed trains, passenger cars, and freight cars due to the dust or splashing generated when the coal dries and loosens.

[0003] The side walls of railway freight cars are typically cleaned manually using wire brushes, scrapers, and brooms, which poses high safety risks. Existing cleaning devices for railway freight cars cannot adaptively adjust to the actual distribution, thickness, and adhesion strength of residual material on the side walls. This results in incomplete cleaning of areas with thick, strongly adhesive residue, while over-cleaning thinner areas leads to energy waste and excessive equipment wear. Furthermore, existing cleaning devices for railway freight cars generally rely on remote control or pre-programmed operation, lacking a mechanism for real-time self-checking and closed-loop feedback of cleaning effectiveness. They cannot automatically determine cleanliness and trigger secondary cleaning after a single operation, often requiring manual re-inspection and re-cleaning, thus failing to achieve truly efficient and reliable unmanned automated cleaning operations.

[0004] It is evident that existing railway freight car cleaning devices suffer from high safety risks associated with manual operation, insufficient automated cleaning, lack of self-inspection mechanisms after cleaning, and failure to trigger secondary cleaning, thus failing to meet the requirements of full automation. Summary of the Invention

[0005] In view of this, the main objective of the present invention is to provide a railway freight car cleaning system and method that can solve the problems of high safety risks of manual operation, insufficient automated cleaning, lack of self-inspection and triggering of secondary cleaning mechanisms after cleaning in the prior art, and inability to meet the requirements of full automation.

[0006] To achieve the above objectives, the technical solution of the present invention is implemented as follows: The railway freight car cleaning system includes: a gantry frame spanning the track and capable of reciprocating along the track's length to sequentially pass through different sections of the carriage; a cleaning device mounted on the gantry frame for cleaning the side walls and roof of the carriage; a first image acquisition unit mounted on the gantry frame for real-time image acquisition of the side walls of the carriage; and a controller mounted on the gantry frame, configured to perform the following steps: based on the images acquired by the first image acquisition unit, using the instantaneous projection position of the cleaning device in the image as a dynamic reference, dynamically dividing the image area into uncleaned and cleaned areas; calculating the unit area ratio coefficient of residual coal in the uncleaned area; adjusting the cleaning parameters of the cleaning device and / or the moving speed of the gantry frame in real-time according to the unit area ratio coefficient; and detecting residual coal in the cleaned area, triggering secondary cleaning if the residual amount exceeds a preset threshold.

[0007] Preferably, the second image acquisition unit is fixedly installed on an independent bracket above the portal frame and is used to acquire images of the top of the carriage. Preferably, the controller is further configured to: identify the type and size of the current carriage by edge extraction and matching with a standard carriage size template based on the image of the carriage top.

[0008] Preferably, the controller is further configured to: fuse the residual coal information on the side walls of the carriage collected by the first image acquisition unit and the residual coal information on the top of the carriage collected by the second image acquisition unit to generate a multi-dimensional cleaning operation grid map; Preferably, the grid map is a 2.5-dimensional top-down grid, and each grid cell contains the following attributes: horizontal coordinate X of the working plane, vertical coordinate Y, height Z of the highest point of remaining coal, and remaining coal quantity level; Preferably, the controller calculates the unit area ratio coefficient of residual coal in the uncleaned area based on the grid map, and adjusts the cleaning parameters of the cleaning device and / or the moving speed of the gantry frame according to the thickness level and spatial distribution of residual coal in the grid map.

[0009] Preferably, the cleaning device is a dry ice cleaner, the cleaning parameters include the airflow of the dry ice cleaner, and the controller adjusts according to the unit area ratio coefficient according to the following rules: Preferably, when the unit area ratio coefficient is greater than a first preset threshold, the airflow is increased and / or the moving speed of the portal frame is reduced; Preferably, when the unit area ratio coefficient is less than a second preset threshold, the airflow is reduced and / or the moving speed of the portal frame is increased; Preferably, the second preset threshold is less than or equal to the first preset threshold.

[0010] Preferably, the preset threshold for triggering secondary cleanup includes: Preferably, the total area of ​​remaining coal blocks in the cleaned area exceeds a preset area percentage threshold; Preferably, there are residual coal pieces in the cleaned area whose area exceeds a preset single-piece area threshold.

[0011] Preferably, when the controller identifies the type of carriage, it performs the following steps: grayscale processing of the image of the carriage top, and extraction of the outer edge contour and inner edge contour using the Canny operator; Preferably, the perimeters of the outer edge contour and the inner edge contour of the top of the carriage are calculated to form a four-dimensional vector; Preferably, the four-dimensional vector is compared with the pre-stored standard carriage size vector using cosine similarity, and the one with the highest similarity is selected as the recognition result.

[0012] Preferably, the image acquired by the second image acquisition unit is preprocessed with grayscale threshold segmentation and morphological filtering to remove interference from the coal pile on edge extraction; Preferably, the edges of the carriage walls are extracted using Canny edge detection and least squares line fitting. Preferably, the location of the remaining coal identified by the second image acquisition unit is mapped to the joint coordinate system of the cleaning device through coordinate calibration transformation, thereby achieving a three-level coordinate calibration transformation.

[0013] Preferably, the gantry frame has an inverted U-shaped structure, and its interior forms a cleaning channel to accommodate the carriage and the track; Preferably, the cleaning device includes a side dry ice cleaning machine fixedly installed on the two side columns of the portal frame and a dry ice cleaning machine fixedly installed on the top crossbeam of the portal frame, which are respectively used to spray dry ice to clean the two side end walls of the carriage and the top of the carriage. Preferably, there are two first image acquisition units, which are respectively installed on the inner side of the two side columns of the portal frame, with the lenses facing the two side end walls of the carriage. The second image acquisition unit is mounted on the top crossbeam of the portal frame via the independent bracket, with the lens pointing vertically downwards at the top of the carriage. Preferably, the controller is an edge computing platform, which is deployed on the gantry frame and is signal-connected to the first image acquisition unit, the second image acquisition unit, and the cleaning device.

[0014] Preferably, it includes a lidar, which is fixedly installed on the ground to detect the position of the carriage; Preferably, the controller is connected to the lidar signal and determines whether the carriage has reached the preset positioning point based on the detection signal of the lidar. If so, it outputs a stop command to stop the carriage.

