Intelligent carriage cleaning method

Abstract

The present application provides an intelligent carriage cleaning method, comprising: matching and selecting appropriate cleaning tools from a predefined tool library according to the characteristics of the cleaning area and the material properties, and realizing automatic assembly of the tools through a robotic arm to improve cleaning adaptability; for each cleaning area, using a reinforcement learning algorithm, combined with the internal environmental constraints of the carriage and the cleaning efficiency goals, learning and optimizing the motion trajectory and operating parameters of the cleaning tool, and formulating a cleaning path; during the cleaning process, using a visual sensor to obtain the distribution changes of residual materials in real time, dynamically updating the material distribution map through target detection and semantic segmentation algorithms, and adjusting the cleaning path in real time according to the updated distribution map; for materials with a high degree of adhesion, using a force feedback sensor to monitor the contact force between the cleaning tool and the material in real time, and dynamically adjusting the pressure and angle of the cleaning tool in combination with a pre-established material property database until the material is cleaned.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to an intelligent carriage cleaning method. Background Art

[0002] In the complex interior environment of train cars, the distribution and adhesion of residual materials vary greatly, posing a significant challenge to the adaptive capabilities of cleaning equipment. The complex and varied interior structure of train cars, with numerous grooves, corners, and narrow spaces, makes it difficult for cleaning tools to reach every trace of residual material. Furthermore, the degree of material adhesion varies greatly, with some materials firmly adhering to the interior walls, making them difficult to clean. Cleaning equipment must be able to flexibly adjust the angle, pressure, and movement of the cleaning tool, implementing differentiated cleaning strategies based on the material characteristics of different areas to achieve thorough cleaning in the shortest possible time. This requires cleaning equipment to possess sophisticated environmental perception and decision-making planning capabilities, capable of analyzing material distribution maps in real time and determining the optimal cleaning path. Furthermore, a diverse set of cleaning tools must be developed, with rapid tool replacement capabilities, to adapt to the complex and changing interior environment of train cars. Summary of the Invention

[0003] The present invention provides an intelligent carriage cleaning method, which mainly includes:

[0004] The three-dimensional point cloud data of the interior of the carriage is obtained through lidar scanning, and the semantic segmentation of the point cloud data is performed in combination with a deep learning algorithm to identify the complex structure inside the carriage and the distribution of residual materials, and obtain the interior environment information model of the carriage; according to the interior environment information model of the carriage, a rule-based reasoning algorithm is used to analyze the position distribution characteristics of the materials, and the degree of adhesion of the materials is detected by a spectral analysis sensor, the interior of the carriage is divided into multiple cleaning areas, and the cleaning priority of each area is determined; according to the characteristics of the cleaning area and the properties of the materials, suitable cleaning tools are matched and selected from the predefined tool library, and the automatic assembly of the tools is realized by the robotic arm to improve the cleaning adaptability; for each cleaning area, a reinforcement learning algorithm is used, combined with the interior environment constraints of the carriage and the cleaning efficiency goals, to learn and optimize the cleaning process. The system uses the motion trajectory and operating parameters of the sweeping tool to formulate a cleaning path; during the cleaning process, it uses visual sensors to obtain the distribution changes of residual materials in real time, and dynamically updates the material distribution map through target detection and semantic segmentation algorithms, and adjusts the cleaning path in real time according to the updated distribution map; for materials with high adhesion, the contact force between the cleaning tool and the material is monitored in real time through force feedback sensors, and the pressure and angle of the cleaning tool are dynamically adjusted in combination with the pre-established material property database until the material is cleaned; during the cleaning operation, the cleaning data under different carriage environments and material properties are recorded and accumulated, and the cleaning path planning and tool selection strategies are continuously optimized through convolutional neural networks and decision tree algorithms to achieve self-learning and evolution of the cleaning equipment, thereby improving its adaptability and cleaning effect.

[0005] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0006] The present invention discloses an intelligent carriage cleaning method. The method acquires three-dimensional point cloud data of the interior of the carriage through laser radar scanning, combines deep learning algorithms for semantic segmentation, identifies the internal structure of the carriage and the distribution of residual materials, and constructs an environmental information model. Based on the model, rule reasoning and spectral analysis are used to divide the cleaning area and determine the priority. According to the characteristics of the area, appropriate cleaning tools are automatically selected and assembled. The cleaning path is optimized by using a reinforcement learning algorithm combined with environmental constraints and efficiency goals. During the cleaning process, the material distribution map is updated in real time and the path is dynamically adjusted. For adhered materials, the parameters of the cleaning tool are adjusted by force feedback sensors. The present invention also improves the adaptability of the cleaning equipment through continuous learning and optimization, thereby achieving intelligent and efficient carriage cleaning. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 The figure is a flow chart of an intelligent carriage cleaning method of the present invention.

[0008] Figure 2 This is a schematic diagram of an intelligent carriage cleaning method according to the present invention.

[0009] Figure 3 This is another schematic diagram of an intelligent carriage cleaning method of the present invention. DETAILED DESCRIPTION

[0010] To further understand the content of the present invention, the present invention is described in detail with reference to the accompanying drawings and examples. The present application is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are intended only to illustrate the relevant invention and are not intended to limit the invention. It should also be noted that, for ease of description, only portions relevant to the invention are shown in the accompanying drawings.

