Train type adaptive cleaning system based on multi-modal perception

Abstract

This invention belongs to the field of mechanical engineering, specifically relating to a train model adaptive cleaning system based on multimodal perception. This system automatically identifies the three-dimensional contours and contamination distribution of different train models by fusing laser point clouds, visible light images, infrared thermal imaging, and onboard communication information. It generates spatially matching cleaning paths and media parameters by combining 3D reconstruction and semantic annotation of key components. Through closed-loop feedback, it dynamically corrects the spray intensity, improving cleaning coverage and avoiding structural interference. It integrates water resource reuse and energy recovery mechanisms, significantly reducing water and energy consumption, achieving efficient, safe, and environmentally friendly fully automated train cleaning. For example, train braking energy recovery technology converts kinetic energy into electrical energy through regenerative braking, significantly improving energy conversion efficiency. Simultaneously, the equipment flushing water reuse system effectively replaces conventional water resources by accumulating and reusing water resources, achieving water conservation and energy reduction.

Description

Technical Field

[0001] This invention belongs to the field of mechanical engineering, specifically relating to a train model adaptive cleaning system based on multimodal perception. Background Technology

[0002] In the field of rail transit operation and maintenance, train cleaning, as a crucial link in ensuring the cleanliness of vehicle appearance, extending vehicle life, and enhancing operational image, has received widespread attention in recent years for its automation and intelligence levels. With the development of sensing technology, industrial automation, and artificial intelligence, the traditional fixed, manual-intervention-based cleaning model is gradually evolving towards integration and automation. Among these advancements, multimodal sensing technology, by integrating data from various sensors such as vision, lidar, and infrared thermography, not only improves environmental understanding and operational capabilities in adaptive robotic arms but also provides new technological pathways for environmental identification and dynamic response in complex industrial scenarios. Multimodal sensing technology demonstrates significant advantages in equipment status monitoring, obstacle detection, and process parameter adjustment, such as achieving comprehensive safety supervision in industrial safety production through real-time monitoring and data fusion technology.

[0003] Among them, the train model adaptive cleaning technology aims to automatically match the optimal cleaning process for different types and formations of trains without manual pre-setting. The core objective of this technology is to dynamically adjust the operating parameters and movement trajectory of the cleaning equipment based on the actual shape, door distribution, and degree of soiling of the train entering the station, ensuring comprehensive cleaning coverage and avoiding mechanical interference. This process requires the system to have high-precision train model recognition capabilities, real-time environmental perception capabilities, and rapid-response control logic, making it an important component of intelligent rail transit maintenance systems.

[0004] While existing technologies have achieved preliminary vehicle model identification and cleaning program invocation based on image recognition, they still have many limitations: the multi-source sensor data fusion mechanism is imperfect, resulting in poor identification stability in rain, snow, strong light, or low-light environments; the cleaning strategy relies on a pre-entered vehicle model template library, making it difficult to flexibly adapt to new train models or non-standard formations; information interaction between subsystems is lagging, causing delays in the perception-decision-execution chain; and the system fails to fully integrate the characteristics of the train surface material and the state of contamination for differentiated spray control, easily leading to water waste or insufficient cleaning. Furthermore, the system lacks online detection and safety avoidance capabilities for dynamic anomalies during the cleaning process (such as train deviation or protrusions encroaching on the limit), limiting its reliability and versatility under complex operating conditions. Therefore, there is an urgent need for a train cleaning system capable of multimodal perception-driven operation, with vehicle model self-learning capabilities and real-time adaptive control functions, to solve the current technical challenges of insufficient intelligence and weak environmental adaptability. Summary of the Invention

[0005] The purpose of this invention is to provide a train model adaptive cleaning system based on multimodal perception, to solve the technical problems of existing train cleaning equipment when dealing with multiple types, multiple train formations, and non-standard train shapes, such as low cleaning coverage, redundant spray paths, high risk of mechanical interference, and serious waste of water resources and chemical cleaning agents. In the current rail transit network, the trend of multiple train models, such as EMUs, metro cars, and intercity trains, operating on the same lines or sharing maintenance facilities is increasingly significant. Traditional fixed cleaning equipment relies on preset cleaning programs and can only adapt to a single or a few standard train model outlines, unable to dynamically identify the actual three-dimensional geometric features and key component distribution of the train to be cleaned. When the train model changes or undergoes partial modification, the cleaning device still operates according to the original trajectory, resulting in cleaning blind spots or over-washing of exposed precision components such as lights, sensors, and pantographs. Long-term accumulation will lead to equipment corrosion, sealing failure, and even functional malfunctions. Furthermore, existing systems have relatively crude control over the supply of cleaning media, failing to implement differentiated spraying based on the actual degree of contamination and material differences of the train body, resulting in low resource utilization efficiency.

[0006] The technical solution of this invention includes: a multimodal data fusion unit, used to simultaneously collect point cloud data, visible light image data, infrared thermal imaging data, and vehicle model identification information obtained through the vehicle communication interface of the train to be cleaned; a three-dimensional contour reconstruction unit, used to construct a high-precision digital twin model of the train's outer surface based on point cloud data, and to perform texture mapping and semantic annotation of key components in combination with visible light images; and a contaminated area identification unit, which, by analyzing visible light images and infrared thermal imaging data, and combining the reflectivity characteristics and temperature field distribution under the baseline clean state, can effectively locate and quantify the type, thickness, and adhesion intensity of pollutants in various areas of the vehicle body, similar to the technical application described in the simulation and analysis of pollutant concentration distribution characteristics in subway cars. The cleaning strategy generation unit is used to plan the activation sequence of non-uniform density cleaning nozzles, the motion trajectory of the robotic arm, and the parameter combination of the cleaning medium based on the 3D contour reconstruction results and the distribution map of the contaminated area. The adaptive execution unit is used to receive cleaning strategy instructions and drive the high-pressure water jet array, rotating brush group and chemical atomizing nozzle to work together to complete the precise cleaning task with spatial matching. The closed-loop feedback correction unit is used to collect the surface image and conductivity signal of the residual water film after cleaning in real time during the cleaning process, evaluate the cleaning effect, and dynamically adjust the cleaning intensity of subsequent sections.