[0015] A method for cleaning railway freight cars includes the following steps: S1: Obtain the position information of the carriage; when the carriage reaches the preset positioning point, control the carriage to stop, and use the stopping position as the absolute zero point of the cleaning operation. S2: Use the second image acquisition unit to acquire images of the top of the carriage, identify the type and size of the carriage, and plan a cleaning path based on the identification results; S3: Combine the images of the two side walls of the carriage acquired by the first image acquisition unit and the image of the top of the carriage acquired by the second image acquisition unit to generate a cleaning operation grid map; S4: Move the gantry frame back and forth along the length of the track, so that the cleaning device passes through different sections of the carriage in sequence; S5: Based on the image acquired by the first image acquisition unit, the instantaneous projection position of the cleaning device in the image is used as a dynamic reference to dynamically divide the image area into an uncleaned area and a cleaned area. S6: Calculate the unit area ratio coefficient of residual coal in the uncleaned area; S7: Adjust the cleaning parameters of the cleaning device and / or the moving speed of the gantry frame in real time according to the unit area ratio coefficient; S8: Detect residual coal in the cleaned area. If the residual amount exceeds a preset threshold, trigger a second cleaning.

[0016] The railway freight car cleaning system and method of the present invention have the following beneficial effects: This railway freight car cleaning system and method utilizes a cleaning device and image acquisition unit mounted on a gantry frame, combined with a controller, to execute the following steps: dynamically dividing the cleaned and uncleaned areas, using the instantaneous projection of the cleaning device in the image as a reference to achieve automatic boundary tracking as the frame moves. It calculates the residual coal percentage coefficient per unit area and adjusts the dry ice airflow and frame movement speed in real time accordingly, performing intensive cleaning and slow movement in thick coal areas, and fast movement and energy saving in thin coal areas. It detects residue in cleaned areas, automatically triggering secondary cleaning when residue exceeds the standard, forming a self-inspection and secondary cleaning triggering mechanism. Furthermore, the system integrates data from top and side cameras to generate a raster map, providing information on residual coal thickness and spatial distribution, and achieving automatic identification and precise positioning of multiple freight car models. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of a railway freight car cleaning system according to the present disclosure; Figure 2 This is a structural schematic diagram of a railway freight car cleaning system according to this disclosure from another perspective; Figure 3 This is a cleaning control flowchart of a railway freight car cleaning method according to the present disclosure; Figure 4 This is a schematic diagram of a grid map for a railway freight car cleaning method according to the present disclosure; Figure 5 This is a flowchart illustrating the carriage type identification method for a railway freight car cleaning method according to this disclosure; Figure 6 A flowchart illustrating the raster map generation and three-level coordinate calibration process for a railway freight car cleaning method according to this disclosure; Figure 7 This is a flowchart illustrating the secondary cleaning decision-making process of a railway freight car cleaning method according to this disclosure; Figure 8 This is a flowchart illustrating an embodiment of a railway freight car cleaning method according to the present disclosure.

[0019] [Explanation of Key Component Symbols] 1. Gantry frame; 2. Cleaning device; 3. First image acquisition unit; 4. Controller; 5. Second image acquisition unit; 6. Independent support bracket; 7. LiDAR; 01. Carriage; 011. Side walls of the carriage; 012. Top of the carriage.

[0020] 001, Outer edge contour; 002, Inner edge contour. Detailed Implementation

[0021] The railway freight car cleaning system and method of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0022] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0023] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0024] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0025] For ease of description, spatial relative terms such as "above," "on top of," "on the upper surface of," "above," etc., are used herein to describe the spatial positional relationship of a device or feature as shown in the figures to other devices or features. It should be understood that spatial relative terms are intended to encompass different orientations in use or operation beyond the orientation of the device as described in the figures. For example, if the device in the figures were inverted, a device described as "above" or "on top of" other devices or structures would subsequently be positioned as "below" or "under" other devices or structures. Thus, the exemplary term "above" can include both "above" and "below." The device may also be positioned in other different ways (rotated 90 degrees or in other orientations), and the spatial relative descriptions used herein will be interpreted accordingly.

[0026] according to Figures 1-8 As shown, the railway freight car cleaning system includes: a gantry frame 1, which has an inverted U-shaped structure and spans across the track. Its interior forms a cleaning channel that accommodates the carriage 01 and the track. The bottom of the gantry frame 1 is equipped with pulleys driven by an electric motor, which can move back and forth along the length of the track to pass through different sections of the carriage 01 in sequence.

[0027] Cleaning device 2, installed on the portal frame 1, is used to clean the side walls 011 and the top 012 of the car body. Cleaning device 2 includes a side dry ice cleaner fixedly installed on the side columns of the portal frame 1 and a dry ice cleaner fixedly installed on the top crossbeam of the portal frame 1. To maintain effective cleaning capacity, the distance between the nozzle of the dry ice cleaner and the car body wall is controlled between 120mm and 250mm to ensure sufficient impact force. It is used to spray dry ice onto the side walls 011 and the top 012 of the car body. It should be noted that the car body 01 is empty of coal during cleaning. Figure 2 As shown, dry ice spraying for cleaning the top 012 of the carriage also includes spraying for cleaning the interior walls of the carriage 01.

[0028] The first image acquisition unit 3 is mounted on the portal frame 1. Specifically, the first image acquisition unit 3 consists of two wide-angle cameras, respectively mounted on the inner sides of the two side columns of the portal frame 1, with their lenses facing the side end walls 011 of the carriage, for real-time acquisition of images of the side end walls 011 of the carriage. The second image acquisition unit 5 is mounted above the top crossbeam of the portal frame 1 via an independent bracket 6, with its lens pointing vertically downwards at the top 012 of the carriage, for real-time acquisition of images of the top 012 of the carriage. The controller 4 communicates with the first image acquisition unit 3 and the second image acquisition unit 5 using the GMSL2 protocol. This protocol supports long-distance, high-bandwidth, and low-latency video stream transmission, ensuring the real-time performance and integrity of the image data.

[0029] Controller 4 is an edge computing platform, utilizing an NVIDIA Jetson AGX Orin 32GB / 64GB module with a computing power of 200 TOPS. The edge computing platform is dynamically deployed on the gantry frame 1 and connected to the first image acquisition unit 3, the second image acquisition unit 5, and the cleaning device 2 via signal connections. Specifically, controller 4 is mounted on the gantry frame 1 and connected via wired communication cables, such as Ethernet or USB cables, to the solenoid valves and motor drivers of the first image acquisition unit 3 and the cleaning device 2, as well as to the second image acquisition unit 5 and the LiDAR 7. Controller 4 is also connected to the driver of the motor that moves the gantry frame 1 to control its speed and direction.

[0030] Controller 4 is configured to perform the following steps: The controller 4 receives real-time image data of the two end walls 011 of the carriage from the first image acquisition unit 3. Since the cleaning device 2 is fixedly installed on the gantry frame 1, its projection position in the image is also fixed. The controller 4 uses the pixel column corresponding to this fixed projection position as a dynamic baseline. As the gantry frame 1 moves from the front to the rear of the carriage 01, the baseline divides the image into two parts: the area in front of the moving direction is the uncleaned area A, and the area behind is the cleaned area B. This dividing line is updated in real-time as the gantry frame 1 moves.