[0011] like Figure 1-3 In this embodiment, an intelligent carriage cleaning method may specifically include:

[0012] S101. Obtain three-dimensional point cloud data of the interior of the vehicle compartment through laser radar scanning, and perform semantic segmentation on the point cloud data in combination with a deep learning algorithm to identify the complex structure and distribution of residual materials inside the vehicle compartment, and obtain an internal environment information model of the vehicle compartment.

[0013] A raw point cloud dataset of the train interior was obtained. A statistical filter from the PCL point cloud library was used to remove noise points from the point cloud. The point cloud was then downsampled using a voxelized grid method to obtain a preprocessed point cloud dataset. The preprocessed point cloud data was then fed into the PointNet++ network. Using semantic segmentation, each point was classified into two categories: the train interior structure and residual material. This yielded a subset of the train interior structure point cloud and a subset of the residual material point cloud. For the subset of the train interior structure point cloud, a region growing algorithm was used for geometric segmentation. Based on growth criteria based on normal vector angle thresholds and curvature thresholds, the point cloud was segmented into distinct structural components. Each component point cloud was then Euclidean clustered to identify independent structural components. For the subset of the residual material point cloud, the number of neighbors for each point was calculated to obtain the local density distribution of the point cloud. The point cloud was then projected onto the Z-axis, and the number of points within different height intervals was counted to generate a height distribution histogram. The location and amount of residual material accumulation were determined based on the peak area of ​​the density distribution and the peak position of the height histogram. The geometric segmentation results of the interior structure are spatially aligned with the residual material distribution to construct a 3D model of the interior environment, including structural components and residual material information. Based on this model, the cleaning robot's motion path is optimized, combining the residual material distribution and structural layout to generate a cleaning trajectory that covers the residual material area and avoids structural obstacles. The cleaning tool is selected and cleaning parameters are adjusted. A point cloud of the cleaned interior is obtained and compared with the original point cloud to quantify the change in residual material volume before and after cleaning. The cleaning effect is evaluated. If the expected cleanliness level is not achieved, the cleaning strategy is adjusted until it meets the requirements.

[0014] Specifically, a lidar scan was used to acquire 3D point cloud data of the interior, generating a raw point cloud dataset. The statistical filter StatisticalOutlierRemoval in the PCL point cloud library was used to remove noise points, and downsampling was performed using the voxelGrid method to obtain a preprocessed point cloud dataset. The PointNet++ network was applied to the preprocessed point cloud data for semantic segmentation, classifying each point into two categories: internal structure and residual material. Based on the semantic segmentation results, a subset of the internal structure point cloud and a subset of the residual material point cloud were extracted. The internal structure point cloud subset was geometrically segmented using a region growing algorithm with normal vector angle thresholds and curvature thresholds as growth criteria to segment the point cloud into different structural components. Each component point cloud was then Euclidean clustered to identify independent structural components. For the residual material point cloud subset, the number of neighbors of each point was calculated to obtain the local density distribution of the point cloud. The point cloud was projected onto the Z-axis, and the number of points within different height intervals was counted to generate a height distribution histogram. The location and amount of residual material accumulation are determined based on the peak area of ​​the density distribution and the peak position of the height histogram. The geometric segmentation results of the interior structure are spatially aligned with the residual material distribution to construct a 3D model of the interior environment, including information about structural components and residual material. Point cloud registration and 3D visualization are implemented using the Open3D library. Based on the interior environment model, automated cleaning operations are planned. Taking into account the residual material distribution and structural layout, the cleaning robot's motion path is optimized to generate a trajectory that covers the residual material area and avoids structural obstacles. Appropriate cleaning tools, such as a high-pressure water gun or a brush, are selected, and parameters such as cleaning pressure and flow rate are adjusted. The 3D model is used to monitor the cleaning process in real time, generating a point cloud of the interior after cleaning. By comparing it with the original point cloud, the change in residual material volume before and after cleaning is quantified to evaluate the cleaning effect. If the expected cleanliness level is not achieved, the cleaning strategy is adjusted, such as increasing the number of cleaning cycles or replacing cleaning tools, until the target is met. Through these steps, a complete closed-loop solution from 3D scanning of the interior to automated cleaning operations is implemented. Point cloud data analysis and modeling provide data support for cleaning decisions, optimizing cleaning paths and parameter selection, and improving vehicle cleaning efficiency and quality. The 3D model can also be used to evaluate and provide feedback on cleaning effectiveness, creating a closed-loop information system.

[0015] S102. Based on the interior environment information model of the carriage, a rule-based reasoning algorithm is used to analyze the location distribution characteristics of the materials, and the degree of adhesion of the materials is detected by a spectral analysis sensor. The interior of the carriage is divided into multiple cleaning areas, and the cleaning priority of each area is determined.

