Indoor ground washing vehicle path planning method and device

Abstract

The invention relates to the technical field of ground washing vehicles, and discloses an indoor ground washing vehicle path planning method and device, and the method comprises the steps: collecting the multi-source data of an indoor environment, carrying out the probability field construction processing of a hierarchical feature map, obtaining a dynamic change region of an initial path sequence in real time, and carrying out the calculation of the dynamic change region. Performing multi-objective optimization on the conflict labeling path sequence, performing deep learning regulation and control on the reconstructed path sequence based on an experience playback mechanism, controlling the ground washing vehicle to execute cleaning and obtain execution data according to the real-time path sequence, and performing dynamic regulation and control on the reconstructed path through a deep learning model, so that the path has real-time offset capability. And the robustness of the path in a complex dynamic scene is ensured. Furthermore, execution data is fed back for self-learning through a closed-loop mechanism, continuous evolution of a path planning strategy is achieved, a data report reflecting the cleaning effect is finally generated, and the cleaning operation quality is continuously guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of floor-cleaning vehicles, and in particular to a path planning method and device for an indoor floor-cleaning vehicle. Background Art

[0002] As cities continue to become more intelligent, intelligent cleaning equipment is gaining widespread adoption in large indoor environments, such as commercial complexes, hospitals, airports, and warehouses. Indoor floor scrubbers, a typical example of automated floor cleaning equipment, rely heavily on their path planning capabilities to directly impact cleaning efficiency, energy consumption, and operational safety. To address dynamic and complex indoor environments and enhance operational flexibility and stability, path planning technology is evolving from traditional static planning to dynamic response and intelligent control. Currently, intelligent path planning, integrating multi-source perception, machine learning, and self-learning mechanisms, has become a research hotspot.

[0003] Existing indoor floor scrubbers mostly use static map construction and preset path generation mechanisms to plan work paths. Although some technologies have introduced dynamic obstacle avoidance or rule-based path correction, there are still problems with response lag and frequent local path conflicts when dealing with dynamically changing areas (such as temporary obstacles and personnel flow), which can easily lead to repeated cleaning, path interruptions, and even equipment stagnation. Especially in places with complex spatial layouts or dense traffic, the path planning system cannot effectively predict and actively avoid potential conflicts, resulting in a decrease in overall operating efficiency.

[0004] Therefore, it is necessary to provide a path planning method and device for an indoor floor washing vehicle to solve the problem that existing indoor floor washing vehicles have low efficiency in handling dynamic space operations. Summary of the Invention

[0005] The main purpose of the present invention is to provide a path planning method and device for an indoor floor washing vehicle, aiming to solve the technical problems mentioned in the above background technology.

[0006] The present invention adopts the following technical solutions: A path planning method for an indoor floor-washing vehicle, comprising: Collecting multi-source data of the indoor environment, clustering the multi-source data, and performing hierarchical modeling on the clustering results to obtain a hierarchical feature map including a weight vector set; Performing a probability field construction process on the hierarchical feature map, and performing a path initialization process on the hierarchical feature map after the probability field construction process to obtain an initial path sequence; Acquire the dynamic change area of ​​the initial path sequence in real time, and perform conflict prediction processing on the initial path sequence based on the dynamic change area to obtain a conflict marked path sequence; Performing multi-objective optimization on the conflict-marked path sequence to obtain a reconstructed path sequence; Performing deep learning control on the reconstructed path sequence based on an experience replay mechanism to obtain a real-time path sequence including a dynamic offset vector; The floor scrubber is controlled to perform cleaning and obtain execution data according to the real-time path sequence, and closed-loop self-learning processing is performed on the execution data based on a closed-loop mechanism to output a cleaning effect report.

