Locomotive outer surface cleaning method and system based on multi-mode perception and intelligent decision

Abstract

The invention discloses a locomotive outer surface cleaning method and system based on multi-modal sensing and intelligent decision, the system comprises a multi-modal sensing module, an intelligent decision control module, a flexible cleaning execution module and a wastewater closed-loop recovery and treatment module, and all the modules cooperate to form a full-closed-loop intelligent cleaning system; three-dimensional point cloud and multispectral image data of a locomotive are collected through a multi-mode sensing module, and precise recognition of stains on the outer surface of the locomotive, self-adaptive planning of a cleaning path and dynamic optimization of cleaning parameters are achieved in combination with an RANSAC algorithm, a pixel-level fusion algorithm, an improved A * path planning algorithm and a BP neural network algorithm. Therefore, the whole-process automatic operation of cleaning the outer surface of the locomotive is realized; through a four-stage physical-chemical combined treatment process of the wastewater closed-loop recovery and treatment module and closed-loop management and control of the water circulation unit, the water resource recycling rate is greater than or equal to 70%, and waste resource treatment is realized.

Description

Technical Field

[0001] This invention belongs to the field of locomotive cleaning technology, and more specifically, relates to a locomotive exterior surface cleaning method and system based on multimodal perception and intelligent decision-making. Background Technology

[0002] In the field of rail transit operation and maintenance, cleaning the exterior surface of locomotives (including EMU trains and subway vehicles) is a crucial step in ensuring operational safety and extending vehicle lifespan. Its technological level directly impacts operation and maintenance efficiency, resource consumption, and environmental benefits. Current mainstream cleaning methods and related technologies have many insurmountable shortcomings and are no longer suitable for the development needs of intelligent, efficient, and green solutions, as detailed below: Traditional manual cleaning methods rely on workers manually operating high-pressure water guns and brushes, which is extremely labor-intensive. A single person can only complete the basic cleaning of 1-2 locomotives per day, and cleaning a single locomotive takes more than 2 hours, resulting in low efficiency. The cleaning quality depends entirely on the operator's experience, and individual differences lead to a stain removal rate that is often less than 80%, resulting in poor stability. There are also safety hazards of falling when working at heights such as on the roof and upper sides of the locomotive. Furthermore, there are no quantitative standards for water consumption, with an average water consumption of 15m³ per locomotive, resulting in serious waste of water resources.

[0003] Through-type fixed cleaning machines employ a "vehicles passing at a constant speed through fixed spray bars and brush curtains" operating mode. While this achieves partial automation, it suffers from the fundamental flaw of "unperceived" and "indiscriminate" cleaning due to a lack of dynamic recognition capabilities. This type of equipment cannot adapt to the contour differences between different locomotive models (such as the streamlined front differences between the Fuxing and Harmony trains), nor can it identify the distribution patterns of dirt on the vehicle body. This results in numerous cleaning blind spots in complex structural areas such as the roof edges, bogies, and undercarriages, leading to insufficient cleaning of heavily soiled areas and over-cleaning of clean areas. Furthermore, each locomotive consumes an average of 20 cubic meters of water, resulting in even greater water waste than high-quality manual cleaning, leading to poor cleaning effectiveness and economic efficiency.

[0004] While existing automated cleaning technologies attempt to overcome the aforementioned limitations, a complete intelligent cleaning system has not yet been formed, and significant technical shortcomings remain: Firstly, the perception dimension is limited, often relying on ordinary 2D vision or simple infrared distance sensors, which can only acquire planar images or rough distance information. This fails to accurately capture the three-dimensional contours of the vehicle, complex curved surfaces, and the chemical composition of stains, resulting in a lack of precise data support for cleaning decisions and the continued existence of "blind cleaning." Secondly, decision-making capabilities are weak, with core parameters such as cleaning path, water pressure, and brush speed often being preset fixed values. These parameters cannot be dynamically adjusted based on the real-time detection of stain type, distribution range, and severity, leading to low intelligence and poor adaptability. Thirdly, the execution mechanisms are rigid, often employing gantry-type fixed structures or low-degree-of-freedom (≤3 degrees of freedom) robotic arms, limiting their range of motion and posture adjustment capabilities. Cleaning coverage is often below 85%, making it difficult to cover streamlined curved surfaces on the roof, complex structures like bogies, and other areas. Fourthly, environmental considerations are lacking, with a general absence of mechanisms for wastewater recycling, treatment, and reuse. Direct discharge of wastewater pollutes the environment and wastes water resources, contradicting the green development trend under the "dual-carbon" goal.

[0005] In summary, existing locomotive exterior surface cleaning technologies face multiple challenges, including "blind cleaning, inefficient cleaning, missed cleaning, and high consumption." There is an urgent need to develop a highly efficient cleaning technology solution that integrates precise perception, intelligent decision-making, full-area coverage, and green environmental protection to address these technical deficiencies. Summary of the Invention

[0006] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a locomotive exterior surface cleaning method and system based on multimodal perception and intelligent decision-making. The method utilizes a multimodal perception module to acquire 3D point cloud and multispectral image data of the locomotive, combined with the RANSAC algorithm, pixel-level fusion algorithm, and an improved A... * The path planning algorithm and BP neural network algorithm enable accurate identification of dirt on the locomotive's exterior surface, adaptive planning of the cleaning path, and dynamic optimization of cleaning parameters. This effectively solves the technical pain points of traditional cleaning technologies, such as cleaning blind spots, unstable quality, and reliance on human experience, significantly improving cleaning accuracy and efficiency. It ensures a dirt removal rate of ≥96% on the locomotive's exterior surface while avoiding damage to the vehicle's substrate and sensitive components. Simultaneously, relying on the UWB positioning system and TOF algorithm, it achieves precise locomotive parking, constructing a fully automated operation system of "positioning-collection-decision-cleaning-recycling," significantly shortening the cleaning cycle and reducing labor costs and safety risks. Furthermore, through a four-level physical-chemical combined treatment process in the wastewater closed-loop recycling and treatment module and closed-loop management of the water circulation unit, it achieves a water resource reuse rate of ≥70%, realizing the resource-based disposal of waste and reducing environmental impact. Moreover, the algorithm of this invention adopts an industrial-grade simplified model, is compatible with conventional industrial servers, has strong system scalability, can meet the cleaning needs of various locomotive models, has real-time monitoring and quality traceability capabilities, reduces long-term operation and maintenance costs, and has significant economic value and promising prospects for widespread application.

[0007] To achieve the above objectives, one aspect of the present invention provides a method for cleaning the outer surface of a locomotive based on multimodal perception and intelligent decision-making, comprising the following steps: S1. The locomotive drives towards the cleaning station, the UWB positioning unit is activated, and the TOF algorithm dynamically guides the locomotive to accurately stop and lock its position, and sends a start signal to the multimodal sensing module. S2. The multimodal perception module is activated, the 3D laser scanner array collects 3D point cloud data of the whole vehicle, and the multispectral camera simultaneously collects multispectral images. The two types of data are transmitted to the intelligent decision control module in real time through redundant links. S3, the intelligent decision-making and control module uses the RANSAC algorithm to achieve spatial alignment between point cloud and image, and generates a 3D cleaning heat map through pixel-level fusion; based on the improved A... * The algorithm plans the optimal path for the two robotic arms, combines a BP neural network to calculate the optimal cleaning parameters, and outputs the path and parameter combination. S4. The high-degree-of-freedom robotic arm carries a composite adaptive cleaning tool and performs collaborative cleaning according to the planned path and parameters. The end vision sensor detects the cleaning quality in real time. If the quality is qualified, the robot moves out. If the quality is not qualified, it returns to step S3 to adjust the parameters and cleans again until it is qualified. S5. After being collected by the collection unit, the cleaning wastewater undergoes a four-stage physical-chemical combined treatment. The water recycling unit completes water quality testing, storage, transportation and flow regulation, and reuses the qualified clean water in the cleaning process to achieve closed-loop utilization of water resources.

[0008] Further, step S3 includes: The 3D laser point cloud data and multispectral image are spatially aligned using a point cloud registration technique based on the RANSAC algorithm. By binding the spatial coordinate information of the 3D laser point cloud with the semantic information of the stain through a pixel-level fusion algorithm, a 3D locomotive cleaning heat map with color gradient and weight value is generated based on the 3D locomotive model. Adopting improved A * The search algorithm takes "shortest total cleaning time" and "lowest total energy consumption" as dual objective functions, and introduces three major constraints: sensitive component avoidance, robotic arm movement range limitation, and collision-free collaboration between the two robotic arms. Based on the cleaning heat map, it dynamically plans the optimal motion trajectory of the two robotic arms. Meanwhile, by using the locomotive cleaning expert knowledge base and the machine learning model based on BP neural network embedded in the multi-parameter collaborative optimization model unit, the model can input the stain type, vehicle body substrate, and surface curvature information of each cleaning point and calculate and output the optimal combination of cleaning parameters in real time. The data structure after pixel-level fusion in step S3 is as follows: ; in: Type of stain; This is a quantification of the severity. For cleaning weight (comprehensive) and , , , These are the weighting coefficients. Weights for stain types; The bi-objective function described in step S3 The expression is: ; in, The time cost function is: , From the starting point to the node The estimated cleaning time, The maximum time a single robotic arm can cover the entire vehicle; Energy cost function: , For the joint speed of the robotic arm, For joint driving force, The energy consumption is obtained by integrating the joint velocity and joint driving force of the robotic arm. This represents the maximum energy consumption threshold. , The cost weight.