[0016] The environmental information model of the interior of the carriage is obtained, and the decision tree algorithm is used to analyze the environmental information model to obtain a material distribution thermodynamic map; the material adhesion information is obtained, and the material adhesion degree is determined according to a preset adhesion degree threshold to obtain a material adhesion distribution map; the material distribution thermodynamic map and the adhesion distribution map are fused to extract the material distribution density characteristics and adhesion degree characteristics in the area; the K-means clustering algorithm is used to cluster the fused feature data, and the optimal number of clusters is determined according to the silhouette coefficient to obtain multiple cleaning areas inside the carriage; for each cleaning area, the mean value of the material distribution density and the mean value of the adhesion degree in the area are calculated respectively, and the two means are normalized to obtain to the distribution density score and adhesion degree score of each area; set the distribution density score weight and the adhesion degree score weight, calculate the comprehensive score of each area by weight, and determine the cleaning priority of each cleaning area according to the comprehensive score; determine the cleaning order of each cleaning area according to the cleaning priority from high to low; input the cleaning order and the position coordinate information of each cleaning area into the cleaning robot; the cleaning robot cleans each cleaning area in order from high to low according to the input cleaning order and position coordinates; during the cleaning process, the vehicle compartment environment information model is updated in real time, and the cleaning task is completed when all areas are cleaned.

[0017] Specifically, based on the interior environment information model, a decision tree algorithm is used to analyze the distribution characteristics of materials within the vehicle compartment, generating a material distribution thermogram. Spectral analysis sensors collect material adhesion information and determine the degree of adhesion based on a preset adhesion threshold to generate a material adhesion distribution map. The material distribution thermogram and adhesion distribution map are fused to extract the material distribution density and adhesion characteristics within each region. The fused feature data is clustered using the K-means algorithm, and the optimal number of clusters is determined based on the silhouette coefficient to identify multiple cleaning zones within the vehicle compartment. For each cleaning zone, the mean material distribution density and mean adhesion level are calculated as two indicators to measure the cleaning priority of that zone. These two indicators are normalized using a min-max process to generate a distribution density score and adhesion level score for each zone. A weight of 6 is assigned to the distribution density score and a weight of 4 is assigned to the adhesion level score. A weighted comprehensive score is calculated for each zone to determine the cleaning priority. The cleaning order for each zone is determined based on the cleaning priority from highest to lowest. The cleaning order and the location coordinates of each cleaning zone are input to the cleaning robot. The cleaning robot cleans each cleaning area in descending order of priority, based on the input cleaning order and location coordinates. During the cleaning process, the vehicle environment information model is updated in real time. Once all areas are cleaned, the cleaning task is complete.

[0018] S103. According to the characteristics of the cleaning area and the material properties, suitable cleaning tools are matched and selected from a predefined tool library, and the tools are automatically assembled through a robotic arm to improve cleaning adaptability.

[0019] Acquire image information of the cleaning area and image information of the material to be cleaned; use a fast random forest algorithm to extract and classify features of the cleaning area image information to obtain regional material and roughness regional characteristic parameters; use a convolutional neural network model to identify the material type of the material to be cleaned image information, and obtain the adhesion and corrosive material attribute parameters corresponding to the material type by querying a predefined material attribute knowledge base; perform similarity matching in a predefined tool feature database based on the regional characteristic parameters and the material attribute parameters to obtain several applicable cleaning tools; perform weighted scoring on the material, wear resistance and cleaning effect performance parameters of the applicable cleaning tools , select the cleaning tool with the highest score as the optimal cleaning tool, and send the model of the optimal cleaning tool to the robotic arm control unit; the robotic arm control unit queries the corresponding tool grabbing position and assembly steps in the predefined robotic arm assembly knowledge base according to the received tool model, controls the robotic arm to grab the target tool from the tool storage area and install it at the execution end, and completes the automatic replacement of the cleaning tool; the robotic arm carries the optimal cleaning tool to carry out cleaning operations on the designated area, adjusts the robotic arm movement speed and contact pressure according to the material of the area and the adhesion of the material, and controls the cleaning force; if there are stubborn stains, adopts repeated cleaning mode to clean, so as to improve the cleaning effect.

[0020] Specifically, based on the acquired image information of the cleaning area, a fast random forest algorithm is used to extract and classify feature parameters such as regional material and roughness; based on the acquired image of the material to be cleaned, a convolutional neural network model is used to identify the material type, and by querying the predefined material attribute knowledge base, the material attribute parameters such as adhesion and corrosivity are obtained; based on the regional feature parameters and material attribute parameters, similarity matching is performed in the predefined tool feature database to screen out several suitable cleaning tools; the performance parameters such as material, wear resistance, and cleaning effect of the applicable tools are weighted and scored, and the tool with the highest score is selected as the optimal cleaning tool. Its model is sent to the robotic arm control unit; the robotic arm control unit queries the corresponding tool grasping position and assembly steps in the predefined robotic arm assembly knowledge base based on the received tool model; the robotic arm accurately grasps the target tool from the tool storage area according to the grasping position information, and installs it at the execution end according to the queried assembly steps to complete the automatic replacement of the tool; after the assembly is completed, the robotic arm carries the cleaning tool to carry out cleaning operations on the designated area. According to the material of the area and the adhesion of the material, the cleaning force is controlled by adjusting the robotic arm movement speed and contact pressure. For stubborn stains, a repeated cleaning mode is adopted to improve the cleaning effect while taking into account the cleaning efficiency.

[0021] S104. For each cleaning area, a reinforcement learning algorithm is used to learn and optimize the motion trajectory and operating parameters of the cleaning tool in combination with the internal environmental constraints of the vehicle compartment and the cleaning efficiency target, and a cleaning path is formulated.