[0007] Furthermore, the step of collecting multi-source data of the indoor environment, clustering the multi-source data, and performing hierarchical modeling on the clustering results to obtain a hierarchical feature map including a weight vector set includes: Collecting data from the indoor environment based on a sensor array to obtain a multi-source original data set, wherein the multi-source original data set includes three-dimensional spatial coordinates, ground friction resistance, and stain density distribution; performing noise filtering on the multi-source original data set to obtain a high-confidence data set; Extracting multi-dimensional features of the high-confidence data set to obtain a feature vector set, and performing hierarchical clustering processing on the high-confidence data set based on the feature vector set to obtain a multi-scale grouped data set; Performing hierarchical modeling on the multi-scale grouped data set to obtain a preliminary feature map; A weight vector assignment process is performed on the multi-scale grouped data set according to the preliminary feature map to obtain a hierarchical feature map including a weight vector set.

[0008] Furthermore, the step of performing probability field construction processing on the hierarchical feature map and performing path initialization processing on the hierarchical feature map after the probability field construction processing to obtain an initial path sequence includes: Performing region segmentation processing on the hierarchical feature map based on spatial clustering to obtain a partition feature matrix; Performing probability field construction processing on the partition feature matrix to obtain an initial probability field model; Performing node gridding processing on the initial probability field model, and adding travel cost and cleaning priority labels to the initial probability field model after the node gridding processing to obtain a probability grid map; Performing path search processing on the probability grid graph through dynamic programming to obtain a candidate path set; Path optimization processing is performed on the candidate path set to obtain a preferred path sequence, and priority marking is performed on the preferred path sequence based on node priority analysis to obtain an initial path sequence.

[0009] Furthermore, the step of acquiring the dynamic change area of ​​the initial path sequence in real time and performing conflict prediction processing on the initial path sequence based on the dynamic change area to obtain the conflict marked path sequence includes: Performing dynamic environment fusion on the initial path sequence to obtain an environment perception data set; Performing incremental differential analysis on the environmental perception data set to obtain a dynamically changing region set; Constructing a spatiotemporal conflict graph for the initial path sequence according to the dynamically changing region set to obtain a spatiotemporal conflict probability graph; Setting a dynamic conflict probability threshold, and performing conflict threshold segmentation processing on the spatiotemporal conflict probability graph based on the dynamic conflict probability threshold to obtain a conflict risk node set; Performing conflict marking processing on the initial path sequence according to the conflict risk node set to obtain a preliminary conflict path sequence; The preliminary conflict path sequence is smoothed and consistency-checked to obtain a conflict-marked path sequence.

[0010] Furthermore, the step of performing multi-objective optimization on the conflict-marked path sequence to obtain a reconstructed path sequence includes: Extracting spatiotemporal distribution features of the conflicting annotated path sequence, and performing segmentation decomposition processing on the conflicting annotated path sequence based on the spatiotemporal distribution features to obtain a path segment set; Constructing a multi-objective evaluation function for the path segment set to obtain a segment objective weight matrix; Performing genetic optimization processing on the path segment set based on the segment target weight moment to obtain a candidate path segment set; Performing dynamic risk assessment on the candidate path segment set based on a Bayesian probability model to obtain a risk path segment set; Performing path integration processing on the risk path segment set to obtain a preliminary reconstructed path sequence; The preliminary reconstructed path sequence is smoothed and optimized by Bezier curve fitting to obtain a reconstructed path sequence.

[0011] Furthermore, the step of performing deep learning control on the reconstructed path sequence based on the experience replay mechanism to obtain a real-time path sequence including a dynamic offset vector includes: Performing state feature extraction processing on the reconstructed path sequence according to the dynamically changing region to obtain a path state feature set; Constructing an experience replay library, and performing sample screening processing on the experience replay library according to the path state feature set to obtain a high-value experience sample set; Training a deep reinforcement learning network based on the high-value experience samples to obtain a path control strategy model; Performing action prediction processing on the reconstructed path sequence according to the path control strategy model to obtain an initial adjusted path sequence; Performing dynamic offset vector generation processing on the initial adjustment path sequence to obtain a dynamic offset vector set; Path reconstruction processing is performed on the initial adjusted path sequence according to the dynamic offset vector set to obtain a real-time path sequence.