[0009] Furthermore, step S3 also includes using heuristic functions. Considering stain weight and distance factors, ensure that the path prioritizes covering heavily soiled areas; heuristic function The expression is: ; in: For nodes To the target area The Euclidean distance; This is the maximum distance between all vehicles. For nodes Cleaning weight; To balance the coefficients, ensure that the path simultaneously takes into account both "proximity to the target area" and "coverage of high-weight stains"; The expression for the sensitive component avoidance constraint in step S3 is: = or or ; in, The three-dimensional coordinate range of the sensitive component. For path nodes; The constraints on the robotic arm's motion range are: path nodes The distance to the coordinates of the robotic arm base is less than or equal to the working radius of the washing arm; The collision-free constraint for the dual robotic arms is: Robotic arm node 1 With the robotic arm 2 nodes The distance between the two is greater than or equal to 0.5m.

[0010] Furthermore, the improved A in step S3 * The search algorithm expression is: ; in, For nodes The overall cost; From the starting point to the node The actual cost; For nodes The heuristic cost to reach the target node.

[0011] Furthermore, step S4 also includes: the composite adaptive cleaning tool automatically switches or combines working modes according to instructions, using a reciprocating cleaning path for heavily soiled areas, a serpentine path for moderately soiled areas, and an intelligent skip + zero-intervention protection mode for clean areas.

[0012] The second aspect of this invention provides a locomotive exterior surface cleaning system based on multimodal perception and intelligent decision-making, used to implement the aforementioned cleaning method. The system includes: a multimodal perception module, an intelligent decision-making control module, a flexible cleaning execution module, and a wastewater closed-loop recycling and treatment module, with each module working collaboratively to form a fully closed-loop intelligent cleaning system. The multimodal perception module is used to construct a complete digital model of the locomotive exterior surface, consisting of "three-dimensional geometry + stain semantics," including a three-dimensional laser scanner array, a multispectral imaging unit, and a UWB ultra-wideband positioning unit. The intelligent decision-making control module is used to generate personalized cleaning solutions based on the perception data from the multimodal perception module, including a multi-source feature fusion and cleaning heatmap generation unit, an adaptive three-dimensional path planning algorithm unit, and a multi-parameter collaborative optimization model unit. The flexible cleaning execution module is used to convert the decision instructions from the intelligent decision-making control module into precise cleaning actions, including a high-degree-of-freedom robotic arm unit and a composite adaptive cleaning tool. The wastewater closed-loop recycling and treatment module is used to realize the recycling of cleaning water resources, including a wastewater collection unit, a multi-stage treatment unit, and a water circulation unit.

[0013] Furthermore, the three-dimensional laser scanner array consists of multiple high-precision three-dimensional laser scanners deployed on both sides and top of the cleaning station to form a 360-degree scanning coverage without blind spots. When the locomotive comes to a stop, the scanners start synchronously and perform high-speed scanning of the locomotive at a scanning speed of 1 million points / second. It can complete the data acquisition of the entire vehicle's outer surface within 60 seconds and generate high-density three-dimensional point cloud data with a point cloud accuracy of ±0.1mm. The multispectral imaging unit includes multiple multispectral cameras deployed corresponding to the 3D laser scanner. Each multispectral camera integrates visible and near-infrared wavelengths; the stain recognition accuracy is ≥90%. The UWB ultra-wideband positioning unit includes multiple UWB positioning base stations deployed along the track direction on the ground of the cleaning station and UWB positioning tags installed on the bottom of the locomotive; it guides the locomotive to stop accurately through the TOF positioning algorithm, with a positioning accuracy of ≤10cm.

[0014] Furthermore, the multi-source feature fusion and cleaning heat map generation unit adopts point cloud registration technology based on RANSAC algorithm to spatially align the three-dimensional laser point cloud data with the multispectral image, and then binds the spatial coordinate information of the three-dimensional laser point cloud with the stain type and severity information of the multispectral image through pixel-level fusion algorithm to generate a three-dimensional locomotive cleaning heat map with semantic information. The heatmap uses a 3D model of the locomotive as its base and visually reflects the severity of stains through color gradients. By binding color labels with weight values, it accurately indicates the specific type, 3D spatial location, and cleaning priority of stains on various parts of the locomotive's outer surface. Red corresponds to heavily soiled areas, orange to moderately soiled areas, yellow to lightly soiled areas, and green to clean areas. The weight values ​​range from 0 to 1, with heavy soiled areas having a weight value ≥ 0.8, moderately soiled areas 0.5-0.8, lightly soiled areas 0.2-0.5, and clean areas ≤ 0.2. The adaptive 3D path planning algorithm unit, based on the generated cleaning heat map, employs an improved A... * The search algorithm performs path planning with "shortest total cleaning time" and "lowest total energy consumption" as dual objective functions. It introduces three major constraints: sensitive component avoidance, robotic arm movement range limitation, and collision-free collaboration between two robotic arms, and dynamically plans the optimal motion trajectory of the high-degree-of-freedom robotic arm unit. The multi-parameter collaborative optimization model unit embeds a locomotive cleaning expert knowledge base and a machine learning model based on a BP neural network.

[0015] Furthermore, the high-degree-of-freedom robotic arm unit employs two seven-degree-of-freedom industrial-grade robotic arms, symmetrically deployed on both sides of the cleaning station. Through collaborative operation, its working range can completely cover the entire cleaning area from the undercarriage bogie to the highest point of the roof. The composite adaptive cleaning tool integrates a high-pressure water jet unit, a soft rotating brush unit, and a cleaning agent precision spraying unit into one unit, weighing ≤8kg; The composite adaptive cleaning tool head has a built-in pressure sensor and attitude sensor, which automatically adjusts the brush extension and water jet angle in complex curved areas. The composite adaptive cleaning tool's end vision sensor collects images of the cleaned area in real time and feeds them back to the intelligent decision-making module. If it detects residual dirt in a certain area, it automatically upgrades the cleaning mode, adjusts the path, and cleans again until it is qualified. After receiving instructions from the intelligent decision control module, the composite adaptive cleaning tool can automatically switch or combine multiple working modes, and the mode switching response time is ≤0.5 seconds.

[0016] Furthermore, the wastewater collection unit includes an anti-slip water collection grid and a diversion channel installed on the floor of the cleaning station, wherein the anti-slip water collection grid has a load-bearing capacity of ≥10 tons / m². 2 The slope of the guide channel is ≥3°. The multi-stage treatment unit uses a combined physical-chemical treatment method to process the wastewater collected by the wastewater collection unit through a sedimentation tank, sand filter, oil-water separator, and precision filter in four stages. The water circulation unit includes a water quality testing module, a medium-water tank storage module, a reuse pipeline transportation module, and a flow control module. It performs water quality testing, standard storage, precise transportation, and flow control on the cleaning wastewater output from the multi-stage treatment unit. The reuse pipeline transportation module is equipped with a pre-filter with a filtration accuracy of 1μm.

[0017] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: (1) The locomotive exterior surface cleaning method and system based on multimodal perception and intelligent decision-making of the present invention uses a three-dimensional laser scanner array and a multispectral imaging unit of the multimodal perception module to collect data collaboratively, and combines the RANSAC algorithm to achieve precise spatial alignment of three-dimensional point clouds and multispectral images. A three-dimensional cleaning heat map with stain semantic information and weight values ​​is generated by a pixel-level fusion algorithm, and combined with the improved A * The dual-objective path planning algorithm and BP neural network multi-parameter optimization model enable accurate identification of dirt on the locomotive exterior surface, adaptive planning of cleaning paths, and dynamic matching of cleaning parameters. This effectively solves the problem of blind spots in cleaning complex curved surfaces and hidden areas in traditional cleaning technologies, significantly improving cleaning accuracy. The dirt removal rate on the locomotive exterior surface is ≥96%, and it can avoid sensitive parts, prevent damage to the vehicle body substrate, and ensure the stability and reliability of cleaning quality.

[0018] (2) The locomotive exterior surface cleaning method and system based on multimodal perception and intelligent decision-making of the present invention relies on the UWB ultra-wideband positioning system and TOF time-of-flight positioning algorithm to achieve precise parking of locomotives ≤10cm. By constructing a fully automated operation system of "positioning-data acquisition-intelligent decision-cooperative cleaning-wastewater recycling", the whole vehicle can be cleaned without human intervention, which greatly shortens the cleaning cycle and improves the cleaning efficiency of a single locomotive by more than 70% compared with traditional manual cleaning. At the same time, it can avoid the safety risks of manual climbing operations, reduce the dependence on human experience, significantly reduce labor costs and management costs, and solve the technical pain points of high labor intensity, high safety risks and low efficiency of traditional manual cleaning.