[0022] Acquire the three-dimensional environmental model data of the interior of the vehicle and the cleaning area division information, and extract the relevant environmental constraint parameters for each cleaning area, wherein the environmental constraint parameters include the area of ​​the cleaning area and the distribution of obstacles; construct a cleaning efficiency evaluation function, which combines multiple optimization objectives such as cleaning coverage, cleaning time, and energy consumption, and uses a weighted summation method to assign weights to each of the optimization objectives, and linearly combine them to form the cleaning efficiency evaluation function; adopt a deep reinforcement learning algorithm, with the motion trajectory and operation parameters of the cleaning tool as decision variables, and the cleaning efficiency evaluation function as the reward function, to obtain an optimized cleaning decision strategy through interactive exploration and learning between the intelligent agent and the environment; during the training process, according to the characteristics of each cleaning area, Given environmental constraints, the exploration behavior of the intelligent agent is guided and constrained, a penalty term is introduced in the reward function, and negative rewards are given for behaviors that produce invalid or dangerous cleaning paths; after multiple rounds of iterative training and optimization, the performance of the decision-making strategy model corresponding to each of the cleaning areas is evaluated on the validation set, and the model with the highest value of the cleaning efficiency evaluation function is selected as the optimal strategy model; during the actual cleaning operation, the current cleaning area information is obtained, it is determined which of the cleaning areas the current cleaning area belongs to, the corresponding decision-making strategy model file is loaded, and the optimal cleaning tool motion trajectory and operation control parameter sequence are planned and generated in real time; the cleaning tool autonomously performs the cleaning operation according to the motion trajectory and operation parameters generated by the plan through the motion control system of the robotic arm or mobile robot.

[0023] Specifically, the system obtains 3D environmental model data of the vehicle interior and cleaning area division information. For each cleaning area, relevant environmental constraint parameters, such as the cleaning area and obstacle distribution, are extracted. A cleaning efficiency evaluation function is constructed, comprehensively considering multiple optimization objectives, such as cleaning coverage, cleaning time, and energy consumption. A weighted summation approach can be used to assign weights to each optimization objective, which are then linearly combined to form the evaluation function. The weights can be adjusted based on actual needs. The deep reinforcement learning algorithm DDPG (Deep Deterministic Policy Gradient) is employed, using the cleaning tool's trajectory and operating parameters as decision variables and the cleaning efficiency evaluation function as the reward function. Through interactive exploration and learning between the agent and the environment, the cleaning decision strategy is gradually optimized. During training, the agent's exploration behavior is guided and constrained based on the specific environmental constraints of each cleaning area. A penalty term can be introduced in the reward function, assigning negative rewards to behaviors that result in invalid or dangerous cleaning paths, thereby guiding the agent towards safe and efficient exploration. After multiple rounds of iterative training and optimization, the performance of the decision-making strategy models corresponding to each cleaning area is evaluated on a validation set, and the model with the highest cleaning efficiency evaluation function value is selected as the optimal strategy model. During actual cleaning operations, based on the current cleaning area information, an if-else conditional statement is used to determine which cleaning area the current cleaning area belongs to. The corresponding decision-making strategy model file is then loaded, and the optimal cleaning tool motion trajectory and operation control parameter sequence are generated in real time. Based on the planned motion trajectory and operation parameters, the cleaning tool autonomously performs the cleaning operation through the motion control system of a robotic arm or mobile robot, achieving intelligent and efficient carriage cleaning.

[0024] S105. During the cleaning process, the visual sensor is used to obtain the distribution changes of the residual materials in real time. The material distribution map is dynamically updated through target detection and semantic segmentation algorithms, and the cleaning path is adjusted in real time according to the updated distribution map.

[0025] Obtain an initial residual material distribution map of the cleaning area, the initial residual material distribution map is obtained by performing a panoramic scan of the cleaning area through a preset visual sensor; divide the initial residual material distribution map into several cleaning units, each of which corresponds to an area of ​​a fixed size; determine the distribution density of residual materials in each cleaning unit based on the divided cleaning units; use a greedy algorithm to plan an initial cleaning path based on the distribution density, the initial cleaning path gives priority to cleaning the cleaning units with large residual material density; control the cleaning equipment to clean the cleaning units along the initial cleaning path; during the cleaning process, collect images of the cleaning area in real time through visual sensors; use the YOLOv3 target detection algorithm to detect residual material targets in the cleaning area image, and obtain the position coordinates of the residual material targets in the cleaning area image; use the position coordinates as prior knowledge and use FCN A semantic segmentation algorithm is used to segment the cleaning area image to obtain the pixel-level distribution of residual materials; the segmentation result is compared with the initial residual material distribution map, and the pixel-by-pixel difference calculation is used to determine whether there is a new residual material area; if the number of new residual material pixels in a certain cleaning unit exceeds a preset threshold, the certain cleaning unit is determined as a new residual material area, and the new residual material area is added to the initial residual material distribution map to obtain an updated residual material distribution map; based on the updated residual material distribution map, the residual material distribution density of each cleaning unit is recalculated; the Dijkstra shortest path algorithm is used, with the cleaning unit as the node and the distance between adjacent cleaning units as the edge weight, to plan an updated cleaning path; the cleaning equipment is controlled to continue cleaning the cleaning unit along the updated cleaning path until all cleaning tasks are completed.