[0031] Controller 4 utilizes a YOLOv5 deep learning model deployed on an edge computing platform. This model is pre-trained using a labeled dataset of residual coal images to perform semantic segmentation on the image of the uncleaned area A, identifying the residual coal pixel region. Controller 4 dynamically determines a calculation window length along the movement direction based on the real-time travel speed of the gantry frame 1. For example, when the vehicle speed is 0.1 m / s, the window length is 0.1 m. Using this window length and the height of the carriage 01 as dimensions, a rectangular sub-region is extracted from the front end of area A. Then, the area ratio occupied by residual coal pixels within this sub-region is calculated to obtain a unit area ratio coefficient of 30%. Of course, the window length increases with increasing vehicle speed and decreases with decreasing vehicle speed to ensure that comparable region samples can be collected at different speeds.

[0032] Based on the unit area ratio coefficient, controller 4 sends analog signals, such as 4-20mA or PWM signals, to the flow regulating valve of cleaning device 2 through a PID control algorithm to increase or decrease the dry ice air flow of cleaning device 2. At the same time, it sends frequency signals to the driver of the motor of gantry frame 1 to increase or decrease the moving speed of gantry frame 1. Both can be adjusted selectively or simultaneously.

[0033] After the gantry frame 1 has traveled a certain distance, the image of the cleaned area B is analyzed again by the controller 4. The controller 4 also uses the YOLOv5 model to detect the total area ratio or the area of ​​a single coal block remaining in area B. If the detected value exceeds a preset threshold, such as a total area ratio > 5% or a single block area > 10cm², the controller 4 automatically triggers a secondary cleaning process: controlling the gantry frame 1 to move in the opposite direction a preset distance, and then re-executes the above cleaning steps. In this way, the railway freight car cleaning system completely replaces manual labor, eliminates major safety hazards such as electric shock and falls from heights, achieves automatic tracking of the cleaning boundary, avoids missed cleaning and repeated cleaning, and improves work efficiency. The airflow and moving speed are adaptively adjusted according to the actual coverage of residual coal, with powerful cleaning in thick coal areas and energy-saving and fast cleaning in thin coal areas, saving dry ice and electricity while ensuring cleanliness, and realizing unmanned automatic cleaning.

[0034] In order to realize the carriage 01 type recognition function and adapt to the external dimensions of different models of freight car carriages 01, so that the cleaning path of the cleaning device 2 can be automatically adapted, the railway freight car cleaning system also includes: a second image acquisition unit 5, which is fixedly installed on an independent bracket 6 above the gantry frame 1, for acquiring images of the carriage top 012. The controller is also configured to: based on the carriage top 012 image, identify the type and size of the current carriage 01 by edge extraction and matching with a standard carriage size template.

[0035] The steps for identifying carriage 01 type are further explained below. After carriage 01 is located by lidar 7 and stops at the preset position, controller 4 triggers the second image acquisition unit 5 to capture a complete image of the carriage top 012. In this system, lidar 7 is a long-range three-dimensional lidar specifically used in the rail transit industry, preferably the Horn-RT series, which has a repeatability of ≤5cm (1σ), accuracy of ≤±5cm, measurement frame rate of ≥10Hz, protection level of IP67, and operating temperature range of -40℃ to 85℃. The preset positioning point is set at the center of the outer surface of the front baffle of carriage 01, and a corner reflector is fixedly installed here as the measurement reference.

[0036] Controller 4 performs grayscale processing on the image of the top of the carriage (012) and extracts the outer edge contour (001) and inner edge contour (002) using the Canny operator. By calculating the perimeters of the outer edge contour (001) and inner edge contour (002), a four-dimensional vector a, b, c, d is obtained. This four-dimensional vector is then compared with a pre-stored database of standard carriage dimensions within the controller using cosine similarity, and the vector with the highest similarity is selected as the current carriage type (01). Based on the identified carriage type (01), controller 4 automatically calls the pre-stored cleaning parameter configuration for that vehicle model, which includes the moving speed of the gantry frame 1.

[0037] It should be noted that this standard freight car size database was pre-built to enable automatic identification of various railway freight car models. The database is compiled based on the vehicle model coding system specified in TB / T3443.3-2016 "Rollercraft and Rolling Stock Type, Model and Number Coding Rules Part 3: Freight Cars", combined with the vehicle clearance dimensions specified in GB146.1 "Standard Gauge Railway Clearance", and the vehicle model technical specifications provided by major freight car manufacturers. Each record in the database contains fields such as vehicle model code, outer contour length, outer contour width, inner contour length, inner contour width, side wall height, and load capacity. The database is stored in SQLite3 format in the 64GB eMMC built-in memory of controller 4, i.e., the NVIDIA Jetson AGX Orin edge computing platform.

[0038] The grayscale processing is further explained below: The controller 4 converts the RGB image acquired by the second image acquisition unit 5 into a grayscale image I_gray, and uses median filtering to remove salt and pepper noise.

[0039] The following is a further explanation of Canny edge extraction: Set the high and low thresholds of the Canny operator, where the high threshold is set to 120 and the low threshold is set to 50. Use these parameters to extract all edges in the image.

[0040] The following further explains the distinction between inner and outer contours: Since the top of the carriage 012 is a rectangular ring, the outer rectangle is the outer edge of carriage 01, and the inner rectangle is the internal opening of carriage 01. After edge detection of the outer edge contour 001 and the inner edge contour 002, two concentric rectangular contours will appear. The controller 4 calculates the area and perimeter of each connected region, identifying the one with the larger perimeter as the outer edge contour 001 and the one with the smaller perimeter as the inner edge contour 002.

[0041] The perimeter calculation is further explained below: The Euclidean distances between all adjacent pixels on the outer edge contour 001 and the inner edge contour 002 are accumulated to obtain the perimeter value. Then, the pixel equivalent (mm / pixel) calibrated by the second image acquisition unit 5 is converted into an actual millimeter value, forming a four-dimensional vector: outer length a, outer width b, inner length c, and inner width d.

[0042] The following provides a set of standard carriage size data examples and a cosine similarity calculation embodiment to illustrate the specific steps in this embodiment of calculating the cosine similarity between the four-dimensional vector and the pre-stored standard carriage size vector, and selecting the largest sim value as the recognition result.

[0043] 1. Specific examples of standard carriage dimensions database:

[0044] 2. To facilitate understanding of the matching principle of cosine similarity, the following is an example of cosine similarity calculation: Let the four-dimensional vector to be identified extracted by the second image acquisition unit 5 from the image of the top of a carriage be: Q=[13950,3230,12480,2885], meaning the measured outer contour length is 13950mm, the outer contour width is 3230mm, the inner contour length is 12480mm, and the inner contour width is 2885mm. Controller 4 calculates the cosine similarity between this vector and the pre-stored standard vectors of the C70, C80, and C64 models in the database:

[0045] Since the similarity score of C64 (0.9449) is higher than that of C70 (0.9052) and C80 (0.8967), the system determines that the current carriage type is C64 general-purpose open wagon.