[0019] (3) The locomotive exterior surface cleaning method and system based on multimodal perception and intelligent decision-making of the present invention, through the four-level physical-chemical combined treatment process of the wastewater closed-loop recycling and treatment module, combined with the return device to realize the recycling and re-treatment of substandard wastewater, combined with the water quality detection, storage regulation and precise delivery functions of the water circulation unit, so that the treated water quality meets the GB / T18920-2020 standard and can be stably reused for cleaning operations, with a water resource reuse rate of ≥70%, which greatly reduces the consumption of fresh water; at the same time, it can realize the resource-based disposal of waste oil and sludge, reduce pollutant emissions, meet the green and low-carbon rail transit operation and maintenance development needs, and reduce the environmental impact and operating costs of cleaning operations.

[0020] (4) The locomotive exterior surface cleaning method and system based on multimodal perception and intelligent decision-making of the present invention adopts an industrial-grade simplified model in the algorithm system, which is compatible with conventional industrial servers (computing speed ≥100GFLOPS) and has strong engineering adaptability; the multimodal perception and intelligent decision-making architecture is compatible with the cleaning needs of different types of locomotives such as EMUs, freight locomotives, and subway cars, and the system has strong scalability; each functional module is equipped with a real-time monitoring and fault warning mechanism, and the components adopt a standardized and corrosion-resistant design, which facilitates daily maintenance and repair while reducing long-term operation and maintenance costs; and the entire cleaning process data is recorded to realize quantitative control and traceability of operation quality, which can provide data support for the refined management of locomotive operation and maintenance; through the real-time feedback of the end vision sensor, the cleaning strategy can be dynamically adjusted to avoid over-cleaning or under-cleaning, ensuring that the cleaning quality of each locomotive meets the preset standard, improving the refined level of locomotive operation and maintenance management, and has broad prospects for large-scale application. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the working state of the locomotive exterior surface cleaning system based on multimodal perception and intelligent decision-making according to an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the control principle of a locomotive exterior surface cleaning system based on multimodal perception and intelligent decision-making, according to an embodiment of the present invention. Figure 3 This is a flowchart of an intelligent cleaning method for a locomotive exterior surface cleaning system based on multimodal perception and intelligent decision-making, according to an embodiment of the present invention. Figure 4 This is a schematic diagram of the structure of the composite end-of-line cleaning tool of the locomotive exterior surface cleaning system based on multimodal perception and intelligent decision-making according to an embodiment of the present invention; Figure 5 This is a schematic diagram of color marking based on a stain heatmap in the locomotive exterior surface cleaning method based on multimodal perception and intelligent decision-making according to an embodiment of the present invention. Figure 6 This is a schematic diagram of the optimized path of the robotic arm in the locomotive exterior surface cleaning method based on multimodal perception and intelligent decision-making according to an embodiment of the present invention. Figure 7 This is a schematic diagram of dynamic path planning based on a stain heat map in the locomotive exterior surface cleaning method based on multimodal perception and intelligent decision-making according to an embodiment of the present invention.

[0022] In all the accompanying drawings, the same reference numerals denote the same technical features, specifically: 1-3D laser scanner, 2-multispectral camera, 3-high degree of freedom robotic arm unit, 4-composite adaptive cleaning tool. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0024] Example 1 Please refer to Figure 1 and Figure 2This embodiment 1 provides a locomotive exterior surface cleaning system based on multimodal perception and intelligent decision-making, including a multimodal perception module, an intelligent decision-making control module, a flexible cleaning execution module, and a wastewater closed-loop recycling and treatment module. These modules work together to form a fully closed-loop intelligent cleaning system. The multimodal perception module is used to construct a complete digital model of the locomotive's exterior surface, consisting of a 3D laser scanner array, a multispectral imaging unit, and a UWB ultra-wideband positioning unit. The intelligent decision-making control module is used to generate personalized cleaning solutions based on the perceived data, including a multi-source feature fusion and cleaning heatmap generation unit, an adaptive 3D path planning algorithm unit, and a multi-parameter collaborative optimization model unit. The flexible cleaning execution module is used to convert decision commands into precise cleaning actions, including a high-degree-of-freedom robotic arm unit 3 and a composite adaptive cleaning tool 4. The wastewater closed-loop recycling and treatment module is used to realize the recycling of cleaning water resources, including a wastewater collection unit, a multi-stage treatment unit, and a water circulation unit. The system collects 3D point cloud and multispectral image data of the locomotive through the multimodal perception module, combining RANSAC algorithm, pixel-level fusion algorithm, and improved A... * The path planning algorithm and BP neural network algorithm enable accurate identification of dirt on the locomotive exterior surface, adaptive planning of cleaning path, and dynamic optimization of cleaning parameters, thereby realizing the fully automated operation of locomotive exterior surface cleaning; through the four-level physical-chemical combined treatment process of the wastewater closed-loop recycling and treatment module and the closed-loop management of the water circulation unit, the water resource reuse rate is ≥70%, realizing the resource-based disposal of waste. This invention employs an integrated closed-loop intelligent architecture encompassing "perception-decision-execution-recovery." It drives intelligent decision-making through multimodal data fusion and controls flexible actuators to complete precise and efficient cleaning operations, fundamentally solving the problems of "blind cleaning, inefficient cleaning, missed cleaning, and high consumption" inherent in existing technologies. The various modules of this invention work collaboratively to form a complete intelligent cleaning system.