[0026] Specifically, to obtain an initial residual material distribution map of the cleaning area, a pre-installed visual sensor can be used to perform a panoramic scan of the cleaning area to acquire an image of the cleaning area. The acquired image of the cleaning area is divided into several cleaning units, each corresponding to an area of ​​fixed size. The initial residual material distribution map is processed based on the divided cleaning units to determine the residual material distribution density within each cleaning unit. Based on the distribution density, a greedy algorithm is used to plan the initial cleaning path, prioritizing cleaning units with high residual material density. The cleaning equipment is controlled to clean the cleaning units along the planned initial cleaning path. During the cleaning process, the visual sensor captures images of the cleaning area in real time, and the YOLOv3 object detection algorithm is used to detect residual material targets in the image. Using the position coordinates of the detected residual material targets in the image as prior knowledge, the image is segmented using the FCN semantic segmentation algorithm to determine the pixel-level distribution of residual material. The segmentation results are compared with the initial residual material distribution map, and pixel-by-pixel difference calculations are used to determine whether new residual material areas exist. If a large number of newly detected residual material pixels appear within a cleaning unit, the cleaning unit is considered to contain a newly detected residual material region and is added to the residual material distribution map. Based on the updated residual material distribution map, the residual material distribution density for each cleaning unit is recalculated. Using the Dijkstra shortest path algorithm, with cleaning units as nodes and the distances between adjacent cleaning units as edge weights, an updated cleaning path is planned. The cleaning equipment is then controlled to continue cleaning the cleaning units along the updated cleaning path until all cleaning tasks are completed.

[0027] S106. For materials with a high degree of adhesion, the contact force between the cleaning tool and the material is monitored in real time through a force feedback sensor, and the pressure and angle of the cleaning tool are dynamically adjusted in combination with a pre-established material property database until the material is cleaned.

[0028] Obtain the adhesion threshold of the material in a pre-established material property database; monitor the contact force between the cleaning tool and the material in real time through a six-dimensional force sensor; compare the contact force monitored in real time with the material adhesion threshold; if the contact force is less than the material adhesion threshold, dynamically adjust the pressure and angle of the cleaning tool using a PID control algorithm so that the contact force gradually increases or reaches the adhesion threshold; optimize the pressure and angle parameters of the cleaning tool using a random forest algorithm based on the material hardness and roughness properties in the material property database; continue to execute the adjustment and optimization process until the contact force stabilizes near the adhesion threshold and the material surface is cleaned; detect the cleaned material surface through a machine vision system to determine whether it meets the preset cleanliness requirements; if the cleanliness requirements are met, the cleaning task is completed; otherwise, return to continue adjusting the cleaning tool until the cleaning is qualified.

[0029] Specifically, the adhesion threshold of the material in the pre-established material property database is obtained, and the contact force between the cleaning tool and the material is monitored in real time by a six-dimensional force sensor. The monitored contact force is compared with the material adhesion threshold. If the contact force is less than the material adhesion threshold, the PID control algorithm is used to dynamically adjust the pressure and angle of the cleaning tool so that the contact force gradually approaches or reaches the adhesion threshold. According to the material hardness, roughness and other properties in the material property database, the random forest algorithm is used to optimize the pressure and angle parameters of the cleaning tool. The above adjustment and optimization process is continuously performed until the contact force stabilizes near the adhesion threshold and the material surface is cleaned. The surface of the material after cleaning is inspected by the machine vision system to determine whether the cleanliness requirements are met. If so, the cleaning task is completed, otherwise it returns to continue adjusting the cleaning tool until the cleaning is qualified.

[0030] S107. During the cleaning process, record and accumulate cleaning data under different carriage environments and material characteristics. Through convolutional neural networks and decision tree algorithms, continuously optimize cleaning path planning and tool selection strategies to achieve self-learning and evolution of cleaning equipment, thereby improving its adaptive capabilities and cleaning effects.

[0031] Acquire a cleaning operation dataset under different compartment environments and material characteristics, the dataset including environmental images, material types and cleaning effect evaluation; preprocess the dataset, including image normalization, data cleaning and annotation; randomly divide the preprocessed dataset into a training set and a test set; use a convolutional neural network to extract features of the environmental images in the training set, and extract multi-scale and multi-level feature representations through convolution layers and pooling layers; fuse the feature representation of the environmental image with other features of the material type to obtain a comprehensive cleaning operation feature vector; use a decision tree algorithm to train the cleaning operation feature vector, and generate a decision tree model that adapts to different situations based on environmental complexity and material adhesion factors; use the test set data to train the trained convolutional neural network and decision tree models. The performance of the tree model is evaluated, and parameters are tuned and the model is improved based on the evaluation results. When the cleaning equipment is performing operations, the camera collects real-time images of the vehicle environment and identifies the material type. The collected data is input into the trained convolutional neural network and decision tree model to adaptively generate the optimal cleaning path and tool selection strategy. The cleaning equipment performs cleaning operations according to the generated strategy and continuously records the cleaning process data through sensors. The recorded cleaning process data is automatically annotated and added to the training data set. At regular intervals, the convolutional neural network and decision tree model are incrementally trained and optimized using the newly added cleaning data to enable the model to continue learning and evolving. The optimized model parameters are updated to the control system of the cleaning equipment, so that the equipment can perform cleaning operations based on the latest decision rules.