[0046] Through the recognition method of the second image acquisition unit 5, automatic identification of various railway freight car models is achieved, improving the system's intelligence and adaptability. Cleaning parameters are automatically matched according to the car model, ensuring that every surface to be cleaned on different sized carriages 01 is covered, avoiding missed areas or interference due to size mismatch. This provides an accurate size reference for coordinate transformation of the raster map. Figure 4 As shown, in the grid map of this specific embodiment, different colors represent different remaining coal thicknesses.

[0047] To avoid a mismatch between the cleaning intensity and the actual state of the remaining coal due to inaccurate description of the thickness and spatial location of the remaining coal in the images, the controller is also used to: fuse the remaining coal information (011) on both sides of the carriage acquired by the first image acquisition unit 3 and the remaining coal information (012) on the top of the carriage acquired by the second image acquisition unit 5. The remaining coal information includes: the area of ​​the remaining coal and its distribution region. A multi-dimensional cleaning operation grid map is generated. The grid map is a 2.5-dimensional top-view grid, and each grid contains the following attributes: the horizontal coordinate (X) and vertical coordinate (Y) of the working plane, the height of the highest point of the remaining coal (Z), and the remaining coal quantity level.

[0048] The following is a further explanation of generating a multidimensional cleaning operation grid map. Since the top 012 of the carriage is often covered with a large amount of residual coal, direct edge detection would extract a large number of false edges. Therefore, this system uses a combination algorithm of grayscale threshold segmentation and morphological filtering for preprocessing: the specific steps are as follows: the controller 4 converts the RGB color image acquired by the second image acquisition unit 5 into a grayscale image. Controller 4 adaptively calculates the segmentation threshold using the Otsu method. The algorithm iterates through gray levels 0-255 to find the gray value with the largest inter-class variance as the optimal threshold. Then, it compares the gray value of each pixel in the grayscale image with the segmentation threshold. Comparison: If the pixel value is greater than the segmentation threshold Treating it as the carriage wall, if the pixel value is less than the segmentation threshold The image is treated as a coal pile, thus obtaining a binary image, which initially distinguishes the coal pile from the carriage wall.

[0049] Then, morphological opening operations are used to eliminate isolated noise points, as isolated noise points may exist in the binarized image due to coal dust, reflections, etc. Controller 4 performs an opening operation on the image to remove these noise points while maintaining the overall shape of the carriage wall. The opening operation formula is:

[0050] Where B is a 3×3 square structural element. Indicates corrosion operation. The expansion operation indicates that corrosion can eliminate isolated black coal pile noise spots, and expansion can restore the integrity of the car body wall area.

[0051] After preprocessing, controller 4 uses the Canny edge detection algorithm to extract the edge point set of the carriage wall. Then, it uses the least squares method to fit these edge points into straight lines to compensate for edge curvature caused by carriage deformation or camera distortion. Specifically, Canny edge detection sets high and low thresholds to extract all strong and weak edges and connects them to form a continuous edge contour. The fitting formula is as follows: The slope is The intercept is , denoted by , where n is the number of valid pixels. The fitted straight line represents the precise edge of the carriage wall, used to subsequently determine the perimeter and dimensions of the carriage's inner and outer contours.

[0052] It should be noted that controller 4 integrates the residual coal information (011) on both side walls of the carriage acquired by the first image acquisition unit 3 and the residual coal information (012) on the top of the carriage acquired by the second image acquisition unit 5 to generate a 2.5D top-view grid map. This grid map takes into account both planar positioning and elevation attributes, but does not include the side wall height coordinates, focusing on the cleaning operation plane and the thickness of the residual coal. Each grid cell contains the following attributes: X-coordinate: Horizontal axis coordinate of the working plane, unit: mm; Y-coordinate: Vertical axis coordinate of the working plane, unit: mm; Z-value: The height of the highest point of remaining coal within this grid relative to the floor of the car body, unit: mm. Remaining coal quantity is defined as levels 1-5: Level 1: No remaining coal; Level 2: Small amount of thin coal; Level 3: Medium thickness; Level 4: Relatively thick; Level 5: Large accumulation. For example, a grid with X=1000, Y=500, Z=15, and remaining coal quantity level=2 indicates that at coordinates 1000mm, 500mm on the working plane, the highest height of remaining coal is 15mm, which is considered a small amount of thin coal.

[0053] The controller 4 maps the remaining coal position identified by the second image acquisition unit 5 to the joint coordinate system of the side cleaning device 2 through a three-level calibration transformation. The specific steps are as follows: In order to achieve accurate coordinate mapping of the remaining coal position identified by the second image acquisition unit to the nozzle of the cleaning device, the system completes three-level joint calibration of camera intrinsic parameters, hand-eye calibration, and actuator calibration during the equipment leaving the factory and on-site deployment stages, and builds a complete coordinate transformation link.

[0054] (I) Calibration Basic Configuration and Matrix Definition This system adopts an eye-to-hand architecture, with the first and second image acquisition units rigidly fixed to the gantry frame and moving synchronously with it. The overall coordinate transformation is divided into three levels: pixel coordinate system → camera coordinate system → gantry frame global coordinate system → cleaning device joint coordinate system, and the mapping is completed entirely by three sets of calibration matrices.

[0055] 1. Camera intrinsic parameter matrix The camera intrinsic parameter distortion calibration is completed in advance to obtain the camera intrinsic parameter matrix K and radial / tangential distortion coefficients, which are used to solve the pixel coordinates to camera 3D coordinates. The expression of the intrinsic parameter matrix is ​​as follows: In the formula: fx and fy are the horizontal and vertical focal lengths of the camera; u0 and v0 are the pixel coordinates of the principal point of the image.

[0056] 2. Hand-eye transformation matrix The goal of hand-eye calibration is to solve for the homogeneous transformation matrix from the camera coordinate system to the global coordinate system of the portal frame. Matrix form: In the formula: R is a 3×3 rotation matrix, and t is a 3×1 translation vector.

[0057] 3. Actuator Transformation Matrix The actuator calibration solves for the transformation matrix from the global coordinate system of the gantry frame to the joint coordinate system of the dry ice washer nozzle. This compensates for installation deviations and structural errors in the nozzles and mechanical supports.