[0025] Furthermore, the multimodal perception module is the foundation for achieving precise cleaning. Through the collaboration of multiple perception technologies, a complete digital model of the locomotive's outer surface, consisting of "three-dimensional geometry + stain semantics," is constructed. The three-dimensional laser scanner array comprises multiple (preferably five) high-precision three-dimensional laser scanners deployed on both sides and the top of the cleaning station, forming a 360-degree scanning coverage without blind spots. When the locomotive comes to a stop, the scanners start simultaneously, performing a high-speed scan of the locomotive at a scanning speed of 1 million points per second. This allows for the acquisition of data on the entire vehicle's outer surface within 60 seconds, generating high-density three-dimensional point cloud data with a point cloud accuracy of ±0.1mm. This data can accurately reconstruct the locomotive's three-dimensional contour model, including details such as the streamlined curved surface of the front, the bogie structure, and the roof curvature, providing a precise spatial reference for subsequent path planning. The multispectral imaging unit includes multiple (preferably five) multispectral cameras 2 deployed corresponding to the 3D laser scanner. Each multispectral camera integrates visible and near-infrared bands, simultaneously acquiring multispectral images of the locomotive's outer surface during scanning. By establishing a stain spectral database containing the spectral characteristics of more than ten common stains such as oil, mud, rust, and bird droppings, a spectral matching algorithm is used to analyze the spectral reflectance characteristics of each pixel in the image, accurately distinguishing and labeling the stain type. Simultaneously, through grayscale value quantification analysis, the severity of the stain is determined, categorized into three levels: heavy, medium, and light, with an accuracy rate ≥90%. The UWB ultra-wideband positioning unit includes multiple (preferably four) UWB positioning base stations deployed along the track direction on the ground of the cleaning station and UWB positioning tags installed on the bottom of the locomotive. When the locomotive enters the station, the UWB positioning base stations and UWB positioning tags interact in real time, and guide the locomotive to stop precisely through the TOF (Time of Flight) positioning algorithm. The positioning accuracy is ≤10cm, ensuring that the data collected by the sensing module is consistent with the working coordinate system of the execution module, providing an accurate working reference for the robotic arm. Furthermore, the intelligent decision-making and control module is key to achieving adaptive cleaning. It generates personalized cleaning plans by collaboratively processing perceived data using multiple algorithms. The multi-source feature fusion and cleaning heatmap generation unit employs point cloud registration technology based on the RANSAC algorithm to spatially align 3D laser point cloud data with multispectral images. Then, a pixel-level fusion algorithm binds the spatial coordinate information of the 3D laser point cloud with the stain type and severity information of the multispectral image, generating a 3D "locomotive cleaning heatmap" with semantic information. This heatmap uses a 3D locomotive model as its base and employs color gradients and weight values ​​of red (heavy stain), orange (medium stain), yellow (light stain), and green (clean) to visually indicate the type, location, and cleaning priority of stains at various locations. This provides a visual basis for subsequent decision-making. Specifically, red corresponds to heavily soiled areas, orange to moderately soiled areas, yellow to lightly soiled areas, and green to clean areas, with color gradients intuitively reflecting the severity of stains. At the same time, by binding color labels with weight values ​​(ranging from 0 to 1, with weight values ​​≥0.8 for heavily soiled areas, 0.5-0.8 for moderately soiled areas, 0.2-0.5 for lightly soiled areas, and ≤0.2 for clean areas), the specific types of stains on the locomotive's outer surface (such as oil stains, mud, rust, bird droppings, etc.), their three-dimensional spatial locations (including stain distribution on complex curved surfaces and structural details), and cleaning priorities (the higher the weight value, the higher the cleaning priority) are accurately indicated. This provides intuitive and accurate visual data support for subsequent adaptive three-dimensional path planning and multi-parameter collaborative optimization. The adaptive 3D path planning algorithm unit, based on the generated cleaning heat map, employs an improved A... *The search algorithm performs path planning with the dual objective functions of "shortest total cleaning time" and "lowest total energy consumption." It introduces three constraints: sensitive component avoidance, robotic arm movement range limitation, and collision-free collaboration between the two robotic arms. The algorithm dynamically plans the optimal motion trajectory of the high-degree-of-freedom robotic arm unit. Specifically, by using a pre-set component coordinate library to avoid sensitive components such as windows, sensors, and pantographs, and by ensuring collision-free collaboration between the two robotic arms, the algorithm dynamically plans the optimal motion trajectory of the two robotic arms. The generated path prioritizes covering the heavily soiled red areas (using a reciprocating cleaning path), followed by the moderately soiled orange areas (using a serpentine path), and intelligently skips the clean green areas, achieving "on-demand cleaning." The path planning efficiency is more than 40% higher than traditional algorithms. The multi-parameter collaborative optimization model unit embeds a locomotive cleaning expert knowledge base (covering the substrate characteristics and stain-adaptive cleaning parameters of mainstream locomotive models such as Fuxing, Harmony, and subway vehicles) and a machine learning model based on a BP neural network. The machine learning model is input to the stain type, vehicle substrate, surface curvature, and other information of each cleaning point in the cleaning heat map. It outputs the optimal combination of cleaning parameters in real time, including: high-pressure water jet pressure (50-150 bar), cleaning agent spray concentration (1%-5%) and flow rate (5-15 L / min), soft brush rotation speed (0-2000 rpm), and ultrasonic cavitation intensity (28-60 kHz). The parameter adjustment response time is ≤0.5 seconds, achieving refined cleaning with "one stain, one policy". Furthermore, the flexible cleaning execution module is responsible for translating decision commands into precise cleaning actions, achieving full-area coverage through a high-degree-of-freedom robotic arm unit. This high-degree-of-freedom robotic arm unit employs two seven-degree-of-freedom industrial-grade robotic arms, symmetrically deployed on both sides of the cleaning station. Each robotic arm has a working radius of 3.5m, and through collaborative operation, its working range can completely cover the entire cleaning area from the undercarriage bogie (lowest point 0.3m above the ground) to the highest point of the roof (highest point 4.5m above the ground). The high-degree-of-freedom robotic arm unit has a repeatability accuracy of ±0.05mm, enabling it to accurately follow the planned path and meet the cleaning needs of complex curved surfaces. The composite adaptive cleaning tool, as the end-effector that directly performs the cleaning action, integrates a high-pressure water jet unit, a soft rotating brush unit, and a cleaning agent precision spraying unit into one unit. Weighing ≤8kg, it can adapt to the load requirements of the robotic arm. After receiving instructions from the intelligent decision control module, the composite adaptive cleaning tool can automatically switch or combine multiple working modes. For example, the "high-pressure water + brush" combination is used for mud and dust stains, the "ultrasound + cleaning agent" combination is used for stubborn oil stains, and "low-pressure water rinsing" is used for lightly soiled areas, achieving targeted cleaning of different stains. At the same time, the soft brush is made of nylon + polyurethane composite material, which can effectively avoid damaging the paint surface of the car body. Furthermore, the wastewater closed-loop recycling and treatment module is used to realize the recycling of cleaning water resources, reducing consumption and emissions; The wastewater collection unit includes an anti-slip water collection grid (with a load-bearing capacity ≥ 10 tons / m²) installed on the ground of the cleaning station. 2 (Adaptable to locomotive weight) and a guide channel with a slope of ≥3° to ensure that all cleaning wastewater is collected without any residual water; The multi-stage treatment unit uses a combined physical-chemical treatment method to process the wastewater collected by the wastewater collection unit through a sedimentation tank (to remove large particulate impurities with a diameter ≥1mm, sedimentation time ≥30 minutes), a sand filter (to filter fine suspended solids with a diameter ≥50μm, filtration accuracy 50μm), an oil-water separator (using centrifugal separation technology, separation efficiency ≥95%, removing floating oil), and a precision filter (security filtration, filtration accuracy 5μm) for four stages of treatment. The water circulation unit is used for water quality testing, compliant storage, precise transportation, and flow control of the cleaning wastewater output from the multi-stage treatment unit, constructing a closed-loop water resource recycling system for the cleaning system. The water circulation unit includes a water quality testing module, a medium-water tank storage module, a reuse pipeline transportation module, and a flow control module. These modules work together to achieve compliant reuse and closed-loop circulation of the cleaning wastewater. The specific structure and functions are as follows: The water quality monitoring module is deployed at the end of the multi-stage treatment unit and at the inlet of the reclaimed water tank. It is equipped with suspended solids concentration sensors, oil content sensors, and pH sensors to monitor key water quality indicators of the treated water in real time (suspended solids concentration ≤5mg / L, oil content ≤1mg / L, pH value 6.0-9.0), ensuring that the effluent meets the GB / T18920-2020 standard for "Water Quality of Urban Wastewater Reuse for Miscellaneous Uses". The monitoring data is fed back to the intelligent decision control module in real time. If the water quality does not meet the standards, the return device is triggered to return the wastewater to the multi-stage treatment unit for reprocessing. Medium water tank storage module: with a volume of 10m³ 3 The corrosion-resistant stainless steel water tank features heat insulation and leak-proof functions, and is used to store compliant clean water. The tank has a built-in level sensor to monitor water level changes in real time; if the water level falls below a preset lower limit (2m), the sensor will detect the water level. 3 When the water level exceeds the preset upper limit (8m), the multi-level processing unit is triggered to accelerate the processing pace. 3 When the water tank overflows, the wastewater treatment process should be suspended to prevent the tank from overflowing. A drain outlet should be installed at the bottom of the tank to regularly remove any accumulated trace impurities and ensure the cleanliness of the stored water. Recycled pipeline delivery module: It adopts an independent corrosion-resistant PVC pipeline system, which is connected to the pre-spray pipeline and the final rinsing pipeline of the cleaning system to form a dedicated reuse channel, avoiding contamination caused by mixing with fresh water pipelines; the pipeline is equipped with a pre-filter with a filtration accuracy of 1μm to further intercept the tiny impurities remaining in the water and protect the nozzles, brushes and other components of the cleaning execution module. Flow control module: Includes a variable frequency booster pump and a flow sensor. Based on the real-time needs of the cleaning operation (such as a large flow rate required in the pre-spray stage and a precise flow rate required in the rinsing stage), the output power of the variable frequency pump is adjusted through the intelligent decision control module to achieve dynamic control of the reuse clean water flow rate (adjustable from 5-15L / min). At the same time, the flow sensor feeds back data to ensure stable delivery flow rate, adapt to the parameter requirements of different cleaning stages, and achieve efficient reuse of water resources. The reflux device includes: Electric three-way valve: deployed at the junction of the outlet and return pipes of each treatment unit; Return pump: a variable frequency centrifugal pump (flow rate adapted to the treatment unit capacity, such as 5-10 m³ / h) is selected and installed on the return pipe; Check valve: installed on the return pipe near the inlet of the treatment unit; Control link: includes a signal receiver and an actuator controller, which receives real-time detection data from water quality monitoring sensors and instructions from the intelligent decision control module to control the switching timing of the electric three-way valve and the start, stop, and speed of the return pump, thereby realizing automated return control; After multi-stage treatment, the wastewater is first verified by the water quality detection module of the water circulation unit to ensure that the water quality meets the GB / T18920-2020 standard for "Urban Wastewater Reuse for Urban Miscellaneous Use". For wastewater that does not meet the standard, a return command is triggered, causing it to return to the multi-stage treatment unit for reprocessing. The qualified clean water is temporarily stored in the intermediate water tank storage module. The water tank has a built-in liquid level monitoring component to regulate the inflow and outflow rhythm in real time to avoid overflow or water shortage. Through an independent reuse pipeline transportation module, the stored clean water is directed to the pre-spray stage or the final rinsing stage of the cleaning system, completely isolated from the fresh water pipeline to prevent cross-contamination. Relying on the flow control module, the water demand of different cleaning stages is dynamically adapted, and the transportation flow and pressure are precisely adjusted to achieve efficient reuse of water resources, reduce the water consumption of the overall cleaning process, and achieve the goal of green and environmentally friendly cleaning.