[0032] Specifically, cleaning operation data under different vehicle cabin environments and material characteristics, including environmental images, material types, and cleaning effect evaluations, is collected to construct a cleaning operation dataset. The dataset is preprocessed, including image normalization, data cleaning, and annotation. The preprocessed dataset is randomly divided into training and test sets. A convolutional neural network is used to extract features from the environmental images in the training set. Multi-scale and multi-level feature representations are extracted through convolutional and pooling layers. Furthermore, other features, such as material type, are fused with image features to generate a comprehensive cleaning operation feature vector. A decision tree algorithm (such as CART or C5) is used to train the cleaning operation feature vector and construct decision rules for cleaning path planning and tool selection. Based on factors such as environmental complexity and material adhesion, a decision tree adapted to different situations is generated. The performance of the trained convolutional neural network and decision tree models is evaluated using the test set data. Parameters are tuned and the model is improved based on the evaluation results to enhance the model's generalization and decision accuracy. While the cleaning equipment is operating, a camera captures real-time images of the vehicle cabin environment and identifies material types. The collected data is fed into a trained convolutional neural network and decision tree model to adaptively generate the optimal cleaning path and tool selection strategy. The cleaning equipment performs cleaning operations according to the generated strategy and continuously records cleaning process data, including cleaning trajectory, tool usage, and cleaning results, through sensors. The recorded data is automatically annotated and added to the training dataset. At regular intervals (e.g., monthly), the convolutional neural network and decision tree models are incrementally trained and optimized using newly added cleaning data, enabling the models to continuously learn and evolve to adapt to the changing cabin environment and material characteristics. The optimized model parameters are updated to the cleaning equipment's control system, enabling the equipment to execute cleaning operations based on the latest decision rules. Furthermore, a human-machine interface displays the cleaning equipment's operating status and cleaning results in real time, facilitating staff oversight and management.

[0033] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An intelligent carriage cleaning method, characterized in that: The method comprises: The three-dimensional point cloud data of the interior of the carriage is obtained through lidar scanning, and the semantic segmentation of the point cloud data is performed in combination with a deep learning algorithm to identify the complex structure inside the carriage and the distribution of residual materials, and obtain the interior environment information model of the carriage; according to the interior environment information model of the carriage, a rule-based reasoning algorithm is used to analyze the position distribution characteristics of the materials, and the degree of adhesion of the materials is detected by a spectral analysis sensor, the interior of the carriage is divided into multiple cleaning areas, and the cleaning priority of each area is determined; according to the characteristics of the cleaning area and the properties of the materials, suitable cleaning tools are matched and selected from the predefined tool library, and the automatic assembly of the tools is realized by the robotic arm to improve the cleaning adaptability; for each cleaning area, a reinforcement learning algorithm is used, combined with the interior environment constraints of the carriage and the cleaning efficiency goals, to learn and optimize the cleaning process. The system uses the motion trajectory and operating parameters of the sweeping tool to formulate a cleaning path; during the cleaning process, it uses visual sensors to obtain the distribution changes of residual materials in real time, and dynamically updates the material distribution map through target detection and semantic segmentation algorithms, and adjusts the cleaning path in real time according to the updated distribution map; for materials with high adhesion, the contact force between the cleaning tool and the material is monitored in real time through force feedback sensors, and the pressure and angle of the cleaning tool are dynamically adjusted in combination with the pre-established material property database until the material is cleaned; during the cleaning operation, the cleaning data under different carriage environments and material properties are recorded and accumulated, and the cleaning path planning and tool selection strategies are continuously optimized through convolutional neural networks and decision tree algorithms to achieve self-learning and evolution of the cleaning equipment, thereby improving its adaptability and cleaning effect.

2. The method according to claim 1, characterized in that The three-dimensional point cloud data of the interior of the carriage is obtained by laser radar scanning, and semantic segmentation of the point cloud data is performed in combination with a deep learning algorithm to identify the complex structure and distribution of residual materials inside the carriage, and obtain an internal environment information model of the carriage, including: Obtain the original point cloud dataset of the interior of the car, use the statistical filter in the point cloud library PCL to remove noise points in the point cloud, and use the voxel grid method to downsample the point cloud to obtain the preprocessed point cloud dataset; The pre-processed point cloud data is input into the PointNet++ network. Through point cloud semantic segmentation, each point is divided into two categories: the internal structure of the carriage and the residual material. The internal structure point cloud subset and the residual material point cloud subset are obtained. For the internal structure point cloud subset of the car body, a region growing algorithm is used for geometric segmentation. Based on the growth criteria of the normal vector angle threshold and the curvature threshold, the point cloud is segmented into different structural components. The point cloud of each component is then subjected to Euclidean clustering to obtain independent structural components. For the residual material point cloud subset, the number of neighboring points of each point is calculated to obtain the local density distribution of the point cloud; Project the point cloud to the Z-axis direction, count the number of points in different height intervals, and generate a height distribution histogram; Determine the accumulation location and quantity of residual materials based on the density distribution peak area and height histogram peak position; The geometric segmentation results of the interior structure of the carriage are spatially aligned with the distribution of residual materials to construct a three-dimensional model of the interior environment of the carriage that includes information on structural components and residual materials. Based on the constructed interior environment model of the carriage, combined with the residual material distribution and structural layout, the motion path of the cleaning robot is optimized to generate a cleaning trajectory that covers the residual material area and avoids structural obstacles. Select a cleaning tool and adjust cleaning parameters; Obtain the point cloud of the interior of the car after cleaning. By comparing it with the original point cloud, quantify the change in the volume of residual material before and after cleaning, and evaluate the cleaning effect. If the expected cleanliness is not achieved, adjust the cleaning strategy until the requirements are met.