[0058] (II) Complete Hand-Eye Calibration Process The hand-eye model of this system satisfies the classical equation AX = XB, where X is the hand-eye matrix to be determined. A is the relative transformation matrix between different poses of the gantry frame, calculated from the real-time displacement data of the travel encoder; B is the relative transformation matrix of the calibration plate in different poses in the camera coordinate system, obtained by calculating the camera extrinsic parameters from the image. The specific calibration steps are as follows: 1. Calibration fixture layout: Use a 9×6 checkerboard calibration plate with a single square side length of 20mm. Place the calibration plate within the working range of the vehicle frame to ensure that the field of view of all cameras completely covers the calibration plate area.

[0059] 2. Multi-pose data synchronous acquisition: The control gantry frame moves sequentially to 15 different working poses, which are evenly distributed throughout the frame's entire stroke range. Two data items are acquired synchronously in each pose: first, the encoder values ​​are read to calculate the relative transformation matrix Ai of the frame in adjacent poses; second, the camera captures an image of the calibration board.

[0060] 3. Image preprocessing and corner extraction: The acquired images are converted to grayscale and filtered for noise reduction. The OpenCV findChessboardCorners function is called to detect chessboard corners, and the cornerSubPix function is used to perform subpixel-level corner optimization to improve positioning accuracy. Invalid images with blurriness, overexposure, or missing corners are removed.

[0061] 4. Matrix Solving: Based on the effective image data, calculate the extrinsic parameters of the calibration board in the camera coordinate system, and calculate the relative transformation matrix Bi between adjacent pose calibration boards; use the Tsai-Lenz algorithm to solve the hand-eye equation, and output the final hand-eye matrix. Furthermore, the rotation matrix is ​​orthogonalized to eliminate numerical calculation errors.

[0062] (III) Calculation Logic of Three-Level Coordinate Transformation The first stage transforms the pixel coordinates to the camera coordinate system of the second image acquisition unit 5, using the pre-calibrated intrinsic parameter calibration matrix K of the second image acquisition unit 5 to convert the pixel coordinates... Convert to 3D coordinates in camera coordinate system The conversion formula is: Among them, the intrinsic parameter calibration matrix , For camera focal length, The pixel coordinates of the image center point. The known fixed distance from the optical center of the camera of the second image acquisition unit 5 to the plane of the carriage.

[0063] After calibration, reprojection error verification is performed using verification data that was not used in calibration. The root mean square error should be less than 0.5 pixels.

[0064] The second stage transforms the camera coordinate system to the global coordinate system of the portal frame 1. Through hand-eye calibration, the transformation relationship between the camera and the portal frame 1 is obtained, including the rotation matrix R and the translation vector t, forming a homogeneous transformation moment. Where R is a 3×3 rotation matrix and t is a 3×1 translation vector. The position of the remaining coal in the camera coordinate system of the second image acquisition unit 5 is transformed to the gantry global coordinate system. .

[0065] The third stage transforms the global coordinate system of the gantry frame 1 to the joint coordinate system of the cleaning device 2. Based on the kinematic model of the side cleaning device 2, the joint coordinate system is established. Through the actuator calibration matrix, the global coordinates are mapped to the pointing point of the dry ice cleaner nozzle, ensuring that the residual coal position identified by the second image acquisition unit 5 can be accurately transmitted to the side cleaning device 6, achieving targeted cleaning.

[0066] Based on the grid map, controller 4 calculates the weighted average of the remaining coal levels in all grids within the uncleaned area A as a unit area percentage coefficient. Simultaneously, based on the thickness level and coordinate position of each grid, controller 4 dynamically adjusts the airflow and gantry movement speed of the cleaning device 2 as it passes through that grid. For example, when the grid level is 4 (thicker), the airflow increases by 30%, and the movement speed decreases by 20%; when the grid level is 2 (thinner coal), the airflow decreases by 20%, and the movement speed increases by 15%.

[0067] To adjust the cleaning intensity based on the amount of remaining coal and avoid energy waste or incomplete cleaning, the cleaning parameters of the cleaning device 2 include the airflow of the dry ice cleaner. In this specific embodiment, the dry ice cleaner includes a dry ice storage tank, a compressed air source, a delivery hose, and nozzles. The dry ice storage tank is fixed to the upper side of the gantry frame 1, and multiple nozzles are arranged on the side of the gantry frame 1 facing the carriage 01. The cleaning range covers the two end walls 011 and the top 012 of the carriage, respectively. The controller 4 controls the mixing ratio and flow rate of dry ice and air through a solenoid valve and a proportional regulating valve. The controller adjusts according to the following rules based on the unit area ratio coefficient: the controller 4 internally sets a first preset threshold. In this specific embodiment, the unit area ratio coefficient is >30% and the second preset threshold is <10%, and the second preset threshold is ≤ the first preset threshold.

[0068] When the unit area ratio coefficient is greater than 30%, the controller 4 uses a PID algorithm to increase the opening of the air flow ratio regulating valve of the dry ice cleaner to 80% of the rated value, while reducing the moving speed of the gantry frame 1 from the rated speed of 0.2m / s to 0.1m / s, a reduction of 50%.

[0069] When the unit area ratio is less than 10%, controller 4 reduces the airflow to 30% of the rated value while increasing the moving speed to 0.3 m / s, a 50% increase. When the ratio is between 10% and 30%, the airflow and moving speed remain at their default values. These adjustments can be performed individually or simultaneously. This allows for on-demand supply of dry ice to the dry ice cleaner: high flow rate and slow speed in thick coal areas to ensure cleaning power, and low flow rate and fast speed in thin coal areas to save dry ice and electricity.

[0070] To prevent residual coal from being left unremoved and to enable secondary cleaning, preset thresholds for triggering secondary cleaning include: the total area of ​​remaining coal blocks in the cleaned area exceeding a preset area percentage threshold, or the presence of a single coal block in the cleaned area exceeding a preset single coal block area threshold. Specifically, after completing one cleaning cycle, controller 4 analyzes the grid map of the cleaned area B. Condition one: calculate the total area of ​​all remaining coal blocks in the connected regions segmented by YOLOv5, divide it by the total area of ​​the cleaned area to obtain the total area percentage. If this percentage exceeds the preset area percentage threshold, it is considered unqualified. Condition two: traverse the connected regions of each remaining coal block, calculate its bounding rectangle area or pixel area. If any coal block has an area greater than the preset single coal block area threshold, it is considered unqualified. When either of these two conditions is met, controller 4 triggers secondary cleaning. The secondary cleaning is executed as follows: the gantry frame 1 moves backward a preset distance, then resumes cleaning forward, and the cleaned area is checked again until both indicators are below the threshold.

[0071] In one specific embodiment, the preset area percentage threshold is set to 5%, and the preset single-piece area threshold is set to 10 cm². When the total area of ​​remaining coal pieces in the cleaned area exceeds 5% of the total area of ​​the area, or when any single piece of coal has an area greater than 10 cm², the system determines that the residual amount exceeds the preset threshold and triggers secondary cleaning. It should be noted that the above thresholds are only examples, and those skilled in the art can adjust the above thresholds according to actual cleaning requirements such as coal type, humidity, and ambient temperature.