[0026] Example 2 like Figure 3 As shown, Embodiment 2 of the present invention provides a locomotive exterior surface cleaning method based on multimodal perception and intelligent decision-making, including the following steps: S1. The locomotive drives towards the cleaning station, the UWB positioning unit is activated, and the TOF algorithm dynamically guides the locomotive to accurately stop and lock its position, and sends a start signal to the multimodal sensing module. S2. The multimodal perception module is activated, the 3D laser scanner array collects 3D point cloud data of the whole vehicle, and the multispectral camera simultaneously collects multispectral images. The two types of data are transmitted to the intelligent decision control module in real time through redundant links. S3, the intelligent decision-making and control module uses the RANSAC algorithm to achieve spatial alignment between point cloud and image, and generates a 3D cleaning heat map through pixel-level fusion; based on the improved A... * The algorithm plans the optimal path for the two robotic arms, combines a BP neural network to calculate the optimal cleaning parameters, and outputs the path and parameter combination. S4. A high-degree-of-freedom robotic arm carries a composite adaptive cleaning tool and performs collaborative cleaning according to the planned path and parameters. The end vision sensor detects the cleaning quality in real time. If the quality is qualified, the robot moves out. If the quality is not qualified, the parameters are adjusted and the cleaning is repeated. S5. After being collected by the collection unit, the cleaning wastewater undergoes a four-stage physical-chemical combined treatment process. The water recycling unit completes water quality testing, storage, transportation, and flow control, and reuses the qualified clean water in the cleaning process, realizing closed-loop utilization of water resources. Further, step S1 specifically includes the locomotive driving towards the cleaning station under the guidance of the dispatching system. The UWB ultra-wideband positioning unit is automatically activated. Multiple UWB positioning base stations deployed along the track direction on the ground of the cleaning station interact in real time with the UWB positioning tag installed on the bottom of the locomotive. The TOF time-of-flight positioning algorithm is used to dynamically adjust the guidance command to guide the locomotive to stop precisely at the preset cleaning position, ensuring that the stopping error is ≤10cm. After stopping, the UWB ultra-wideband positioning unit of the cleaning system works in conjunction with the intelligent decision control module to automatically lock the locomotive position and send a start signal to the multimodal perception module. The scheduling system mentioned in step S1 is an existing mature system in the rail transit operation and maintenance system used to coordinate locomotive operation scheduling and task allocation. It is an external supporting system that works in conjunction with the cleaning system of the present invention. Its core function is to send driving instructions to the target locomotive according to the rail transit operation and maintenance plan (such as locomotive maintenance schedule, cleaning demand priority, etc.), and guide the locomotive to drive precisely from the operating track or parking area to the dedicated cleaning station of the present invention, so as to provide the prerequisites for the start of subsequent cleaning operations.

[0027] Furthermore, in step S1, the UWB ultra-wideband positioning unit is automatically activated. Through real-time interaction between the UWB positioning base station deployed on the ground along the track direction at the cleaning station and the UWB positioning tag installed on the bottom of the locomotive, the TOF time-of-flight positioning algorithm is used to dynamically adjust the guidance instructions and guide the locomotive to accurately stop at the preset cleaning position. Further, step S2 specifically includes the following steps: after the multimodal perception module receives the start signal, it triggers the synchronous start of the 3D laser scanner array unit and the multispectral imaging unit; multiple high-precision 3D laser scanners in the 3D laser scanner array unit perform a 360-degree high-speed scan of the locomotive at a scanning speed of 1 million points / second, completing the acquisition of data on the outer surface of the entire vehicle within 60 seconds, generating high-density 3D point cloud data with a point cloud accuracy of ±0.1mm; multiple multispectral cameras in the multispectral imaging unit (integrating two bands: visible light (400-760nm) and near-infrared (760-1100nm)) synchronously acquire multispectral images of the outer surface of the locomotive; the 3D point cloud data and multispectral image data are sent to the intelligent decision control module in real time through a redundant transmission link (fiber optic + industrial Ethernet); Further, step S3 specifically includes: the intelligent decision control module performs collaborative processing on the collected multimodal data (including 3D point cloud data and multispectral image data): firstly, the 3D laser point cloud data and multispectral images are spatially aligned using point cloud registration technology based on the RANSAC algorithm; then, the spatial coordinate information of the 3D laser point cloud and the semantic information of the stains are bound together using a pixel-level fusion algorithm to generate a 3D locomotive cleaning heat map with color gradient and weight values ​​based on the locomotive 3D model; subsequently, an improved A... * The search algorithm uses "shortest total cleaning time" and "lowest total energy consumption" as dual objective functions. It introduces three major constraints: sensitive component avoidance, robotic arm movement range limitation, and collision-free collaboration between the two robotic arms. Based on the cleaning heat map, it dynamically plans the optimal motion trajectory of the two robotic arms. At the same time, through the locomotive cleaning expert knowledge base embedded in the multi-parameter collaborative optimization model unit and the machine learning model based on BP neural network, it takes the stain type, vehicle body substrate, and surface curvature information of each cleaning point as input, calculates and outputs the optimal cleaning parameter combination in real time, completes adaptive three-dimensional path planning and multi-parameter collaborative optimization, and finally outputs the optimal cleaning path and parameter combination. In step S3, the RANSAC algorithm (Random Sample Consensus Algorithm) is used to solve the spatial coordinate mapping problem between 3D laser point cloud data (spatial geometric information) and multispectral image data (2D semantic information). This ensures that the stain semantic information (type, severity) of each pixel can be accurately bound to the corresponding spatial location in the 3D point cloud, laying a foundation for coordinate unification for subsequent heatmap generation. The RANSAC algorithm eliminates outlier matching points (outsiders) through random sampling and solves for the optimal spatial transformation matrix, achieving rigid alignment between the two types of data. Furthermore, in step S3, the pixel-level fusion algorithm is used to bind the spatially aligned 3D point cloud coordinates with the stain semantic information (type, severity) of the multispectral image at the pixel level, generating a semantic 3D cleaning heat map, and realizing a unified representation of "geometric location + stain attributes". The process involves binding the spatial coordinates of the 3D laser point cloud with the semantic information of the stains using a pixel-level fusion algorithm to generate a 3D locomotive cleaning heatmap with color gradients and weight values, based on a 3D locomotive model. The severity of stains is analyzed using grayscale values ​​from multispectral images, and quantitative indicators are defined. (Cleaning priority weight, value range 0-1):

[0028] in: For the first The grayscale value of each pixel (grayscale value in the near-infrared band of a multispectral image; the heavier the stain, the lower the grayscale value); The maximum grayscale value for the clean area (default is 255); The minimum grayscale value for heavily polluted areas (default is 50); Corresponding to heavy pollution (red), Corresponding to medium sludge (yellow), Corresponding to light pollution (orange), Corresponding to cleanliness (green); For each point in the 3D point cloud The corresponding multispectral image pixels are obtained using the RANSAC algorithm. The merged data structure is as follows: in: The type of stain (output by a spectral matching algorithm, with values ​​such as oil = 1, mud = 2, rust = 3, bird droppings = 4, etc.); The severity level is quantified (calculated using the formula above); For cleaning weight (comprehensive) and , , , where are weighting coefficients. Weighting by stain type, such as oil stains mud and dust );

[0029] After fusion, a 3D heatmap is generated using a voxelized mesh (voxel size 0.5mm × 0.5mm × 0.5mm). The color of each voxel is determined by... Mapping (red → yellow → orange → green); Furthermore, the improved A described in step S3 * The search algorithm is used to achieve on-demand cleaning with "shortest total cleaning time" and "lowest total energy consumption" based on a 3D cleaning heat map, avoiding sensitive parts and planning a collision-free collaborative trajectory for two robotic arms. The dual objective function The expression is: in, The time cost function (normalized to 0-1): , From the starting point to the node The estimated cleaning time, The maximum time for a single robotic arm to cover the entire vehicle (preset to 10 minutes). Energy cost function (normalized to 0-1): , For the joint speed of the robotic arm, For joint driving force, The energy consumption is obtained by integrating the joint velocity and joint driving force of the robotic arm. This represents the maximum energy consumption threshold. , Weighted by cost (prioritizing cleaning efficiency); Step S3 also includes using heuristic functions Considering stain weight and distance factors, ensure that the path prioritizes covering heavily soiled areas; heuristic function The expression is: in: For nodes To the target area Euclidean distance between heavily polluted and moderately polluted areas; Maximum distance for the entire vehicle (default value) ); For nodes Cleaning weights (heatmap weight values); To balance the coefficients, ensure that the path simultaneously takes into account both "proximity to the target area" and "coverage of high-weight stains"; Furthermore, the expression for the sensitive component avoidance constraint in step S3 is as follows: or or