[0007] Furthermore, the multimodal data fusion unit is deployed on both sides of the track and on the top gantry, and includes a two-dimensional laser scanner array, an industrial-grade CMOS camera group, an uncooled infrared focal plane detector, and a wireless radio frequency readout module. The laser scanner emits a fan-shaped beam at a sampling frequency of 100Hz, scanning along a direction perpendicular to the track, and can acquire high-density point cloud data of the track profile within a range of 0 to 4.5 meters in real time. The visible light camera adopts a global shutter mode, automatically triggering the supplementary lighting array when the ambient light is below 50 lux, ensuring that the image signal-to-noise ratio is not less than 36dB. The infrared thermal imaging unit has a spatial resolution of 320×240 pixels, can identify the minimum distance between two adjacent targets, and has a temperature measurement accuracy of ±1.5℃, which enables it to effectively detect abnormal thermal resistance areas caused by oil contamination. The onboard communication interface uses a train communication network protocol conforming to the IEC 61375 standard to parse the vehicle model, train formation number, and special protection area code broadcast by the TCMS system in real time.

[0008] As one embodiment of the present invention, the three-dimensional contour reconstruction unit performs the following processing flow: First, motion distortion compensation is performed on the original point cloud, and coordinate system one is realized using the pre-embedded coded reference targets in the track section; then, voxel mesh filtering is used for downsampling, and the point cloud density is compressed to 40% of the original data while retaining the curvature abrupt change features; then, a KD tree index is constructed through K-nearest neighbor search, and the moving least squares method is applied for surface smoothing reconstruction; finally, a triangular mesh model with normal vector attributes is output, and its spatial reconstruction error is less than ±3mm. The semantic annotation process of key components is implemented through a convolutional neural network, which is trained with 100,000 labeled samples and can identify 28 typical structures such as headlight components, air conditioner outdoor units, obstacle clearers, and antenna covers, and mark their bounding boxes and suggested avoidance distances on the three-dimensional model.

[0009] Furthermore, the contaminated area identification unit adopts a dual-channel feature fusion mechanism: the visible light channel extracts the saturation decay index and gray-level co-occurrence matrix contrast parameters in the HSV color space to identify dry pollutants such as soil and dust; the infrared channel calculates the surface heat capacity change rate of the same physical location before and after cleaning, and this parameter has a non-linear positive correlation with the oil film thickness; the two types of feature vectors are reduced in dimensionality by principal component analysis and then input into a support vector machine classifier to output five levels of contamination level judgment results, namely: level 0 (clean), level 1 (light dust), level 2 (moderate scale), level 3 (heavy oil), and level 4 (solidified adhesive).

[0010] In one embodiment of the present invention, the cleaning strategy generation unit constructs a three-dimensional spatial cleaning priority field function, which defines the required cleaning energy density for each surface micro-element with millimeter-level resolution. The input variables of the priority field include the local radius of curvature, the material pressure resistance threshold database index, the pollution level assessment value, and the historical cleaning frequency statistics. For concave areas with a radius of curvature of less than 150 mm, the jet pressure is automatically increased by 15% to overcome the water flow adhesion effect. For aluminum alloy skin areas, the maximum impact pressure is limited to no more than 8 MPa. For pollution areas of level 3 or above, a pulsed cavitation jet mode is activated, with a frequency set to 120 Hz and a duty cycle of 60%. All strategy parameters are solved using a dynamic programming algorithm to obtain the optimal solution sequence, ensuring that the total water consumption is minimized while meeting safety constraints.

[0011] Furthermore, the adaptive execution unit includes a three-degree-of-freedom servo robotic arm array, with 16 linearly arranged intelligent nozzle modules integrated at the end of each arm. Each nozzle module has a built-in micro solenoid valve, eddy current atomization chamber, and pressure sensor, with a response time of less than 20ms. The robotic arm's motion trajectory is generated using seventh-order spline interpolation to ensure acceleration continuity and avoid inertial impact. The nozzle activation sequence is linked to the train's speed to achieve synchronous spatial spraying. The chemical cleaning agent supply system selects the corresponding formula according to the type of pollution. Alkaline cleaning agents are used for grease decomposition, and acidic cleaning agents are used for mineral deposition removal. The amount of each spray is determined by the integral of the pollution area and thickness, with a control error within ±5%.

[0012] In one embodiment of the present invention, the closed-loop feedback correction unit is equipped with a binocular vision detection station and a contact conductivity probe array at the cleaning outlet. The binocular camera reconstructs the surface water film thickness distribution map after cleaning using a stereo matching algorithm. Combined with water film continuity detection, when a local residual water film exceeding 0.1 mm is detected and its duration is greater than 3 seconds, it is determined that the cleaning is incomplete. According to the WS310-2016 standard, the conductivity of purified water used for the final rinsing of instruments for cleaning and disinfection should be ≤15 μS / cm (25℃). The conductivity probe measures the ion concentration of the flowing water. If it is higher than this benchmark value, it indicates the presence of cleaning agent residue. The above abnormal signal triggers a local rewash mechanism, and the control system calls on the backup nozzle group to perform targeted additional rinsing of the area until the feedback signal returns to the normal range.

[0013] Furthermore, the system integrates an energy recovery subsystem for collecting cleaning wastewater and performing three-stage filtration: a first-stage cyclone sedimentation removes particulate matter, a second-stage ultrafiltration membrane retains colloidal substances, and a third-stage activated carbon adsorbs organic residues; the treated water quality meets the GB / T 19923-2005 standard for urban miscellaneous water use, with a reuse rate of not less than 75%; at the same time, the high-pressure pump set is equipped with an energy recovery device, which uses the pressure potential energy of the discharged wastewater to drive a hydraulic motor and reverse-assisted the operation of the water supply pump, reducing the overall energy consumption of the system by 22%.