[0072] To avoid inconsistent cleaning starting points, which could affect grid map coordinate alignment and cleaning coverage, the system also includes a LiDAR 7, fixedly installed on the ground, used to detect the position of the carriage 01. Correspondingly, corner reflectors are fixed at specific positions on the bottom or side of the carriage 01 as the positioning reference for the LiDAR 7. The controller 4 is connected to the LiDAR 7 via a fieldbus (Profinet). Specifically, when the carriage 01 is traveling on an empty track, the LiDAR 7 continuously emits a laser beam and receives the echoes returned from the corner reflectors, measuring the distance of the carriage 01 relative to a preset positioning point in real time. The LiDAR 7 sends the measurement data to the controller 4 at a frequency of 100Hz. The controller 4 internally sets a preset positioning point. When the controller 4 determines that the error between the current distance value and the preset value is less than ±2cm, it determines that the carriage 01 has reached the preset positioning point, and then outputs a stop command to the traction system of the carriage 01 through the I / O interface, causing the carriage 01 to stop. After stopping, the controller 4 takes the current position of the carriage 01 as the absolute zero point for subsequent cleaning operations, and the origin of the grid map is also set accordingly. The grid size is determined based on the effective cleaning diameter of the dry ice nozzle.

[0073] The side cleaning device 6 in operation is a wide-width fan-shaped flat nozzle, fixed at the center of the side of the gantry. At this point, the nozzle pressure is relatively weak, suitable for situations where residual coal adhesion is low. To handle situations with more residual coal, it is necessary to switch to intelligent mode, changing the nozzle to a round nozzle to achieve higher spray pressure. In this case, the cleaning point needs to be located using the generated cleaning grid map. When the vehicle frame moves to the same plane as the cleaning position, the position of the dry ice cleaner 6 is adjusted up and down to ensure the nozzle is directly facing the area to be cleaned, achieving powerful cleaning.

[0074] A method for cleaning railway freight cars includes the following steps: S1: Obtain the position information of carriage 01. When carriage 01 reaches the preset positioning point, control carriage 01 to stop and use the stopping position as the absolute zero point of the cleaning operation. S2: Use the second image acquisition unit 5 to acquire images of the top 012 of the carriage, identify the type and size of the current carriage 01, and plan the cleaning path based on the identification results; S3: Combine the images of the two side walls 011 of the carriage acquired by the first image acquisition unit 3 and the top of the carriage 012 acquired by the second image acquisition unit 5 to generate a cleaning operation grid map. S4: Move the gantry frame 1 back and forth along the length of the track, so that the cleaning device 2 passes through different sections of the carriage 01 in sequence; S5: Based on the image acquired by the first image acquisition unit 3, the instantaneous projection position of the cleaning device 2 in the image is used as a dynamic reference to dynamically divide the image area into an uncleaned area and a cleaned area. S6: Calculate the unit area ratio coefficient of residual coal in the uncleaned area; S7: Adjust the cleaning parameters of the cleaning device 2 and / or the moving speed of the gantry frame 1 in real time according to the unit area ratio coefficient. S8: Detect residual coal in the cleaned area. If the residual amount exceeds the preset threshold, trigger a second cleaning.

[0075] The following is a complete embodiment: taking a coal train composed of multiple C64 type general open wagons as an example, the railway freight car cleaning system and method of the present invention will be fully described.

[0076] I. System Initialization and Database Loading Before starting the cleaning operation, the operator starts the system, and controller 4 executes the self-test and initialization program: 1. Hardware self-test: The controller 4 detects the communication status with the two cameras of the first image acquisition unit 3, the one camera of the second image acquisition unit 5, the lidar 7, the dry ice cleaner of the cleaning device 2, and the drive motor of the gantry frame 1, and confirms that all devices are online and in normal condition.

[0077] 2. Calibration Data Loading: Controller 4 loads the following calibration data from its local eMMC memory:

[0078] 3. Vehicle Model Database Loading: Controller 4 loads the standard vehicle body size database into memory. The database contains common vehicle models such as C70, C80, C64, P70, and NX70. Each record contains a four-dimensional vector and a pre-calculated vector magnitude vec_norm.

[0079] II. Carriage Positioning and Parking Control 1. Train enters the work area: The train, driven by the traction system, travels on the track at a speed of 0.1 m / s and enters the freight yard cleaning work area. A fixed ground-mounted lidar 7 (Honn-RT series) with a measurement frequency of 20 Hz and a repeatability of 5 cm continuously emits a laser beam to detect the train's position.

[0080] 2. Carriage Positioning Detection: The lidar 7 transmits distance measurement data to the controller 4 at a frequency of 20Hz. When the corner reflector on the front baffle of the first C70 open wagon enters the lidar's detection range, the radar echo signal strength is significantly enhanced.

[0081] 3. Stop Detection: The controller calculates the arithmetic mean of 10 consecutive frames of measurement data over 0.5 seconds and uses Kalman filtering to smooth noise. When the error between the filtered measurement value and the preset positioning point is ≤ ±2cm, the controller determines that the carriage has reached the preset positioning point and outputs a stop command through the I / O interface, stopping the carriage.

[0082] 4. Absolute Zero Point Setting: After the carriage stops, controller 4 sets this position as the absolute zero point for the cleaning operation. The X=0mm mark is set at the front baffle of the carriage, and the Y=0mm mark is set at the center line of the carriage. The origin of the grid map is set accordingly, and the grid size is set to 40mm×40mm, which is matched with half of the effective cleaning diameter of the dry ice nozzle, which is 80mm.

[0083] III. Automatic Recognition of Carriage Type 1. Top Image Acquisition: After the carriage stops, the controller 4 triggers the second image acquisition unit 5 to capture a complete image of the top 012 of the carriage, with a resolution of 1920×1080 pixels.

[0084] 2. Image preprocessing: Controller 4 performs the following processing on the original image sequentially:

[0085] 3. Edge Extraction: The Canny edge detection algorithm was used to extract the carriage outline. The Otsu adaptive thresholding method was used: For the C70 carriage top image in this example, the Otsu method calculated the following: High threshold Thigh = 145 (grayscale level), Low threshold Tlow = 145 × 0.4 = 58. Hysteresis edge connections were performed using the 8-neighborhood rule to fill in edge breaks.

[0086] 4. Contour recognition and differentiation: By calculating the area and perimeter of the connected region, the outer edge contour 001 and the inner edge contour 002 are distinguished: the outer edge contour 001 with the larger perimeter is the outer edge contour of the carriage, and the inner edge contour 002 with the smaller perimeter is the inner edge contour of the opening inside the carriage.