3. The method according to claim 1, characterized in that The system uses a rule-based reasoning algorithm based on the interior environment information model to analyze the location distribution characteristics of the materials, and uses a spectral analysis sensor to detect the degree of material adhesion, divides the interior of the carriage into multiple cleaning areas, and determines the cleaning priority of each area, including: Obtaining an environmental information model of the interior of the carriage, analyzing the environmental information model using a decision tree algorithm, and obtaining a material distribution thermodynamic map; Obtaining material adhesion information, determining the material adhesion degree according to a preset adhesion degree threshold, and obtaining a material adhesion distribution map; fusing the material distribution thermodynamic map and the adhesion distribution map to extract material distribution density characteristics and adhesion degree characteristics within the region; The K-means clustering algorithm is used to cluster the fused feature data, and the optimal number of clusters is determined according to the silhouette coefficient to obtain multiple cleaning areas inside the car. For each cleaning area, the mean value of the material distribution density and the mean value of the adhesion degree in the area are calculated respectively, and the two means are normalized to obtain the distribution density score and adhesion degree score of each area; Setting a distribution density score weight and an adhesion degree score weight, weightedly calculating a comprehensive score for each area, and determining a cleaning priority level for each cleaning area based on the comprehensive score; Determining a cleaning order for each cleaning area according to the cleaning priority levels from high to low; inputting the cleaning sequence and the position coordinate information of each cleaning area into the cleaning robot; The cleaning robot performs cleaning operations on each of the cleaning areas in descending order of priority according to the input cleaning order and position coordinates; During the cleaning process, the vehicle compartment environment information model is updated in real time. When all areas are cleaned, the cleaning task is completed.

4. The method according to claim 1, wherein The system matches and selects appropriate cleaning tools from a predefined tool library based on the characteristics of the cleaning area and the material properties, and automatically assembles the tools through a robotic arm to improve cleaning adaptability, including: Obtain image information of the cleaning area and image information of the material to be cleaned; A fast random forest algorithm is used to extract and classify the image information of the cleaned area to obtain regional material and roughness regional feature parameters; A convolutional neural network model is used to identify the material type of the image information of the material to be cleaned, and the adhesive and corrosive material property parameters corresponding to the material type are obtained by querying a predefined material property knowledge base; Based on the regional characteristic parameters and the material attribute parameters, similarity matching is performed in a predefined tool feature database to obtain several applicable cleaning tools; Performing a weighted score on the material, wear resistance, and cleaning effect performance parameters of the applicable cleaning tools, selecting the cleaning tool with the highest score as the optimal cleaning tool, and sending the model of the optimal cleaning tool to the robotic arm control unit; The robotic arm control unit searches the predefined robotic arm assembly knowledge base for the corresponding tool grabbing position and assembly steps based on the received tool model, controls the robotic arm to grab the target tool from the tool storage area and install it at the execution end, completing the automated replacement of the cleaning tool; The robotic arm carries the optimal cleaning tool to clean the designated area, and adjusts the movement speed and contact pressure of the robotic arm to control the cleaning force according to the material of the area and the adhesion of the material; If there are stubborn stains, use the repeated cleaning mode to improve the cleaning effect.

5. The method according to claim 1, wherein For each cleaning area, the reinforcement learning algorithm is used to learn and optimize the motion trajectory and operating parameters of the cleaning tool in combination with the internal environmental constraints of the vehicle compartment and the cleaning efficiency target, and a cleaning path is formulated, including: Acquire the three-dimensional environment model data of the interior of the vehicle and the cleaning area division information, and extract the relevant environmental constraint parameters for each cleaning area, wherein the environmental constraint parameters include the area of ​​the cleaning area and the obstacle distribution; Constructing a cleaning efficiency evaluation function, which combines multiple optimization objectives such as cleaning coverage, cleaning time, and energy consumption, assigns weights to each of the optimization objectives in a weighted summation manner, and linearly combines them to form the cleaning efficiency evaluation function; A deep reinforcement learning algorithm is used, with the motion trajectory and operating parameters of the cleaning tool as decision variables and the cleaning efficiency evaluation function as the reward function. Through interactive exploration and learning between the intelligent agent and the environment, an optimized cleaning decision strategy is obtained; During the training process, the exploration behavior of the agent is guided and constrained according to the specific environmental constraints of each cleaning area, and a penalty term is introduced into the reward function to give negative rewards for behaviors that produce invalid or dangerous cleaning paths; After multiple rounds of iterative training and optimization, the performance of the decision strategy model corresponding to each cleaning area is evaluated on the validation set, and the model with the highest cleaning efficiency evaluation function value is selected as the optimal strategy model; During the actual cleaning operation, the current cleaning area information is obtained, the current cleaning area is determined, the corresponding decision strategy model file is loaded, and the optimal cleaning tool motion trajectory and operation control parameter sequence are planned and generated in real time; The cleaning tool autonomously performs cleaning operations according to the motion trajectory and operation parameters generated by the plan through the motion control system of the robotic arm or mobile robot.