[0014] In one embodiment of the present invention, the cleaning strategy generation unit is connected to the railway dispatch management system, obtains train entry plans and vehicle type information in advance, and loads the corresponding initial contour template and default cleaning parameters in advance; when the measured data deviates from the preset template by more than 10%, the full scan process is automatically started, and the new vehicle type features are stored in the local database for subsequent use; the system supports remote firmware upgrades and expert diagnosis access, and all operation logs and performance indicators are uploaded to the cloud monitoring platform for easy maintenance traceability and optimization iteration.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: This system integrates multi-source heterogeneous data, including laser point clouds, optical imaging, and onboard communication, to achieve fully automated, high-precision identification of the three-dimensional geometric features of non-cooperative target trains. It overcomes the limitations of traditional cleaning equipment that relies on manual settings or mechanical limits, enabling a single unit to be compatible with CRH series, urban rail A / B type trains, and various export models, increasing the vehicle type adaptability to more than three times that of the original system. The combination of three-dimensional contour reconstruction and semantic segmentation technology for key components constructs a spatial perception-driven cleaning and protection mechanism, effectively avoiding direct scouring of sensitive areas such as the pantograph sliding plate and driver's cab side windows, reducing the equipment damage accident rate to zero. The quantitative assessment of pollution levels and the graded response strategy transform the supply of cleaning media from experience-based, extensive control to evidence-driven, precise delivery. According to industry monitoring reports, the Chinese locomotive engine cleaning agent market will show a significant growth trend between 2025 and 2030, which is closely related to technological advancements and improved environmental standards in the train cleaning system industry. For example, the application of automated intelligent water-saving cleaning circulation systems achieves water conservation and improved cleaning efficiency through air-water mixing and water recycling, thereby significantly reducing operating costs and environmental impact. The closed-loop feedback correction mechanism introduces the concept of process quality control, ensuring the consistency and reliability of cleaning results through online detection and dynamic compensation, solving the problem of unstable cleaning quality caused by fluctuations in environmental temperature and humidity or differences in the aging characteristics of pollutants. The energy recovery and water resource recycling design aligns with the trend of green manufacturing development; the system's overall energy efficiency is 25% higher than similar equipment, and annual carbon emissions are reduced by approximately 18 tons, demonstrating significant economic and social benefits. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall technical solution architecture proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the integration of three-dimensional contour reconstruction and pollution area identification in this invention. Detailed Implementation

[0017] Please refer to Figure 1 and Figure 2 This invention provides a train model-adaptive cleaning system based on multimodal perception, aiming to achieve high-precision, low-damage, and resource-optimized automated cleaning of the outer surfaces of non-standardized, multi-formation, and heterogeneous trains. The system constructs a three-dimensional digital twin model of the train to be cleaned by fusing laser point cloud data, visible light images, infrared thermal imaging, and onboard communication data. Combined with quantitative analysis of contamination levels, it dynamically generates spatially matched cleaning paths and media supply strategies, significantly improving cleaning coverage and resource utilization efficiency while ensuring the safety of critical components. The system is deployed on a fixed gantry structure in railway maintenance bases or depot access areas, covering both sides and the overhead space of the tracks, forming a full-line coverage perception and execution network.

[0018] The system begins operation when the train enters the cleaning section at a constant speed of 2 to 8 kilometers per hour. At this time, the multimodal data fusion unit initiates a synchronous acquisition process to obtain multidimensional information streams of the train in motion. This unit consists of a two-dimensional laser scanner array distributed on the columns on both sides of the track, an industrial-grade CMOS camera group arranged on the top crossbeam, an uncooled infrared focal plane detector, and a wireless radio frequency readout module. The laser scanner uses a 905-nanometer wavelength semiconductor laser source to emit a fan-shaped beam perpendicular to the track direction, with a scanning angle range of ±75 degrees and a sampling frequency set to 100 Hz. A single scan can acquire profile point cloud data along the height direction of the car body from 0 meters to 4.5 meters above the track surface, with the point spacing controlled within 3 millimeters to ensure complete capture of the underside skirt, the area above the bogies, and the roof contour. All laser scanning equipment has been factory calibrated and spatially calibrated on-site using a three-axis attitude sensor. The rotation and translation parameters between its measurement coordinate system and the track reference coordinate system have been pre-determined and stored in the system configuration database.

[0019] Before the point cloud data is transmitted to the 3D contour reconstruction unit, motion distortion compensation processing is required. Because the train is in continuous motion, the original point cloud exhibits stretching or compression in space. To eliminate this effect, the system pre-embeds a set of coded optical reference targets on the track bed. The targets use a black and white checkerboard pattern, with clearly defined corner geometric features and high reflectivity. When the train passes, a visible light camera simultaneously captures images of the targets. A corner detection algorithm extracts their sub-pixel-level coordinates, and combined with known physical distances, the actual pose of the train relative to a fixed coordinate system at the current moment is deduced. This pose information serves as a key reference for timestamp alignment, projecting point cloud fragments collected at different times onto the same stationary coordinate system to complete spatiotemporal registration. The registered point cloud sequence enters the voxel mesh filtering stage. The system divides the 3D space into cubic voxel units with sides of 10 mm. Each voxel retains only one representative point closest to the sensor, thus preserving details in areas of abrupt curvature changes (such as the nose cone of the train and window frame corners) while compressing the original point cloud density to 40%, effectively reducing the computational load for subsequent processing.

[0020] The filtered point cloud is organized using a KD-tree index structure, supporting efficient K-nearest neighbor search operations. The system sets the search radius to 25 mm, querying at least 12 neighboring points within the neighborhood of each point to estimate local geometric features. Based on this, a moving least squares method is applied for surface fitting and reconstruction. This method assumes that the local surface can be approximated by a low-order polynomial function, and solves for the best-fitting plane or quadratic surface through weighted least squares optimization, with the weighting factor decreasing Gaussian with increasing distance. A normal vector consistency constraint is introduced during the fitting process to prevent spurious connections at edges or in hole regions. The final output is a triangular mesh model with vertex normal vector attributes. Its spatial reconstruction error, verified by a third-party laser tracker, has a mean of less than ±3 mm, meeting industrial-grade modeling accuracy requirements. The mesh model is further processed using hole-filling and topology repair algorithms to ensure no open boundaries or self-intersecting surfaces, forming a closed, watertight digital shell.

[0021] Meanwhile, visible light image data is simultaneously captured by six global shutter CMOS cameras deployed on the top and sides of the gantry. The cameras have a resolution of 2048×1536 pixels, a frame rate of 30 frames per second, and lens focal lengths of 12mm, 25mm, and 50mm respectively, depending on the field of view requirements, covering multi-level imaging of the entire vehicle from macroscopic outlines to local details. Image acquisition is controlled by ambient lighting conditions. The system's built-in illuminance sensor monitors the surrounding brightness in real time. When the brightness drops below 50 lux, the LED supplementary lighting array is automatically activated, with the light source color temperature set to 5500 Kelvin and the total luminous flux not less than 8000 lumens, ensuring clear images with a signal-to-noise ratio of not less than 36 dB are obtained even in low-light scenarios such as at night or at tunnel exits. All image data is stored in RAW format to avoid artifacts introduced by JPEG compression, providing a high-quality input source for subsequent texture mapping and contamination identification.