[0030] in, The three-dimensional coordinate range of sensitive components (windows, pantographs). For path nodes; The constraint expression for the robotic arm's motion range is: path node The distance to the coordinates of the robot arm base is less than or equal to the working radius of the robot arm, such as a working radius of 3.5m; The expression for the collision-free constraint condition of the dual robotic arms is: Robotic arm node 1 With the robotic arm 2 nodes Distance between the two , To maintain a safe distance and avoid collisions with the robotic arm; Furthermore, the improved A in step S3 * The search algorithm expression is: in, For nodes The overall cost; From the starting point to the node The actual cost (i.e. the cost that has already been incurred); For nodes The heuristic cost (i.e. the estimated cost) to reach the target node. Through iterative calculation ,choose The smallest node expands until all non-clean areas are covered, with the final output being: Robotic Arm 1 Path: (Prioritizes coverage of heavily soiled areas such as the edge of the roof and the center of the vehicle body, using a reciprocating path); Robotic arm path 2: (Prioritize coverage of heavily contaminated areas such as bogies and the lower part of the car body, using a reciprocating path); The path planning efficiency of this invention is higher than that of traditional A * The algorithm is improved by 40%, and the path planning time for a single robotic arm is ≤2 seconds. Furthermore, the BP neural network algorithm described in step S3 outputs the optimal combination of cleaning parameters (high-pressure water pressure, cleaning agent concentration, brush speed, etc.) in real time based on input information such as the type of stain at the cleaning point, the vehicle body substrate, and the surface curvature, to achieve "one solution for each stain". In a specific embodiment of the present invention, the BP neural network structure of the BP neural network algorithm is as follows: Input layer: 3 neurons, namely (Stain type, unique thermal code: oil stain = [1,0,0,0], dirt = [0,1,0,0], etc.) (Body body base material: aluminum alloy = 1, stainless steel = 2, fiberglass = 3) (Surface curvature) (Unit: 1 / m) Hidden layers: 2 layers, 10 neurons per layer, using the ReLU activation function. ; Output layer: 5 neurons, namely (High-pressure water, bar) (Cleaning agent concentration, %) (Flow rate, L / min) (Brush rotation speed, rpm) (Ultrasonic frequency, kHz); The formula for inter-layer mapping (taking the input layer to the first hidden layer as an example) is expressed as follows: in: For the first hidden layer The output of each neuron; For the input layer The first hidden layer contains neurons. The weights of each neuron; For the first hidden layer Bias terms for each neuron; It is the ReLU activation function; The output layer mapping formula is expressed as follows: in: For the second hidden layer The output of each neuron; For the second hidden layer The nth neuron to the output layer The weights of each neuron; For the output layer Bias terms for each neuron; For the Sigmoid function (normalizes the output to 0-1); , These are the maximum and minimum values ​​of the output parameters; Error backpropagation formula (training process): using mean squared error (MSE) as the loss function: in, These are the actual optimal parameters (provided by the expert knowledge base). Predict the network output; update the weights and biases using gradient descent: in, The learning rate is the network prediction error after training. Parameter adjustment response time Second; The intelligent decision-making and planning algorithm system of this invention achieves efficient operation through temporal coordination, data coordination, and engineering adaptation. Temporally, it follows the sequence of RANSAC algorithm → pixel-level fusion algorithm → improved A... *The algorithm follows a progressive logic of BP neural network, with the output of the preceding algorithm directly serving as the input data for the subsequent algorithm. At the data level, the RANSAC algorithm is used to unify the coordinates of the 3D point cloud and multispectral image, and pixel-level fusion is performed to form data with both "location + semantic" dual-dimensional information, providing core support for subsequent cleaning path planning and multi-parameter collaborative optimization. In engineering applications, all algorithms adopt industrial-grade simplified models to avoid complex calculations and ensure that the execution time of a single round of the full-process algorithm is ≤4 seconds, which can adapt to the real-time processing requirements of industrial servers with a computing speed of ≥100GFLOPS.