[0012] Furthermore, the steps of controlling the floor scrubber to perform cleaning and obtain execution data according to the real-time path sequence, performing closed-loop self-learning processing on the execution data based on a closed-loop mechanism, and outputting a cleaning effect report include: Controlling the floor scrubber to perform a cleaning task according to the real-time path sequence to obtain an execution data set; Performing dynamic feature extraction processing on the execution data set based on a temporal convolutional network to obtain a dynamic feature vector group; performing update correction processing on the hierarchical feature map according to the dynamic feature vector group to obtain an updated feature map; Performing closed-loop feedback analysis on the updated feature map based on Markov decision making to obtain a feedback control parameter set for controlling the path and action strategy; Dynamically adjusting the real-time path sequence according to the feedback control parameter set to obtain an optimized path sequence; The execution effect of the optimized path sequence is evaluated to obtain a cleaning effect report including an updated self-learning model, and the updated self-learning model is fed back to the hierarchical modeling process to form a closed-loop adaptive system.

[0013] The present invention also provides a path planning device for an indoor floor-washing vehicle, comprising: An acquisition module is used to collect multi-source data of the indoor environment, perform clustering on the multi-source data, and perform hierarchical modeling on the clustering results to obtain a hierarchical feature map including a weight vector set; A construction module is used to perform a probability field construction process on the hierarchical feature map, and perform a path initialization process on the hierarchical feature map after the probability field construction process to obtain an initial path sequence; A prediction module is used to obtain the dynamic change area of ​​the initial path sequence in real time, and perform conflict prediction processing on the initial path sequence based on the dynamic change area to obtain a conflict marked path sequence; An optimization module, configured to perform multi-objective optimization on the conflict-marked path sequence to obtain a reconstructed path sequence; A control module, configured to perform deep learning control on the reconstructed path sequence based on an experience replay mechanism to obtain a real-time path sequence including a dynamic offset vector; The control module is used to control the floor scrubber to perform cleaning and obtain execution data according to the real-time path sequence, perform closed-loop self-learning processing on the execution data based on a closed-loop mechanism, and output a cleaning effect report.

[0014] The present invention further provides a computer device, comprising a processor, a memory, and a computer program stored in the memory and running on the processor, wherein the processor implements the above method when executing the computer program.

[0015] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the above method when executed by a processor.

[0016] Beneficial effects: In the present invention, multi-source data of the indoor environment is collected and clustered, and then the clustering results are hierarchically modeled to construct a hierarchical feature map containing a weight vector set, thereby realizing a high-dimensional abstract expression of the indoor environment. On this basis, the path pass probability in different areas is quantitatively modeled through probability field construction processing, and the path initialization processing is performed in combination with the model to generate an initial path sequence with global feasibility and regional adaptability, thereby improving the basic accuracy of path planning and the environmental adaptability.

[0017] By perceiving the dynamic changes in the areas involved in the initial path sequence in real time, it is possible to identify the impact of potential changes on path traffic, predict conflicts based on this, and generate an information-based path sequence with conflict annotations. Based on a multi-objective optimization algorithm, it takes into account path length, obstacle avoidance efficiency, and execution continuity, and obtains a path result with global optimization characteristics. The reconstructed path is dynamically regulated through a deep learning model, giving the path the ability to offset in real time and ensuring the path's robustness in complex dynamic scenarios. Furthermore, through a closed-loop mechanism, execution data feedback is used for self-learning, achieving the continuous evolution of the path planning strategy, and ultimately generating a data report reflecting the cleaning effect to continuously ensure the quality of the cleaning operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a schematic diagram of the steps of a path planning method for an indoor floor-cleaning vehicle according to the present invention; Figure 2 This is a schematic structural diagram of a path planning device for an indoor floor-cleaning vehicle according to the present invention; Figure 3 is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention; The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0019] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0020] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "transverse", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise" and the like, indicating positions or positional relationships, are based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the said features. In the description of the present invention, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined.