[0022] The texture mapping process attaches a high-resolution color image to the surface of the previously generated triangular mesh model according to perspective projection. The system employs a spatial registration technique based on feature point matching. First, SIFT feature descriptors are extracted from the image, and corresponding points are searched in the 3D model projection image to establish a mapping relationship between pixel coordinates and 3D vertices. For areas with occlusion or blind spots (such as the equipment compartment under the vehicle), the system calls a multi-view stereo vision fusion algorithm to perform depth completion using the overlapping areas of adjacent camera views, ultimately generating a continuous and seamless color texture map of the entire vehicle body with a resolution of over 16 pixels per square centimeter.

[0023] After texture mapping is completed, the semantic annotation module for key components initiates the recognition process. This module is implemented based on a deep convolutional neural network architecture, with the backbone network using a ResNet-50 structure. After pre-training on the ImageNet dataset, fine-tuning the model parameters achieved a training accuracy of 93.8% and a validation accuracy of 93.44%. The total number of training samples exceeded 100,000, covering the CR400AF / BF China Standard EMU, CRH series, urban rail A / B type trains, and various export models. The annotation categories include 28 typical structures such as headlight assemblies, air conditioning outdoor units, obstacle clearers, antenna covers, pantograph bases, driver's cab side windows, and brake cylinder covers. The annotation format is a dual output of instance segmentation masks and bounding boxes, ensuring positioning accuracy and classification reliability. During the network inference stage, the input is multi-view texture image patches of the vehicle to be processed, and the output is the three-dimensional spatial position, orientation, and suggested avoidance distance of each component. For example, for the pantograph sliding plate area, the system marks its 3D bounding box and sets a minimum safe distance of 150 mm; for the driver's cab windshield, the maximum permissible impact pressure threshold is defined as 6 MPa, and direct contact with the rotating brush assembly is prohibited. All semantic tags are embedded in the 3D model file in the form of metadata for use by the cleaning strategy generation unit.

[0024] Infrared thermal imaging data is acquired by an uncooled infrared focal plane detector, which operates at room temperature without cryogenic cooling. This detector offers advantages such as low cost, small size, light weight, low energy consumption, long lifespan, cost savings, and rapid response. The device has a resolution of 320×240 pixels, a spectral response range of 8 to 14 micrometers, a temperature measurement accuracy of ±1.5 degrees Celsius at 20 degrees Celsius, and a thermal sensitivity of less than 50 milliklvin. The detector is mounted on the side of the gantry, avoiding direct sunlight and interference from high-temperature equipment to ensure measurement accuracy. The system records the ambient background temperature field before the train enters, serving as a reference. As the train passes, it continuously acquires infrared radiation images of the train's surface, focusing on monitoring areas with abnormal thermal resistance caused by oil contamination. Clean metal surfaces have high emissivity (approximately 0.92), while oil films significantly reduce local emissivity and alter thermal conductivity, resulting in surface temperatures higher than surrounding areas under the same operating conditions. The system identifies suspected contamination areas with a temperature difference exceeding 3 degrees Celsius by comparing the measured temperature field with a standard thermal distribution model preset based on vehicle type, and maps the coordinates of these areas to the corresponding positions in the three-dimensional model, serving as the initial input clues for the contamination area identification unit.

[0025] The onboard communication interface connects to the Train Communication Network (TCN) system via wireless radio frequency, adhering to the IEC 61375 standard protocol specification. The operating frequency band is 2.4 GHz, and the communication range covers the entire cleaning section. The IEC 61375 standard defines the communication requirements of the Train Communication Network (TCN), ensuring that equipment from different manufacturers can operate under a unified communication architecture, achieving seamless integration. This standard not only specifies the protocols for hardware and software interfaces but also includes aspects such as data communication, network management, and fault handling. The system actively initiates connection requests and receives vehicle model codes, train formation numbers, traction unit locations, and special protection zone instructions broadcast by the TCMS. For example, a CRH380B EMU reports that a precision weather sensor is installed on the roof of its third carriage, requiring the high-pressure spray to be disabled during cleaning. This information is parsed into structured data fields, including carriage number, equipment type, three-dimensional coordinate offset, and protection level, and written to the system's operating context database in real time. If wireless communication fails, a backup plan is activated: the system automatically matches the most similar vehicle model record in the local vehicle model template library based on the three-dimensional contour reconstruction results and prompts for manual confirmation.

[0026] The four types of data—a 3D geometric model constructed from point clouds, visible light texture images, infrared temperature anomaly maps, and vehicle communication identification information—are spatiotemporally aligned and semantically associated within the multimodal data fusion unit. The alignment mechanism employs a joint registration algorithm based on timestamps and pose interpolation. The system records the acquisition time of each frame of point cloud, image, and infrared data, and performs linear interpolation based on the train's speed to ensure that data from different sensors accurately correspond to the same vehicle body cross-section in space. Semantic association is achieved through coordinate transformation, converting the coordinates of the protected area provided by vehicle communication from the vehicle's local coordinate system to the global cleaning system coordinate system, and superimposing them onto the 3D model. The final output is a multidimensional digital twin integrating geometry, material, temperature, functional attributes, and operational constraints, serving as the foundational data source for subsequent cleaning decisions.

[0027] The contaminated area identification unit receives the integrated data stream from the multimodal data fusion unit and performs pollutant type identification and level assessment tasks. The identification process employs a dual-channel feature fusion mechanism, processing information from the visible light and infrared channels separately. The visible light channel focuses on analyzing visual degradation features in the HSV color space. The system extracts the saturation decay index for each area of ​​the vehicle body, defined as the ratio of the current pixel saturation value to the baseline saturation under clean conditions for the same vehicle model. When the index is below 0.7, it is determined that there are obvious colored pollutants. Simultaneously, the contrast parameter of the gray-level co-occurrence matrix is ​​calculated to reflect local texture roughness changes, used to distinguish between floating dust (low contrast) and solidified dirt (high contrast). These two types of features constitute the first set of input vectors.

[0028] The infrared channel focuses on differences in thermodynamic response. The system collects the steady-state temperature distribution of the vehicle surface before cleaning and establishes expected temperature benchmarks for each area under simulated conditions. In actual measurements, if the surface temperature of a certain area is more than 2 degrees Celsius higher than the benchmark and lasts for more than 10 seconds, it is determined that there is a coating with degraded thermal conductivity. Furthermore, the system calculates the rate of change of surface heat capacity at that location after a short heating pulse (such as sunlight or equipment heating), defined as the ratio of the temperature rise slope to the input heat. Experimental data shows that this parameter has a non-linear positive correlation with oil film thickness, which is consistent with the experimental principle of estimating the size of oleic acid molecules using the oil film method. The measurement method and calculation of oil film thickness can provide an empirical formula for estimating contaminant thickness. For example, on aluminum alloy panels, when the rate of change of heat capacity decreases to 60% of the clean state, the corresponding average oil film thickness is approximately 0.15 mm.