6. The method according to claim 1, characterized in that During the cleaning process, the visual sensor is used to obtain the distribution changes of residual materials in real time, the material distribution map is dynamically updated through target detection and semantic segmentation algorithms, and the cleaning path is adjusted in real time according to the updated distribution map, including: Obtaining an initial residual material distribution map of the cleaning area, wherein the initial residual material distribution map is obtained by performing a panoramic scan of the cleaning area using a preset visual sensor; Dividing the initial residual material distribution map into a plurality of cleaning units, each of the cleaning units corresponding to an area of ​​a fixed size; Determining the distribution density of residual materials in each cleaning unit according to the divided cleaning units; A greedy algorithm is used to plan an initial cleaning path according to the distribution density, wherein the initial cleaning path prioritizes cleaning units with a high density of residual materials; Controlling the cleaning device to perform a cleaning operation on the cleaning unit along the initial cleaning path; During the cleaning process, the visual sensor collects images of the cleaning area in real time; Using the YOLOv3 target detection algorithm to detect the residual material target in the cleaning area image, and obtain the position coordinates of the residual material target in the cleaning area image; Using the position coordinates as prior knowledge, the FCN semantic segmentation algorithm is used to segment the cleaning area image to obtain the pixel-level distribution of residual materials; Comparing the segmentation result with the initial residual material distribution map, and determining whether there is a new residual material area by pixel-by-pixel difference calculation; If the number of newly added residual material pixels in a certain cleaning unit exceeds a preset threshold, the cleaning unit is determined as a newly added residual material area, and the newly added residual material area is added to the initial residual material distribution map to obtain an updated residual material distribution map; Recalculating the residual material distribution density of each cleaning unit according to the updated residual material distribution map; Using Dijkstra's shortest path algorithm, with the cleaning units as nodes and the distances between adjacent cleaning units as edge weights, an updated cleaning path is planned; The cleaning device is controlled to continue cleaning the cleaning unit along the updated cleaning path until all cleaning tasks are completed.

7. The method according to claim 1, characterized in that For materials with high adhesion, the contact force between the cleaning tool and the material is monitored in real time by a force feedback sensor, and the pressure and angle of the cleaning tool are dynamically adjusted in combination with a pre-established material property database until the material is cleaned, including: Obtaining a material adhesion threshold value from a pre-established material property database; The contact force between the cleaning tool and the material is monitored in real time through a six-dimensional force sensor; comparing the contact force monitored in real time with the material adhesion threshold; If the contact force is less than the material adhesion threshold, the pressure and angle of the cleaning tool are dynamically adjusted using a PID control algorithm so that the contact force gradually increases or reaches the material adhesion threshold; Optimizing the pressure and angle parameters of the cleaning tool using a random forest algorithm based on the material hardness and roughness attributes in the material property database; Continuing the adjustment and optimization process until the contact force stabilizes near the adhesion threshold and the material surface is cleaned; The surface of the cleaned material is inspected by a machine vision system to determine whether it meets the preset cleanliness requirements; If the cleanliness requirements are met, the cleaning task is completed; Otherwise, return and continue to adjust the cleaning tool until the cleaning is qualified.

8. The method according to claim 1, characterized in that During the cleaning process, the cleaning data under different compartment environments and material characteristics are recorded and accumulated. Through convolutional neural networks and decision tree algorithms, the cleaning path planning and tool selection strategies are continuously optimized to achieve self-learning and evolution of the cleaning equipment, thereby improving its adaptive capabilities and cleaning effects, including: Acquire a cleaning operation dataset under different vehicle compartment environments and material characteristics, the dataset including environment images, material types, and cleaning effect evaluation; Preprocessing the dataset, including image normalization, data cleaning and annotation; The preprocessed dataset is randomly divided into training set and test set; Using a convolutional neural network to extract features from the environmental images in the training set, and extracting multi-scale and multi-level feature representations through convolutional layers and pooling layers; Fusing the feature representation of the environment image with other features of the material type to obtain a comprehensive cleaning operation feature vector; A decision tree algorithm is used to train the cleaning operation feature vector, and a decision tree model adapted to different situations is generated according to factors such as environmental complexity and material adhesion; Use the test set data to evaluate the performance of the trained convolutional neural network and decision tree models, and perform parameter tuning and model improvement based on the evaluation results; When the cleaning equipment is performing operations, the camera collects real-time images of the carriage environment and identifies the material type; The collected data is input into the trained convolutional neural network and decision tree model to adaptively generate the optimal cleaning path and tool selection strategy; The cleaning equipment performs cleaning operations according to the generated strategy and continuously records cleaning process data through sensors; The recorded cleaning process data is automatically annotated and added to the training dataset; At regular intervals, new cleaning data is used to incrementally train and optimize the convolutional neural network and decision tree models, enabling the models to continuously learn and evolve. The optimized model parameters are then updated to the control system of the cleaning equipment, enabling the equipment to perform cleaning operations based on the latest decision rules.

Images