[0021] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; they may refer to mechanical connections, direct connections, or indirect connections through an intermediate medium; they may refer to internal communication between two components or the interaction between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0022] In the present invention, unless otherwise expressly specified or limited, a first feature being "above" or "below" a second feature may include the first and second features being in direct contact, or may include the first and second features not being in direct contact but being in contact via another feature between them. Furthermore, a first feature being "above," "above," and "above" a second feature may include the first feature being directly above or obliquely above the second feature, or may simply mean that the first feature is higher in level than the second feature. A first feature being "below," "below," and "below" a second feature may include the first feature being directly below or obliquely below the second feature, or may simply mean that the first feature is lower in level than the second feature.

[0023] Reference Figure 1 The present invention proposes a path planning method for an indoor floor washing vehicle, comprising the following steps: S1: Collect multi-source data of the indoor environment, perform clustering on the multi-source data, and perform hierarchical modeling on the clustering results to obtain a hierarchical feature map including a weight vector set; In step S1, the floor scrubber is equipped with a variety of sensors, including at least lidar, ultrasonic sensors and optical sensors, to collect multi-source data of the indoor environment in real time, covering key information such as the ground friction coefficient, stain density distribution and spatial geometric constraints. The ground friction coefficient reflects the resistance characteristics of different areas, the stain density distribution provides the priority basis for cleaning tasks, and the spatial geometric constraints define the boundaries of the areas where the floor scrubber can pass. The collection of multi-source data ensures a comprehensive perception of the indoor environment.

[0024] The multi-source data is further preprocessed using adaptive threshold filtering to remove noise and retain high-confidence data points. The threshold is automatically adjusted based on the dynamic characteristics of the data distribution, effectively removing outliers caused by sensor errors or environmental interference while retaining data points that reflect the true environmental characteristics. A hierarchical clustering algorithm is used to group the filtered data. This algorithm divides the indoor environment into multiple regions with similar characteristics based on the spatial correlation and stain characteristics of the data. For example, it distinguishes areas with high stain density from areas with low friction coefficients. This generates a multi-scale feature map and recursively clusters the data points into clusters at different levels by analyzing the Euclidean distance and feature similarity between the data points, thereby constructing a hierarchical feature representation.

[0025] Based on the generation of multi-scale feature maps, a weight vector is further assigned to each data point, which includes three dimensions: friction resistance, stain intensity, and regional connectivity. Friction resistance reflects the energy consumption required by the floor scrubber to travel in the area, stain intensity indicates the priority of the cleaning task, and regional connectivity describes the spatial accessibility of the area to other areas. To ensure the comparability and consistency between weights in different dimensions, the weight values ​​are mapped to a unified range through normalization, which is achieved through linear transformation or softmax function, thereby eliminating dimensional differences and enhancing cross-dimensional compatibility. The entire calculation process adopts multi-threaded parallel processing technology. By assigning data cleaning, clustering, and weight assignment tasks to multiple computing threads, the processing efficiency is significantly improved, reaching a processing speed of 100,000 data points per second, ensuring the real-time processing process in complex indoor environments.

[0026] The final output is a hierarchical feature map containing a set of spatial region weight vectors, which represents the indoor environment in a gridded form. Each grid point is associated with a weight vector, which comprehensively describes the environmental characteristics of the area. The hierarchical feature map not only retains the local details of the spatial region, but also reflects the global relationship between regions through clustering results, providing high-precision input data for probability field construction and path initialization.

[0027] S2: performing a probability field construction process on the hierarchical feature map, and performing a path initialization process on the hierarchical feature map after the probability field construction process to obtain an initial path sequence; In step S2, based on the hierarchical feature map and the weight vector set, an initial path sequence is generated through probability field construction processing and path initialization processing. This ensures that path planning takes into account both cleaning efficiency and driving feasibility by comprehensively analyzing environmental characteristics and combining optimization algorithms. Specifically, based on the hierarchical feature map, the feature map contains a grid representation of the indoor environment. Each grid point is associated with a weight vector. The weight vector contains information such as stain intensity, friction resistance and regional connectivity. This information provides multi-dimensional input for probability field construction. The probability field model is constructed through weighted superposition calculation. Specifically, the field intensity distribution of the probability field is determined by the stain intensity and friction resistance in the weight vector. The stain intensity reflects the priority of the cleaning task, and the friction resistance represents the energy consumption demand of the floor scrubber in the area. By weighted combination of these two dimensions, a probability field representing the cleaning priority and difficulty of passage of each area is generated and presented in the form of a grid. The field intensity value of each grid point reflects the attraction or repulsion of the area to the floor scrubber path selection.