[0029] The two types of feature vectors are normalized and then concatenated into a joint feature matrix with a dimension of 1×N, where N is the total number of features. To eliminate redundancy and improve classification efficiency, the system uses principal component analysis for dimensionality reduction, selecting the top M principal components with a cumulative contribution rate exceeding 95% as the final feature input. The dimensionality-reduced data is then fed into a support vector machine classifier. The kernel function is a radial basis function, the penalty coefficient C is set to 1.0, and the gamma parameter γ is set to 0.1. The parameter combination is optimized through cross-validation. The classifier outputs a five-level pollution level judgment result: Level 0 indicates a clean surface, requiring no additional treatment; Level 1 indicates light dust, requiring only low-pressure rinsing; Level 2 indicates moderate scale buildup, requiring increased water flow shear force; Level 3 indicates heavy oil stains, requiring chemical decomposition; Level 4 indicates solidified adhesives (such as bird droppings or asphalt drips), requiring cavitation jet stripping.

[0030] The classification results are overlaid onto the surface of the 3D model in a color-coded format to form a heat map of pollution distribution. Each triangular facet is assigned a pollution level label, which, along with the pollutant type (dust, oil stains, mineral deposits, etc.) and estimated thickness value, constitutes the input elements of the cleaning strategy generation unit. The system also maintains a historical cleaning frequency statistics table, recording the number of times each vehicle or similar vehicle type has been cleaned in the past 30 days, which is used to adjust the energy density baseline value for this cleaning. For example, trains that frequently operate in coastal salt spray environments will automatically increase the treatment intensity by one level, even if the current pollution level is level 2, to prevent corrosion accumulation.

[0031] The cleaning strategy generation unit constructs a three-dimensional spatial cleaning priority field function based on the 3D contour reconstruction results and the contaminated area distribution map. This function defines the required cleaning energy density for each micro-element on the vehicle surface at millimeter-level resolution, in joules per square centimeter. The input variables of the priority field include the local radius of curvature, the material pressure resistance threshold database index, the contamination level assessment value, and the historical cleaning frequency correction coefficient. The system discretizes the entire vehicle surface into square grid cells with a side length of 5 mm, and each cell independently calculates its target cleaning parameters.

[0032] The local radius of curvature is calculated using the vertex normal vector field of the 3D model. The system employs discrete differential geometry to fit quadratic surfaces to the neighborhood triangular patches of each grid center point, extracting the principal curvatures k1 and k2, with a radius of curvature r = 1 / max(|k1|, |k2|). When r is less than 150 mm, it can be identified as a strongly concave region (such as the inner cavity of a car headlight or the recess of a nameplate). Such regions are prone to water vortices and cavitation shielding, leading to cleaning dead zones. Therefore, the system automatically increases the spray pressure of this unit by 15% and compensates for flow field unevenness by increasing the nozzle residence time by 20%. For flat areas with r greater than 800 mm, the system appropriately reduces the pressure to save energy.

[0033] The material pressure resistance threshold database stores the mechanical strength parameters of common car body materials. For example, aluminum alloy 6005A-T6 is a medium-strength heat-treatable alloy with excellent corrosion resistance, widely used in the manufacture of high-speed trains and subway train bodies. Its yield strength and tensile strength parameters are recorded in detail. Stainless steel SUS301L, due to its wear resistance, impact resistance, and corrosion resistance, is used in the manufacture of lightweight urban rail vehicle bodies, and its tensile strength and other performance parameters are also accurately measured. The database is automatically loaded by vehicle model identification and supports manual correction. When the system identifies an area as aluminum alloy skin, it will forcibly limit the maximum impact pressure to no more than 8 MPa to prevent plastic deformation; if it is a composite material skirt, a flexible brush group priority contact mode is activated to avoid the hard bristles scratching the surface.

[0034] The pollution level assessment directly determines the setting of the basic cleaning energy. Level 0 areas do not activate any nozzles; Level 1 areas use 3 MPa water pressure with a flow rate of 0.5 liters per minute; Level 2 areas increase to 5 MPa and 1.2 liters per minute; Level 3 areas activate a pulsed cavitation jet mode with a frequency set at 120 Hz and a duty cycle of 60%, utilizing the micro-jet impact effect generated by the periodic interruption of high-speed water flow to enhance oil removal; Level 4 areas are further enhanced with a chemical pre-soaking step, first spraying a special swelling agent to soften the contaminants, then supplemented by mechanical removal using a high-frequency vibrating brush assembly.

[0035] All strategy parameters are solved using a dynamic programming algorithm to find the optimal solution sequence. In the multi-stage decision model for optimizing the cleaning process, the state space is defined as the set of remaining untreated areas, while the action space involves the combinations of nozzles that can be activated in the next moment and the pose of the robotic arm. The cost function of this model comprehensively considers factors such as water consumption, chemical dosage, operation time, and equipment wear to improve processing efficiency and reduce costs. The algorithm uses a reverse recursive approach, backtracking from the endpoint to the starting point to find the state transition path that minimizes the total cost. Constraints include: a maximum concurrent number of nozzles of no more than 128, a robotic arm acceleration limited to within 3 meters per second squared, and an activation interval of no less than 50 milliseconds between adjacent nozzles to prevent electromagnetic interference. The final output is a spatiotemporally linked cleaning command sequence, updated precisely every 50 milliseconds to update the nozzle on / off state and the target coordinates of the robotic arm.

[0036] The adaptive execution unit receives cleaning strategy instructions and drives the high-pressure water jet array, rotating brush assembly, and chemical atomizing nozzles to coordinate their movements, completing a precise cleaning operation with spatial matching. The core of the execution unit is a three-degree-of-freedom servo robotic arm array, consisting of six sets located on both sides and the top of the track. Each set has an arm span of 2.8 meters and a repeatability accuracy better than ±0.1 mm. The robotic arm joints are driven by AC permanent magnet synchronous motors and equipped with high-resolution photoelectric encoders and harmonic reducers to ensure smooth movement and rapid response. An intelligent nozzle module array is connected to the arm end flange. This array consists of 16 nozzles arranged linearly with a nozzle spacing of 150 mm, covering a width of 2.4 meters, capable of cleaning an entire side wall of the carriage in one operation.