[0031] Further, step S4 specifically includes: the decision command from the intelligent decision control module is sent to the high-degree-of-freedom robotic arm unit of the flexible cleaning execution module via the control bus; the seven-degree-of-freedom industrial-grade robotic arm of the high-degree-of-freedom robotic arm unit carries a composite adaptive cleaning tool to perform collaborative cleaning operations according to the planned path and parameters; the composite adaptive cleaning tool automatically switches or combines working modes according to the instructions, using a reciprocating cleaning path for heavily soiled areas, a serpentine path for moderately soiled areas, and an intelligent skip + zero-intervention protection mode for clean areas; during the cleaning process, the image of the cleaning area is collected in real time by the vision sensor built into the end effector and fed back to the intelligent decision control module for quality detection; if qualified, the machine leaves; if residual stains are found, the process returns to step S3 to adjust the cleaning parameters in real time and clean again until the cleaning requirements are met. Step S4 involves using a reciprocating cleaning path for heavily soiled areas, specifically including: The robotic arm, carrying a composite adaptive cleaning tool, moves along a planned reciprocating path. The path coverage density is 1.5 times that of the medium-soil area, and the spacing between adjacent paths is ≤5cm, ensuring that heavily soiled areas are thoroughly cleaned. The number of reciprocations is dynamically adjusted according to the stain weight: 3 reciprocations are performed when the weight value is ≥0.9, and 2 reciprocations are performed when the weight value is 0.7-0.9, ensuring that stubborn stains are fully removed. The composite adaptive cleaning tool automatically activates a four-in-one collaborative mode of "precise cleaning agent spraying + ultrasonic cavitation + high-pressure water jet + soft rotating brush," and the parameters of each unit are precisely matched according to the output of the multi-parameter optimization model. Cleaning agent spraying unit: Sprays at the optimal concentration of 1%-5% (e.g., 3%-4% for heavy oil stains, 4%-5% for heavy rust stains), with a flow rate of 8-15L / min, ensuring full contact and reaction between the cleaning agent and the stains; Ultrasonic cavitation unit: Activates the optimal frequency of 28-60kHz (40-50kHz for stubborn oil stains, 28-35kHz for heavy mud and dust), using cavitation effect to break the adhesion between the stains and the vehicle body substrate; High-pressure water jet unit: Outputs 100-150bar high pressure, with the water jet sprayed in a fan shape (spray angle 30°). (°-60°) The brush is angled at 15°-30° to the direction of brush movement to assist in rinsing away the detached stains; Soft rotating brush unit: Rotates at high speed of 1500-2000rpm, the nylon + polyurethane composite material brush closely conforms to the curved surface of the car body, enhancing the cleaning effect through mechanical friction while avoiding scratching the paint surface; During operation, the cleaning agent is first sprayed to soak (pre-action time 2-3 seconds), then the ultrasonic waves are activated synchronously with the brush, and the high-pressure water jet rinses away the detached stains in real time, forming a "soaking-detaching-rinsing" closed loop to ensure that the stain removal rate in heavily soiled areas is ≥98%; Step S4 employs a serpentine path for the medium-contaminated area, specifically including: The robotic arm moves at a constant speed along a serpentine path (0.8-1.2 m / s), with adjacent paths spaced ≤ 8 cm, balancing cleaning efficiency and coverage. It completes thorough cleaning in a single serpentine scan without reciprocating. The composite adaptive cleaning tool automatically switches to a two-in-one collaborative mode of "high-pressure water jet + soft rotating brush," and in some scenarios (such as light oil stains), a low-concentration cleaning agent is sprayed in conjunction. The parameters are configured as follows: High-pressure water jet unit: outputs medium pressure of 50-100 bar, water jet angle of 45°-90°, flow rate of 5-10 L / min, balancing cleaning power and water consumption; Soft rotating brush unit: rotates at medium speed of 800-1500 rpm, adaptively adjusting brush pressure according to the curvature of the vehicle body surface (pressure of 0.3-0.5 MPa on curved protrusions, pressure of 0.5-0.8 MPa on flat areas); Cleaning agent spray unit: activates only when the dirt contains oil, concentration of 1%-3%, flow rate of 5-8 L / min, achieving efficient cleaning without ultrasonic assistance; During operation, the high-pressure water jet and brush work synchronously. The water jet first initially washes away floating dirt, the brush friction removes firmly attached stains, and the subsequent water flow promptly carries away the stains to prevent secondary adhesion; Step S4 employs an intelligent skip + zero-intervention protection mode for clean areas, specifically including: The robotic arm quickly traverses the area according to the planned path, with a movement speed of 1.5-2.0 m / s. The path maintains a safe distance of ≥10 cm from the boundary of the clean area to avoid ineffective work. The composite adaptive cleaning tool automatically shuts down all decontamination function units (high-pressure water, cleaning agents, ultrasonic waves, and brushes all stop working), maintaining only the posture adjustment and position movement of the robotic arm to ensure that it does not cause unnecessary rinsing or contact with the clean area. For clean areas containing sensitive components such as windows and pantographs, the vision sensor built into the tool's end reconfirms the area boundary, and the robotic arm's movement trajectory is further offset by ≥5 cm. At the same time, all fluid jet units are shut down to completely avoid damage to sensitive components caused by high-pressure water and cleaning agents (such as aging of window sealing strips, short circuits in pantograph electrical components, etc.). The composite adaptive cleaning tool receives instructions from the intelligent decision module via EtherCAT industrial Ethernet, with a mode switching response time of ≤0.5 seconds, ensuring rapid switching at the boundary between heavily contaminated, medium-contaminated, and clean areas without transitional ineffective work. Curved surface adaptation: The composite adaptive cleaning tool head has built-in pressure and posture sensors, which automatically adjust the brush extension (0-50mm) and water jet angle in complex curved surface areas (such as the streamlined surface of the front of the car and the concave and convex structure of the bogie) to ensure that the cleaning mode is accurately adapted to the shape of the car body. Quality feedback linkage: The vision sensor at the end of the composite adaptive cleaning tool collects images of the cleaning process in real time and feeds them back to the intelligent decision module. If residual dirt is detected in a certain area (not meeting the cleanliness standard), the cleaning mode is automatically upgraded (e.g., medium dirt mode is upgraded to heavy dirt mode), and the path is adjusted to reciprocating to clean again until it meets the standard. Furthermore, step S5 specifically includes: all wastewater generated during the cleaning process is collected by the wastewater collection unit of the wastewater closed-loop recycling and treatment module and enters the multi-stage treatment unit. The wastewater is treated using a combination of physical and chemical methods, passing through a sedimentation tank, sand filter, oil-water separator, and precision filter for four stages of treatment. The water circulation unit performs water quality testing on the clean water after treatment by the multi-stage treatment unit, ensuring that the water meets the standards, storing it, accurately transporting it, and controlling its flow. Finally, the water is reused specifically in the pre-spraying process or the final rinsing process to achieve closed-loop recycling of water resources. In step S5, the primary treatment is a sedimentation tank (physical sedimentation treatment), which uses gravity sedimentation to remove large suspended solid impurities from the wastewater, reduce the load on subsequent treatment units, and prevent large particles from clogging subsequent filtration equipment. Equipment structure and parameters: Rectangular reinforced concrete sedimentation tank with an effective volume ≥5m³ and a design hydraulic retention time ≥30 minutes; a sludge discharge hopper with a slope ≥2° is installed at the bottom of the tank to facilitate the centralized discharge of settled impurities; an overflow weir is installed at the top of the tank to ensure that wastewater flows evenly through the tank and to guarantee the sedimentation effect; The treatment process is as follows: the collected wastewater flows into the sedimentation tank by gravity through the diversion channel, and the water flow velocity is controlled at 0.05-0.1m / s. During the slow flow, large particles of impurities with a diameter ≥1mm (such as silt, metal scraps, large pieces of dirt and residue) settle to the bottom of the tank under gravity. The settled sludge is discharged periodically (every 8 hours) by the sewage pump. After dewatering by filter press, the sludge can be recycled and disposed of as general industrial solid waste (such as for roadbed filling). Treatment objective: Remove more than 80% of large particulate suspended impurities from wastewater, with an effluent suspended solids concentration ≤100mg / L, ensuring stable operation of the subsequent sand filter; The secondary treatment is a sand filter (physical filtration treatment), which uses quartz sand as the filter medium. By utilizing the interception and sieving effect of the medium, it removes small suspended particles and colloidal substances in the wastewater, further purifying the water quality. Equipment structure and parameters: A pressure sand filter is selected, with double-layer quartz sand as the filter media (upper layer coarse sand particle size 1.2-2.0mm, lower layer fine sand particle size 0.5-1.0mm), and filter layer thickness ≥800mm; the design filtration speed is 8-12m / h, and the working pressure is 0.2-0.4MPa; equipped with a backwashing system (air backwashing + water backwashing), the backwashing cycle is 24 hours, and the backwashing time is 15-20 minutes to ensure filter media regeneration; Treatment process: Wastewater treated in the sedimentation tank is pumped into the sand filter and flows from top to bottom through the quartz sand filter layer. Fine suspended solids with a diameter ≥50μm (such as fine silt, dirt powder, colloidal particles, etc.) are trapped by the filter media. During the filtration process, the pressure difference between the inlet and outlet water is monitored in real time. When the pressure difference is ≥0.1MPa, the backwashing program is started. The impurities attached to the surface of the filter media are peeled off and discharged by a combination of air scrubbing and high-pressure water rinsing, restoring the filtration performance of the filter media. Treatment objective: Remove more than 90% of fine suspended impurities from wastewater, with effluent suspended solids concentration ≤10mg / L and water turbidity ≤5NTU, creating conditions for subsequent oil-water separation; The tertiary treatment is an oil-water separator (physical separation + chemical auxiliary treatment). Based on the density difference between oil and water, it combines centrifugal separation technology to enhance the separation effect. At the same time, a demulsifier is added to help break up the oil-water emulsion system, effectively removing floating oil, dispersed oil and some emulsified oil from the wastewater. Equipment structure and parameters: It adopts a centrifugal oil-water separator with a processing flow rate ≥5m³ / h and a separation efficiency ≥95%; the equipment has a built-in centrifugal separation chamber (speed 3000-5000r / min) and a demulsification reaction zone, and the demulsifier dosage is 5-10mg / L (a non-ionic demulsifier, such as a polyether compound, is selected to avoid secondary pollution). Treatment process: The wastewater after sand filtration first enters the demulsification reaction zone and is thoroughly mixed with the added demulsifier (mixing time ≥ 5 minutes) to break the oil-water emulsion state; then it enters the centrifugal separation chamber, where, under the action of centrifugal force, oil sludge with a density less than water (such as lubricating oil and diesel oil residue on locomotive surfaces) is thrown to the inner wall of the separator, collected, and discharged through the oil collection tank. The recovered waste oil can be reused as industrial fuel after purification treatment; the treated water flows out from the bottom of the separator and enters the subsequent precision filter. Treatment objective: Remove more than 95% of oily substances from wastewater, with an oil content of ≤5mg / L in the effluent, to avoid oil contamination of subsequent filter media and ensure the safety of water reuse; The fourth stage of treatment is a precision filter (deep filtration treatment), which uses a high-precision filter membrane as the filter medium to intercept the tiny suspended solids, colloids, bacteria and some organic matter remaining after the previous treatment, so as to achieve deep purification of water quality and ensure that the effluent meets the reuse standards. Equipment structure and parameters: A safety-type precision filter is selected, with a filtration accuracy of 5μm. The filter element material is a polypropylene pleated membrane (acid and alkali resistant, corrosion resistant, suitable for cleaning wastewater quality); the equipment is designed to withstand a pressure of 0.3-0.5MPa, with a single filter element flow rate ≥1m³ / h. The number of filter elements is configured according to the treatment flow rate (ensuring the total treatment capacity matches the wastewater generation); a differential pressure monitoring device is provided to monitor the filter element clogging status in real time. Treatment process: Wastewater treated by the oil-water separator enters the precision filter, where the filter element removes tiny impurities with a diameter ≥5μm (such as residual fine particles, trace oil droplets remaining after demulsification, and suspended bacteria in the water) through deep interception. When the pressure difference between the inlet and outlet water is ≥0.15MPa, the filter element is replaced, and the used filter element is disposed of in accordance with the hazardous waste regulations. Treatment objective: To further remove residual suspended impurities and trace amounts of oil from wastewater, with effluent suspended solids concentration ≤5mg / L and oil content ≤1mg / L, meeting the GB / T18920-2020 standard for "Urban Wastewater Reuse for Urban Miscellaneous Uses" and safe for reuse in the pre-spraying or final rinsing stages of the cleaning system of this invention. The cleaning system of the present invention connects a sedimentation tank, a sand filter, an oil-water separator, and a precision filter in series through a pre-designed pipeline layout. It combines a "gravity flow + pressurization" composite transportation method (wastewater from upstream equipment (such as sedimentation tank) to downstream equipment (such as sand filter) can be transported by gravity through a guide channel, and the links that require pressure increase (such as from sand filter to oil-water separator) are transported stably through a built-in pressurization pump) to ensure continuous and uninterrupted water flow between each treatment device. Each treatment unit is equipped with water quality monitoring sensors (suspended solids concentration sensor and oil content sensor) at its outlet to collect effluent water quality data (such as suspended solids concentration and oil content) in real time. The monitoring data is then transmitted to the intelligent decision control module in real time via an industrial bus. The intelligent decision control module compares and judges the collected effluent water quality data in real time based on the built-in water quality compliance thresholds (such as suspended solids ≤100mg / L in sedimentation tank effluent and oil content ≤5mg / L in oil-water separator effluent). When the intelligent decision control module detects that the effluent data of a certain treatment device exceeds the compliance threshold, it immediately generates a reflux control command, triggering the corresponding reflux device (such as an electric three-way valve or reflux pump) to start. The reflux device guides the substandard wastewater from the outlet pipe of the device back to the inlet of the device, and re-enters the treatment unit for secondary treatment (such as the reflux wastewater in the sedimentation tank undergoing re-settling, or the reflux wastewater in the sand filter undergoing re-filtration). Only when the sensor detects that the effluent water quality meets the standards again will the intelligent decision control module shut down the reflux device and allow the wastewater to enter the next treatment device, ensuring that the water quality after each stage of treatment meets the requirements of subsequent processes. The processing capacity of the entire multi-stage processing unit of this invention is ≥10m. 3 / h, capable of adapting to a single locomotive (water consumption ≤8m³ / h) 3 The treatment capacity for cleaning wastewater is ≤30 minutes, ensuring rapid recycling and reuse of water resources.

[0032] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for cleaning the outer surface of a locomotive based on multimodal perception and intelligent decision-making, characterized in that, Includes the following steps: S1. The locomotive drives towards the cleaning station, the UWB positioning unit is activated, and the TOF algorithm dynamically guides the locomotive to accurately stop and lock its position, and sends a start signal to the multimodal sensing module. S2. The multimodal perception module is activated, the 3D laser scanner array collects 3D point cloud data of the whole vehicle, and the multispectral camera simultaneously collects multispectral images. The two types of data are transmitted to the intelligent decision control module in real time through redundant links. S3, the intelligent decision-making and control module uses the RANSAC algorithm to achieve spatial alignment between point cloud and image, and generates a 3D cleaning heat map through pixel-level fusion; based on the improved A... * The algorithm plans the optimal path for the two robotic arms, combines a BP neural network to calculate the optimal cleaning parameters, and outputs the path and parameter combination. S4. The high-degree-of-freedom robotic arm carries a composite adaptive cleaning tool and performs collaborative cleaning according to the planned path and parameters. The end vision sensor detects the cleaning quality in real time. If the quality is qualified, the robot moves out. If the quality is not qualified, it returns to step S3 to adjust the parameters and cleans again until it is qualified. S5. After being collected by the collection unit, the cleaning wastewater undergoes a four-stage physical-chemical combined treatment. The water circulation unit completes water quality testing, storage, transportation and flow regulation, and reuses the qualified clean water for the cleaning process, realizing the closed-loop utilization of water resources.