[0087] 5. Perimeter Calculation and Four-Dimensional Vector Generation: Calculate the perimeter of the outer edge contour 001 and the inner edge contour 002 respectively. Convert the perimeter to actual millimeter values ​​according to the pixel equivalent of 0.5mm / pixel calibrated by the camera, and obtain the four-dimensional vector: Q=[13950,3230,12480,2885], unit: mm.

[0088] 6. Cosine similarity matching: Controller 4 compares the vector to be identified, Q, with the pre-stored standard vectors in the database one by one:

[0089] Among them, the C64 model has the highest similarity, 0.9449, which falls within the medium matching range of 0.90~0.95. The controller determines that the current carriage type is a C64 general open wagon and stores the identification record in the queue to be reviewed.

[0090] 7. Cleaning parameter call: Based on the identified vehicle model C64, the controller 4 automatically calls the pre-stored cleaning parameters for this vehicle model: nominal moving speed of the gantry frame: 0.10m / s, nominal dry ice air flow rate: 60%, nominal distance between the nozzle and the truck bed wall: 200mm, cleaning path planning: the frame moves from X=0 to X=13438mm.

[0091] IV. Raster map generation: The controller 4 integrates the remaining coal information from the first image acquisition unit 3 and the second image acquisition unit 5 to generate a 2.5-dimensional raster map.

[0092] 1. Multi-channel image synchronous acquisition: Controller 4 triggers the following simultaneously through the hardware triggering function of the GMSL2 protocol: left first image acquisition unit 3L: captures the image of the left end wall of the carriage; right first image acquisition unit 3R: captures the image of the right end wall of the carriage; and second image acquisition unit 5: captures the 012 image of the top of the carriage. The time synchronization error of the three images is <1ms, ensuring the spatiotemporal consistency of the raster map fusion.

[0093] 2. YOLOv5 Semantic Segmentation: Controller 4 runs a TensorRT-accelerated YOLOv5 model on the NVIDIA Jetson AGX Orin platform to perform semantic segmentation on three images, identify coal residue areas, and achieve a model inference latency of <8ms / frame.

[0094] 3. Three-level coordinate calibration conversion: Convert the identified residual coal pixel coordinates into nozzle joint coordinates: Level 1: Pixel coordinates → Camera coordinate system; Level 2: Camera coordinate system → Portal frame global coordinate system; Level 3: Global coordinate system → Nozzle joint coordinate system; By using the actuator calibration matrix Tcamnozzle to transform, the joint coordinates of the target point that the side nozzle needs to be aligned with are obtained.

[0095] 4. Raster Map Generation: Controller 4 discretizes the remaining coal position in the global coordinate system into a 40mm×40mm grid. Each grid records: X coordinate (along the track direction, in mm); Y coordinate (perpendicular to the track direction, in mm); Z value (height of the highest remaining coal point, in mm); and remaining coal quantity level (1-5). A partial example of the generated raster map is shown below.

[0096] V. Cleaning Operations and Dynamic Parameter Adjustment 1. Cleaning begins: The gantry frame 1 starts from position X=0 and moves along the track towards the rear of the vehicle at a nominal speed of 0.10 m / s. The cleaning device 2 is activated, with the side dry ice cleaner and the top dry ice cleaner working simultaneously.

[0097] 2. Dynamic window calculation and pre-aiming control: Controller 4 monitors the moving speed of the portal frame 1 in real time and dynamically adjusts the length of the calculation window for the proportion coefficient.

[0098] Meanwhile, controller 4 introduces a pre-aiming control mechanism: it reads grid map data within the pre-aiming distance ahead and starts adjusting parameters 200ms before the chassis enters the thick coal zone.

[0099] 3. Calculation of Unit Area Proportion Coefficient: Controller 4 extracts the grid corresponding to the current cleaning area and calculates the weighted average of the remaining coal quantity levels as the unit area proportion coefficient. When entering the X=80-160mm area, there are grids with remaining coal quantity levels 4 and 5 in this area, and the calculated proportion coefficient is 65%.

[0100] 4. Cleaning parameters are adjusted in real time. Based on the proportion coefficient, controller 4 adjusts according to the following rules:

[0101] After adjustment, the dry ice cleaning machine uses a higher flow rate and stronger impact to perform a strong wall-hugging cleaning of the thick coal area, while the vehicle frame slows down to ensure sufficient cleaning time.

[0102] 5. Rapid passage through thin coal areas: When the gantry frame 1 travels to the cleaned area X=0-80mm, the controller 4 detects that there is only a small amount of thin coal with a proportion coefficient of only 5%, and immediately reduces the air flow to 30% and increases the frame speed to 0.25m / s, realizing long-distance fast sweeping and saving dry ice and electricity.

[0103] VI. Post-cleaning self-inspection and secondary cleaning: 1. Residual Detection in Cleaned Areas: After the gantry frame 1 passes through a certain section, the controller 4 analyzes the grid map of the cleaned area B to detect residual coal. Condition 1 is the percentage of total area: the total area of ​​residual coal blocks in all connected domains is calculated as the percentage of the total cleaned area. For example, in a 10mm × 10mm area, the percentage of residual coal blocks detected is 6%. Condition 2 is the area of ​​a single piece: the connected domains of each residual coal block are traversed, and the area of ​​its circumscribed rectangle is calculated. A residual coal block with an area of ​​12cm² is detected.

[0104] 2. Determination of exceeding the standard and triggering of secondary cleaning: Preset thresholds: (1) Total area percentage threshold: 5%, (2) Single area threshold: 10cm². Since 6%>5% in condition one and 12cm²>10cm² in condition two both exceed the standard, controller 4 determines that the residual amount exceeds the preset threshold and automatically triggers secondary cleaning.

[0105] 3. Secondary cleaning execution: Controller 4 executes the following secondary cleaning process: controls the gantry frame 1 to move 500mm in the opposite direction, retreats to the front of the excessive area, temporarily increases the dry ice airflow to 90%, further decelerates the frame to a speed of 0.04m / s, and re-cleans the excessive area forward.

[0106] 4. Re-inspection after secondary cleaning: After the secondary cleaning is completed, controller 4 performs a residue detection on the same area again: when the total area percentage drops to 1.2% (<5%); when the largest single area drops to 3cm² (<10cm²), both indicators are below the threshold, the cleaning is deemed qualified, and the gantry frame 1 continues to move forward.

[0107] VII. Continuous Multi-Carriage Operation: After cleaning the first carriage, controller 4 records the cleaning completion status and outputs a traction start command, causing the train to move forward one carriage distance. LiDAR 7 then detects the position of the second carriage again, repeating steps S1 to S6 to achieve continuous unmanned automatic cleaning of multiple carriages.

[0108] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention.