[0037] Each intelligent nozzle module is an independently controllable unit, incorporating a miniature solenoid valve, a vortex atomization chamber, a pressure sensor, and a temperature compensation circuit. The solenoid valve has a response time of less than 20 milliseconds and supports PWM pulse width modulation control, enabling continuous flow regulation. The vortex atomization chamber features a spiral guide channel, causing the liquid to form a conical atomization angle under centrifugal force. The spray angle can be electrically adjusted between 30 and 110 degrees to accommodate surfaces with varying curvatures. The pressure sensor has a range of 0 to 15 MPa and an accuracy class of 0.5, providing real-time feedback of the outlet pressure value to the control system for closed-loop pressure regulation. The module housing is made of corrosion-resistant 316L stainless steel, and the internal flow channels are mirror-polished to prevent impurity deposition.

[0038] The robotic arm's trajectory is generated by seventh-order spline interpolation, ensuring continuous position, velocity, acceleration, and jerk throughout the entire process, avoiding mechanical resonance or water hammer effects caused by abrupt changes. Trajectory planning takes into account the train's speed; the system monitors the train speed in real time via an encoder and dynamically adjusts the robotic arm's lateral scanning rate within a ±10% fluctuation range to achieve synchronized spatial spraying. For example, when the train speed is 5 km / h, the robotic arm's horizontal movement speed is matched to 1.39 m / s, keeping the nozzles stationary relative to the train body and ensuring that each micro-element receives a constant dose of cleaning energy.

[0039] The chemical cleaning agent supply system selects the appropriate formula based on the type of contamination. The system is equipped with three independent storage tanks, respectively containing an alkaline cleaning agent (pH 12.5, containing sodium hydroxide, sodium carbonate, and other alkaline substances, suitable for the saponification and decomposition of greases), an acidic cleaning agent (pH 2.0, mainly composed of hydrochloric acid and sulfuric acid, used to remove rust and mineral deposits), and a neutral protective agent (containing siloxane film-forming substances, used for rust prevention after cleaning). The selection logic is driven by the output of the contamination area identification unit: when the contaminant is determined to be grease, the alkaline agent is used; for minerals, the acidic agent is used; for mixed types, the two are mixed in a specific ratio. The spray volume for each application is determined by the integral of the contamination area and thickness, calculated using the following formula: in, Total spray volume, in liters; For the first The area of ​​each contaminated grid, in square meters; This represents the equivalent thickness of pollutants within the grid, in meters. This is an empirical coefficient, set to 1.2, used to compensate for atomization loss and wind dispersion. This represents the total number of contaminated grids. The control system monitors the output in real time through a mass flow meter and adjusts the metering pump speed in a closed loop to keep the actual spray volume control error within ±5%.

[0040] During the cleaning process, the closed-loop feedback correction unit initiates a quality assessment process at the outlet. This unit includes a binocular vision inspection station and a contact conductivity probe array. The binocular camera has a resolution of 1920×1080 pixels, a baseline distance of 600 mm, and uses active structured light-assisted illumination to project random speckle patterns to enhance the matching ability in areas with weak texture. The system reconstructs the surface water film thickness distribution map after cleaning using a stereo matching algorithm, with a matching window size of 9×9 pixels, a parallax search range of 64 pixels, and a sub-pixel interpolation accuracy of 1 / 8 pixel. When a local residual water film exceeds 0.1 mm and lasts for more than 3 seconds, it is determined that the cleaning is incomplete, possibly due to the water flow failing to sufficiently remove hydrophobic contaminants.

[0041] The conductivity probe array consists of 12 sets of stainless steel probes arranged in a grid pattern at a height of 1.2 meters above the track, making slight contact with the flowing water on the roof surface. The probe spacing is 400 mm, covering a width of 4.4 meters. Each probe set has a built-in constant current source excitation circuit and a high-gain amplifier to measure the conductivity of flowing water, with a range of 0 to 2000 microsiemens per centimeter and a resolution of 1 microsiemens per centimeter. The system's set reference value is 150 microsiemens per centimeter. If the measured value remains above this threshold for more than 5 seconds, it indicates a risk of cleaning agent residue, which may lead to subsequent corrosion problems.

[0042] The two types of abnormal signals mentioned above trigger a local rewash mechanism. The control system immediately calls upon the backup nozzle group to perform targeted additional flushing of the area. The backup nozzles are located 1.5 meters behind the main array, offering greater flexibility and allowing for individual adjustment of the spray angle and pressure. For areas with residual water film, a low-pressure fan-shaped spray (2 MPa, 90-degree diffusion angle) is activated to drive away the accumulated water using shear water flow; for areas with residual cleaning agent, the system switches to pure water rinsing mode, with the duration dynamically extended according to the degree of exceedance until the conductivity returns to a safe range. The rewash process is also subject to closed-loop monitoring, performing a maximum of three rounds of compensation operations. If the standard is still not met, an alarm is issued and the event log is recorded.

[0043] The energy recovery subsystem is integrated at the back end of the cleaning station, responsible for wastewater collection and reuse. Cleaning wastewater is collected via a ditch and flows to a primary sedimentation tank for first-stage cyclone sedimentation treatment. The tank is equipped with a guide tube and a conical separation chamber, utilizing centrifugal force to rapidly settle particles such as sand and metal shavings with a diameter greater than 50 micrometers, achieving a removal rate of up to 85%. The supernatant flows into a secondary ultrafiltration unit, employing hollow fiber membrane modules with a pore size of 0.02 micrometers and an operating pressure of 0.15 MPa, effectively trapping colloidal substances, bacteria, and large organic molecules, resulting in an effluent turbidity of less than 1 NTU. Finally, it enters a tertiary activated carbon adsorption tower, filled with coconut shell-based granular activated carbon with an iodine value of not less than 900 mg / g and a residence time of 15 minutes, adsorbing residual organic matter and odor molecules.

[0044] Through real-time and continuous monitoring by online instruments, the treated water quality meets the GB / T 19923-2005 standard for urban miscellaneous water use. These instruments can monitor key water quality indicators in real time, such as suspended solids, chemical oxygen demand, petroleum hydrocarbons, and pH value, ensuring that these indicators do not exceed 10 mg / L, 50 mg / L, 5 mg / L, and 6.5 to 9.0, respectively. The treated water is stored in a recycled water tank and then resupplyed to the front end of the cleaning system via a variable frequency pump, achieving a measured reuse rate of over 75%. The high-pressure pump set is equipped with an energy recovery device, employing a hydraulic turbine structure. It utilizes the pressure potential energy of the discharged wastewater to drive a hydraulic motor, which in turn assists the main water supply pump, achieving an energy recovery efficiency of 38% and reducing the overall system energy consumption by 22%.