2. The locomotive exterior surface cleaning method based on multimodal perception and intelligent decision-making according to claim 1, characterized in that, Step S3 includes: The 3D laser point cloud data and multispectral image are spatially aligned using a point cloud registration technique based on the RANSAC algorithm. By binding the spatial coordinate information of the 3D laser point cloud with the semantic information of the stain through a pixel-level fusion algorithm, a 3D locomotive cleaning heat map with color gradient and weight value is generated based on the 3D locomotive model. Adopting improved A * The search algorithm uses "shortest total cleaning time" and "lowest total energy consumption" as dual objective functions, and introduces three major constraints: sensitive component avoidance, robotic arm movement range limitation, and collision-free collaboration between the two robotic arms. Based on the cleaning heat map, it dynamically plans the optimal motion trajectory of the two robotic arms. Meanwhile, by using the locomotive cleaning expert knowledge base and the machine learning model based on BP neural network embedded in the multi-parameter collaborative optimization model unit, the model can input the stain type, vehicle body substrate, and surface curvature information of each cleaning point and calculate and output the optimal combination of cleaning parameters in real time. The data structure after pixel-level fusion in step S3 is as follows: ; in: The type of stain; This is a quantification of the severity. For cleaning weight (comprehensive) and , , , These are the weighting coefficients. Weights for stain types; The bi-objective function described in step S3 The expression is: ; in, The time cost function is: , From the starting point to the node The estimated cleaning time, The maximum time a single robotic arm can cover the entire vehicle; Energy cost function: , For the joint speed of the robotic arm, For joint driving force, The energy consumption is obtained by integrating the joint velocity and joint driving force of the robotic arm. This represents the maximum energy consumption threshold. , The cost weight.

3. The locomotive exterior surface cleaning method based on multimodal perception and intelligent decision-making according to claim 2, characterized in that, Step S3 also includes using heuristic functions Considering stain weight and distance factors, ensure that the path prioritizes covering heavily soiled areas; heuristic function The expression is: ; in: For nodes To the target area The Euclidean distance; This is the maximum distance between all vehicles. For nodes Cleaning weight; To balance the coefficients, ensure that the path simultaneously takes into account both "proximity to the target area" and "coverage of high-weight stains"; The expression for the sensitive component avoidance constraint in step S3 is: = or or ; in, The three-dimensional coordinate range of the sensitive component. For path nodes; The constraints on the robotic arm's motion range are: path nodes The distance to the coordinates of the robotic arm base is less than or equal to the working radius of the washing arm; The collision-free constraint for the dual robotic arms is: Robotic arm node 1 With the robotic arm 2 nodes The distance between the two is greater than or equal to 0.5m.

4. The locomotive exterior surface cleaning method based on multimodal perception and intelligent decision-making according to claim 3, characterized in that, Improved A in step S3 * The search algorithm expression is: ; in, For nodes The overall cost; From the starting point to the node The actual cost; For nodes The heuristic cost to reach the target node.

5. A locomotive exterior surface cleaning method based on multimodal perception and intelligent decision-making according to any one of claims 1-4, characterized in that, Step S4 also includes: the composite adaptive cleaning tool automatically switches or combines working modes according to instructions, using a reciprocating cleaning path for heavily soiled areas, a serpentine path for medium-soiled areas, and an intelligent skip + zero-intervention protection mode for clean areas.

6. A locomotive exterior surface cleaning system based on multimodal perception and intelligent decision-making, characterized in that, The cleaning method described in any one of claims 1-5 comprises: a multimodal perception module, an intelligent decision control module, a flexible cleaning execution module, and a wastewater closed-loop recycling and treatment module; the multimodal perception module is used to construct a complete digital model of the locomotive's outer surface, consisting of "three-dimensional geometry + stain semantics," including a three-dimensional laser scanner array, a multispectral imaging unit, and a UWB ultra-wideband positioning unit; the intelligent decision control module is used to generate personalized cleaning solutions based on the perception data from the multimodal perception module, including a multi-source feature fusion and cleaning heatmap generation unit, an adaptive three-dimensional path planning algorithm unit, and a multi-parameter collaborative optimization model unit; the flexible cleaning execution module is used to convert the decision instructions from the intelligent decision control module into precise cleaning actions, including a high-degree-of-freedom robotic arm unit and a composite adaptive cleaning tool; the wastewater closed-loop recycling and treatment module is used to realize the recycling of cleaning water resources, including a wastewater collection unit, a multi-stage treatment unit, and a water circulation unit.

7. A locomotive exterior surface cleaning system based on multimodal perception and intelligent decision-making according to claim 6, characterized in that, The three-dimensional laser scanner array consists of multiple high-precision three-dimensional laser scanners deployed on both sides and top of the cleaning station, forming a 360-degree scanning coverage without blind spots; the scanning speed is 1 million points / second, which can complete the data acquisition of the entire vehicle's outer surface within 60 seconds, generating high-density three-dimensional point cloud data with a point cloud accuracy of ±0.1mm; The multispectral imaging unit includes multiple multispectral cameras deployed corresponding to the 3D laser scanner. Each multispectral camera integrates visible and near-infrared wavelengths; the stain recognition accuracy is ≥90%. The UWB ultra-wideband positioning unit includes multiple UWB positioning base stations deployed along the track direction on the ground of the cleaning station and UWB positioning tags installed on the bottom of the locomotive; it guides the locomotive to stop accurately through the TOF positioning algorithm, with a positioning accuracy of ≤10cm.

8. A locomotive exterior surface cleaning method based on multimodal perception and intelligent decision-making according to claim 7, characterized in that, The multi-source feature fusion and cleaning heat map generation unit adopts point cloud registration technology based on RANSAC algorithm to spatially align three-dimensional laser point cloud data with multispectral images. Then, it uses a pixel-level fusion algorithm to bind the spatial coordinate information of three-dimensional laser point cloud with the stain type and severity information of multispectral images to generate a three-dimensional locomotive cleaning heat map with semantic information. The heatmap uses a 3D model of the locomotive as its base and visually reflects the severity of stains through color gradients. By binding color labels with weight values, it accurately indicates the specific type, 3D spatial location, and cleaning priority of stains on various parts of the locomotive's outer surface. Red corresponds to heavily soiled areas, orange to moderately soiled areas, yellow to lightly soiled areas, and green to clean areas. The weight values ​​range from 0 to 1, with heavy soiled areas having a weight value ≥ 0.8, moderately soiled areas 0.5-0.8, lightly soiled areas 0.2-0.5, and clean areas ≤ 0.

2. The adaptive 3D path planning algorithm unit, based on the generated cleaning heat map, employs an improved A... * The search algorithm performs path planning with "shortest total cleaning time" and "lowest total energy consumption" as dual objective functions. It introduces three major constraints: sensitive component avoidance, robotic arm movement range limitation, and collision-free collaboration between two robotic arms, and dynamically plans the optimal motion trajectory of the high-degree-of-freedom robotic arm unit. The multi-parameter collaborative optimization model unit embeds a locomotive cleaning expert knowledge base and a machine learning model based on a BP neural network.

9. A locomotive exterior surface cleaning method based on multimodal perception and intelligent decision-making according to claim 8, characterized in that, The high-degree-of-freedom robotic arm unit uses two seven-degree-of-freedom industrial-grade robotic arms. Through collaborative operation, its working range can completely cover the entire cleaning area from the undercarriage bogie to the highest point of the roof. The composite adaptive cleaning tool integrates a high-pressure water jet unit, a soft rotating brush unit, and a cleaning agent precision spraying unit into one unit, weighing ≤8kg; The composite adaptive cleaning tool head has a built-in pressure sensor and attitude sensor, which automatically adjusts the brush extension and water jet angle in complex curved areas. The composite adaptive cleaning tool's end vision sensor collects images of the cleaned area in real time and feeds them back to the intelligent decision-making module. If it detects residual dirt in a certain area, it automatically upgrades the cleaning mode, adjusts the path, and cleans again until it is qualified. After receiving instructions from the intelligent decision control module, the composite adaptive cleaning tool can automatically switch or combine multiple working modes, and the mode switching response time is ≤0.5 seconds.

10. A method for cleaning the outer surface of a locomotive based on multimodal perception and intelligent decision-making according to claim 9, characterized in that, The wastewater collection unit includes an anti-slip water collection grid and a diversion channel installed on the ground of the cleaning station. The anti-slip water collection grid has a load-bearing capacity of ≥10 tons / m². 2 The slope of the guide channel is ≥3°. The multi-stage treatment unit uses a combined physical-chemical treatment method to process the wastewater collected by the wastewater collection unit through a sedimentation tank, sand filter, oil-water separator, and precision filter in four stages. The water circulation unit includes a water quality testing module, a medium-sized water tank storage module, a reuse pipeline transportation module, and a flow control module. It performs water quality testing, standard storage, precise transportation, and flow control on the cleaning wastewater output from the multi-stage treatment unit. The pipeline of the reuse pipeline transport module is equipped with a pre-filter with a filtration accuracy of 1μm.

Images