Claims

1. A railway freight car cleaning system, characterized in that, include: The gantry frame (1) spans across the track and can move back and forth along the length of the track to pass through different sections of the carriage (01) in sequence; A cleaning device (2) is installed on the gantry frame (1) and is used to clean the side walls (011) and the top (012) of the carriage. The first image acquisition unit (3) is installed on the portal frame (1) and is used to acquire images of the two side walls (011) of the carriage in real time. The controller (4), mounted on the portal frame (1), is configured to perform the following steps: Based on the image acquired by the first image acquisition unit (3), the instantaneous projection position of the cleaning device (2) in the image is used as a dynamic reference to dynamically divide the image area into an uncleaned area and a cleaned area. Calculate the unit area ratio coefficient of residual coal in the uncleaned area; Based on the unit area ratio coefficient, the cleaning parameters of the cleaning device (2) and / or the moving speed of the gantry frame (1) are adjusted in real time. The cleaned area is tested for residual coal. If the residual amount exceeds a preset threshold, a second cleaning is triggered.

2. The railway freight car cleaning system according to claim 1, characterized in that, Also includes: The second image acquisition unit (5) is fixedly installed on the independent bracket (6) above the portal frame (1) and is used to acquire images of the top (012) of the carriage. The controller is also configured to identify the type and size of the current carriage (01) by edge extraction and matching with a standard carriage size template based on the image of the carriage top (012).

3. The railway freight car cleaning system according to claim 2, characterized in that, The controller is also used to: fuse the coal remaining information on the two side walls (011) of the carriage collected by the first image acquisition unit (3) and the coal remaining information on the top of the carriage (012) collected by the second image acquisition unit (5) to generate a multi-dimensional cleaning operation grid map; The grid map is a 2.5-dimensional top-down grid, and each grid cell contains the following attributes: x-coordinate of the working plane, y-coordinate of the working plane, height of the highest point of remaining coal, and level of remaining coal quantity. The controller calculates the unit area ratio coefficient of residual coal in the uncleaned area based on the grid map, and adjusts the cleaning parameters of the cleaning device (2) and / or the moving speed of the gantry frame (1) according to the thickness level and spatial distribution of residual coal in the grid map.

4. The railway freight car cleaning system according to claim 3, characterized in that, The cleaning device (2) is a dry ice cleaner, and the cleaning parameters include the airflow of the dry ice cleaner. The controller adjusts the airflow according to the unit area ratio coefficient according to the following rules: When the unit area ratio coefficient is greater than the first preset threshold, increase the airflow and / or reduce the moving speed of the portal frame (1); When the unit area ratio coefficient is less than the second preset threshold, reduce the airflow and / or increase the moving speed of the portal frame (1); The second preset threshold is less than or equal to the first preset threshold.

5. The railway freight car cleaning system according to claim 4, characterized in that, The preset thresholds for triggering secondary cleanup include: The total area of ​​remaining coal blocks in the cleaned area exceeds a preset area percentage threshold. Or, there may be residual coal in the cleaned area whose area exceeds a preset single-piece area threshold.

6. The railway freight car cleaning system according to claim 5, characterized in that, When the controller identifies the type of the carriage (01), it performs the following steps: grayscale processing of the image of the top (012) of the carriage, and extracting the outer edge contour (001) and inner edge contour (002) using the Canny operator; Calculate the perimeter of the outer edge contour (001) and the inner edge contour (002) of the top of the carriage (012) to form a four-dimensional vector; The four-dimensional vector is compared with the pre-stored standard carriage size vector using cosine similarity, and the one with the highest similarity is selected as the recognition result.

7. The railway freight car cleaning system according to claim 6, characterized in that, The image acquired by the second image acquisition unit (5) is preprocessed with grayscale threshold segmentation and morphological filtering to remove the interference of coal pile on edge extraction; The edges of the carriage walls were extracted using Canny edge detection and least squares line fitting. The remaining coal position identified by the second image acquisition unit (5) is mapped to the joint coordinate system of the cleaning device (2) through coordinate calibration transformation, thereby realizing the three-level coordinate calibration transformation.

8. The railway freight car cleaning system according to claim 7, characterized in that, The gantry frame (1) has an inverted U-shaped structure, and its interior forms a cleaning channel to accommodate the carriage (01) and the track; The cleaning device (2) includes a side dry ice cleaning machine fixedly installed on the two side columns of the portal frame (1) and a dry ice cleaning machine fixedly installed on the top crossbeam of the portal frame (1), which are used to spray dry ice to clean the two side end walls (011) of the carriage and the top (012) of the carriage, respectively. The first image acquisition unit (3) consists of two units, which are respectively installed on the inner side of the two side columns of the portal frame (1), with the lens facing the two side end walls (011) of the carriage. The second image acquisition unit (5) is mounted above the top crossbeam of the portal frame (1) via the independent bracket (6), with the lens pointing vertically downward toward the top of the carriage (012). The controller (4) is an edge computing platform, which is deployed on the gantry frame (1) and is connected to the first image acquisition unit (3), the second image acquisition unit (5) and the cleaning device (2).

9. The railway freight car cleaning system according to claim 8, characterized in that, It also includes a lidar (7), which is fixedly installed on the ground and used to detect the position of the carriage (01); The controller (4) is connected to the laser radar (7) and determines whether the carriage (01) has reached the preset positioning point based on the detection signal of the laser radar (7). If so, it outputs a stop command to stop the carriage (01).

10. A method for cleaning railway freight cars based on the railway freight car cleaning system according to any one of claims 2 to 9, characterized in that, Includes the following steps: S1: Obtain the position information of the carriage (01). When the carriage (01) reaches the preset positioning point, control the carriage (01) to stop and use the stopping position as the absolute zero point of the cleaning operation. S2: Use the second image acquisition unit (5) to acquire an image of the top of the carriage (012), identify the type and size of the current carriage (01), and plan the cleaning path according to the identification result; S3: Combine the images of the two side walls (011) of the carriage acquired by the first image acquisition unit (3) and the image of the top of the carriage (012) acquired by the second image acquisition unit (5) to generate a cleaning operation grid map; S4: Move the gantry frame (1) back and forth along the length of the track, so that the cleaning device (2) passes through different sections of the carriage (01) in sequence; S5: Based on the image acquired by the first image acquisition unit (3), the instantaneous projection position of the cleaning device (2) in the image is used as a dynamic reference to dynamically divide the image area into an uncleaned area and a cleaned area; S6: Calculate the unit area ratio coefficient of residual coal in the uncleaned area; S7: Adjust the cleaning parameters of the cleaning device (2) and / or the moving speed of the gantry frame (1) in real time according to the unit area ratio coefficient. S8: Detect residual coal in the cleaned area. If the residual amount exceeds a preset threshold, trigger a second cleaning.

Images