[0045] The system is networked with the railway dispatching and management system, receiving train arrival plans via TCP / IP protocol. This information includes train number, estimated arrival time, train type, number of trains, and affiliated line. The control system pre-loads the corresponding initial profile template and default cleaning parameters based on this information, shortening preparation time. When the measured data deviates from the preset template by more than 10% (e.g., during trial operation of a new model or temporary modification), a full scan process is automatically initiated, and the new model's characteristics are stored in the local database for later retrieval. The system supports remote firmware upgrades, downloading update packages via an encrypted tunnel, and employing a dual-partition boot mechanism to ensure rollback in case of upgrade failure. Expert diagnostic terminals can access the system through authorized authentication to view real-time operating data and historical logs. All operation records, performance indicators, and fault alarms are uploaded to a cloud monitoring platform, with a storage period of no less than 3 years, facilitating maintenance traceability and optimization iteration.

[0046] This embodiment achieves a fully closed-loop intelligent cleaning process from perception, modeling, decision-making to execution and feedback through the close collaboration of the aforementioned units. The system no longer relies on preset programs but dynamically generates cleaning strategies based on real-world data, fundamentally solving the adaptation problem under mixed-vehicle traffic conditions. Three-dimensional contour reconstruction and semantic segmentation technology for key components construct a spatial perception-driven protection mechanism, effectively avoiding direct scouring of sensitive areas and reducing the equipment damage accident rate to zero. Through quantitative assessment of contamination levels and a graded response strategy, the supply of cleaning media has shifted from experience-based extensive control to evidence-driven precise delivery. For example, in the railway transportation sector, research on the upgraded washing effect of EMU exterior cleaning machines shows that water consumption per unit train has decreased by 38%, and cleaning agent consumption has decreased by 42%, significantly improving cleaning efficiency and reducing resource consumption. The closed-loop feedback correction mechanism introduces the concept of process quality control, ensuring the consistency of cleaning results through online detection and dynamic compensation. The energy recovery and water resource recycling design aligns with the trend of green manufacturing development; the system's overall energy efficiency is 25% higher than similar equipment, and annual carbon emissions are reduced by approximately 18 tons, demonstrating significant economic and social benefits. For example, the design of railway washing wastewater treatment and reuse projects has achieved the recycling of washing wastewater by optimizing process operation parameters, which not only improves economic efficiency but also reduces environmental impact.

[0047] Current train cleaning equipment generally uses fixed spray arrays and pre-programmed robotic arm trajectories, making it suitable only for a single train model. When the train model changes, manual re-teaching of the path or replacement of the mold is required, leading to low efficiency and a high risk of errors. While some high-end equipment uses laser ranging for rough contour detection, it lacks semantic understanding and cannot identify functional components or perform differentiated processing. Cleaning medium control relies entirely on timers or simple photoelectric switches, failing to consider actual contamination levels and resulting in significant resource waste. Furthermore, no system possesses process quality feedback and dynamic compensation capabilities; cleaning results depend entirely on post-cleaning manual sampling, leading to large quality fluctuations.

[0048] The core breakthrough of this solution lies in constructing a technological chain encompassing multimodal perception, digital twin, intelligent decision-making, and closed-loop execution. By integrating four-dimensional data from laser, optics, infrared, and communication, a three-dimensional model with both geometric accuracy and semantic information is generated, providing a cognitive foundation for refined operations. The contamination identification unit integrates visual and thermodynamic features, achieving a leap from "seeing" to "identifying." The cleaning strategy generation unit introduces priority field functions and dynamic programming algorithms, transforming complex cleaning tasks into computable optimization problems. The intelligent nozzle module array of the adaptive execution unit achieves media delivery control with millimeter-level spatial resolution. The closed-loop feedback correction unit establishes a "detection-evaluation-compensation" quality assurance loop, enabling the system to self-regulate. The energy recovery subsystem further extends the sustainability dimension of the technological chain, forming a complete green cleaning solution.

[0049] in, The overall energy efficiency improvement coefficient of the system; The water treatment reuse rate is set at 0.75. The pressure energy recovery efficiency is set to 0.38. The transmission loss rate during the pipeline and pumping process is taken as 0.12. Substituting this value into the calculation yields... =0.75×0.38×(1−0.12)=0.25, which means the overall energy efficiency of the system is improved by 25%, consistent with the measured data.

[0050] In this embodiment, the quantitative model of the system's overall energy efficiency is expressed mathematically. (Left side of the formula) Representing the net energy efficiency gain from resource recycling, it is a core indicator for measuring green performance. The three items on the right show the product relationship between water treatment reuse efficiency, pressure energy recovery efficiency, and inherent system losses, respectively. The removal efficiency is determined by the three-stage treatment process: cyclone sedimentation removes 85% of suspended solids, ultrafiltration removes 92% of colloids, activated carbon removes 78% of COD, and the overall water reuse rate reaches 75%. The efficiency is determined by the mechanical efficiency of the hydraulic turbine (80%) and the overall efficiency of the energy conversion chain (hydraulic-mechanical-electric) (47.5%), with a measured value of 38%. Including factors such as pipeline frictional resistance, valve throttling, and pump internal leakage, the value was determined to be 12% through fluid dynamics simulation and field calibration. This formula not only verifies the composition of the energy-saving effect but also points the way for subsequent optimization: increasing the ultrafiltration membrane flux can improve... Improving turbine blade design can enhance Optimizing pipeline layout can reduce .

[0051] The technical solution described in this embodiment has been deployed and continuously tested in a high-speed rail depot. The test subjects covered four main train models: CR400AF, CRH2A, CRH6F, and Metro Type B trains, with a cumulative cleaning of over 1200 trains. Actual test data shows that the average cleaning time per train remained stable at 28 minutes, a 14% reduction compared to traditional equipment; the cleaning blind spot area decreased from 9.7% to below 0.3%; the number of accidental contact incidents with the pantograph area caused by high-pressure water jets decreased from an average of 3 per quarter to 0; the water consumption per train decreased from an average of 1.8 cubic meters to 1.12 cubic meters, a reduction of 38%; cleaning agent consumption decreased from 0.95 kg per train to 0.55 kg, a reduction of 42%; the wastewater reuse rate remained stable between 76% and 78%; and the system achieved an average annual power saving of 67,000 kWh, equivalent to a reduction of 18.2 tons of carbon emissions. All performance indicators met or exceeded the design targets, verifying the engineering feasibility and practical value of this invention.

[0052] The system also demonstrates good robustness under extreme conditions. In low-temperature winter conditions (ambient temperature -15℃), the heating system automatically activates to maintain the water supply pipe temperature above 5℃, preventing freezing and blockage; electric heating tape is installed at the nozzle outlet to ensure normal atomization. During sandstorms, the pre-air filter automatically switches to high-resistance mode to protect the optical lens; the infrared detector uses a dynamic background suppression algorithm to eliminate the influence of suspended particles on temperature measurement. In the event of a communication interruption, the system degrades to a fully autonomous mode, relying solely on laser and image data to complete vehicle identification and cleaning operations. Although the protection accuracy decreases slightly, the basic cleaning functions can still be safely operated.

Claims

1. A train model adaptive cleaning system based on multimodal perception, characterized in that, include: The multimodal data fusion unit is used to simultaneously collect point cloud data, visible light image data, infrared thermal imaging data, and vehicle model identification information provided by the vehicle communication interface of the train to be cleaned. A three-dimensional contour reconstruction unit is used to construct a digital twin model of the train's outer surface based on the point cloud data, and to perform texture mapping and semantic annotation of key components in conjunction with the visible light image data. The pollution area identification unit is used to analyze the visible light image data and infrared thermal imaging data to locate and quantify the type and adhesion intensity of pollutants in various areas of the vehicle body. The cleaning strategy generation unit is used to plan the activation sequence of non-uniform density cleaning nozzles, the motion trajectory of the robotic arm, and the combination of cleaning medium parameters based on the digital twin model and the pollution area distribution map. An adaptive execution unit is used to receive the cleaning strategy instructions and drive the high-pressure water jet array, rotating brush group and chemical atomizing nozzle to work together to complete a precise cleaning operation with spatial matching. The closed-loop feedback correction unit is used to acquire surface images and residual water film conductivity signals in real time during the cleaning process, evaluate the cleaning effect, and dynamically adjust the cleaning intensity of subsequent sections.

2. The train model adaptive cleaning system based on multimodal perception according to claim 1, characterized in that, The multimodal data fusion unit is deployed on both sides of the track and on the top gantry, and includes a two-dimensional laser scanner array, an industrial-grade CMOS camera group, an uncooled infrared focal plane detector, and a wireless radio frequency readout module.

3. The train model adaptive cleaning system based on multimodal perception according to claim 2, characterized in that, The three-dimensional contour reconstruction unit includes: The motion distortion compensation subunit is used to compensate for motion distortion of the original point cloud using coded reference targets pre-embedded in the track section. The surface smoothing reconstruction sub-unit is used to downsample and fit the surface of the compensated point cloud, and output a triangular mesh model with normal vector properties. The key component semantic annotation subunit is used to identify typical structures of headlamp components, air conditioner outdoor units, obstacle clearers, and antenna covers through convolutional neural networks, and to mark their bounding boxes and suggested avoidance distances on the triangular mesh model.

4. The train model adaptive cleaning system based on multimodal perception according to claim 3, characterized in that, The contaminated area identification unit includes: The visible light feature extraction subunit is used to extract the color space saturation decay index and gray-level co-occurrence matrix contrast parameters from the visible light image data. An infrared feature extraction subunit is used to calculate the surface heat capacity change rate from the infrared thermal imaging data; The pollution level classification subunit is used to fuse the visible light features and infrared features and input them into the classifier to output multi-level pollution level determination results.

5. The train model adaptive cleaning system based on multimodal perception according to claim 4, characterized in that, The cleaning strategy generation unit includes: The cleaning priority field construction sub-unit is used to define the required cleaning energy density for each micro-element on the vehicle body surface at millimeter-level resolution. Its input variables include local radius of curvature, material pressure resistance threshold database index, contamination level assessment value, and historical cleaning frequency statistics. The optimal solution sequence solving subunit is used to solve the cleaning nozzle activation sequence, robotic arm motion trajectory, and cleaning medium parameter combination that minimizes the total water consumption under the condition of satisfying safety constraints using a dynamic programming algorithm.

6. The train model adaptive cleaning system based on multimodal perception according to claim 5, characterized in that, The adaptive execution unit includes: A three-degree-of-freedom servo robotic arm array, with multiple intelligent nozzle modules linearly arranged at the end of the arm; Each of the aforementioned smart nozzle modules incorporates a miniature solenoid valve, a vortex atomization chamber, and a pressure sensor. The chemical cleaning agent supply subsystem is used to select the corresponding formula according to the type of pollution and determine the spraying amount based on the integral of the pollution area and thickness.

7. The train model adaptive cleaning system based on multimodal perception according to claim 6, characterized in that, The closed-loop feedback correction unit includes: Binocular vision inspection station is used to reconstruct the thickness distribution map of the water film on the surface after cleaning using a stereo matching algorithm; A contact conductivity probe array is used to measure the ion concentration of flowing water. The local flushing trigger subunit is used to call up the backup nozzle group to perform targeted additional flushing on the area when the water film thickness or ion concentration exceeds a preset threshold.

8. The train model adaptive cleaning system based on multimodal perception according to claim 7, characterized in that, It also includes an energy recovery subsystem for collecting cleaning wastewater and performing multi-stage filtration treatment. The treated water meets the standards for urban miscellaneous water use and is reused for cleaning operations. The energy recovery subsystem is also equipped with an energy recovery device that uses the pressure potential energy of the discharged wastewater to assist the operation of the water supply pump.

9. The train model adaptive cleaning system based on multimodal perception according to claim 8, characterized in that, The cleaning strategy generation unit is connected to the railway dispatch management system to obtain train arrival plans and vehicle type information in advance, and to preload the corresponding initial contour template and default cleaning parameters. When the deviation between the measured data and the preset template exceeds a preset threshold, the system will automatically start the full scan process and store the new vehicle type features in the local database. For example, the patent "Data Processing Method and Device" applied for by Beijing Urban Construction Intelligent Control Technology Co., Ltd. involves a train automatic driving system, which predicts and adjusts train operation by obtaining train speed information and historical control data. This is consistent with the train information processing logic in our system.

10. The train model adaptive cleaning system based on multimodal perception according to claim 9, characterized in that, The onboard communication interface uses a train communication network protocol conforming to the IEC 61375 standard to parse the vehicle model, train formation quantity, and special protection zone code broadcast by the TCMS system in real time. If communication fails, it automatically matches the most similar vehicle model record in the local vehicle model template library based on the three-dimensional contour reconstruction results.

Images