[0028] To optimize the construction of the probability field, a Markov random field (MRF) model is adopted to balance cleaning requirements and driving efficiency by iteratively adjusting probability values. Specifically, the Markov random field defines the neighborhood relationship between grid points and combines the regional connectivity in the weight vector to calculate the conditional probability distribution of each grid point. This process gradually converges through iterative optimization algorithms (such as belief propagation or graph cut algorithms), so that the probability field can accurately reflect the priority distribution of cleaning tasks while avoiding path infeasibility caused by local high resistance areas. The optimized probability field not only retains the high priority characteristics of high-stain areas, but also reduces the risk of mutation in path planning through smoothing processing.

[0029] In the path initialization processing stage, the algorithm uses dynamic programming to search for low-resistance, high-priority node sequences in the probability field to generate an initial path sequence. The dynamic programming method first regards each grid point in the probability field as a node, and calculates the transfer cost between nodes based on the field strength value and connectivity. Then, a recursive search is used to find the optimal path from the starting point to the end point. During the search process, the cleaning priority and driving efficiency are comprehensively considered, and areas with high stain intensity but low friction resistance are given priority, while avoiding inaccessible obstacle areas. Through three sub-steps of field strength calculation, node expansion and path scoring, the field strength calculation is used to update the priority of each node, node expansion is used to explore possible path branches, and path scoring sorts the paths according to the total cost (including energy consumption and cleaning effect).

[0030] The initial path sequence is obtained, along with the probability value and priority label of each node. The probability value reflects the field strength of the node in the probability field, and the priority label assigns the priority of the cleaning task to each node based on the stain intensity and difficulty of passage. At the same time, the node label reserves flexibility for dynamic adjustment of the path.

[0031] S3: acquiring a dynamic change region of the initial path sequence in real time, and performing conflict prediction processing on the initial path sequence based on the dynamic change region to obtain a conflict-marked path sequence; In step S3, based on the initial path sequence and node priority labels, dynamically changing areas are acquired in real time and conflict prediction processing is performed to generate a conflict-annotated path sequence. Through dynamic environmental perception and probabilistic modeling, real-time tracking of mobile obstacles in the indoor environment and accurate prediction of path conflicts are achieved, providing a key basis for subsequent path optimization. First, based on the initial path sequence, which contains a series of nodes sorted by priority, each node is accompanied by a probability value and priority label, reflecting the priority of the cleaning task and the feasibility of passage. By collecting sensor data of the indoor environment in real time and comparing it with the environmental characteristics of the initial path sequence, the dynamically changing areas are extracted using incremental differential analysis. Incremental differential analysis identifies the appearance or position change of mobile obstacles (such as pedestrians, pets, or temporarily placed objects) in the environment by comparing the current environmental data with the static feature map at the previous moment. At the same time, the speed distribution and motion trajectory of the obstacles are calculated in combination with the sensor data.

[0032] After identifying dynamically changing areas, a Gaussian mixture model (GMM) is used to construct a spatiotemporal conflict map to predict the potential interference areas of moving obstacles within the next 10 seconds. The Gaussian mixture model generates a probability density distribution by performing multimodal fitting on the obstacle's position and velocity data, characterizing the possible spatiotemporal trajectory of each dynamic obstacle. This model's advantage lies in its ability to handle uncertainty in multi-target scenarios, such as multiple pedestrians moving at different speeds. By weightedly superimposing the trajectory of each obstacle, a comprehensive spatiotemporal conflict map is generated, in which each grid point is annotated with the probability of a path conflict within the next 10 seconds. The conflict probability is calculated based on the obstacle's speed, direction, and proximity to nodes in the initial path sequence. Probabilistic inference further quantifies the interference risk value of each node. Data differencing is used to update the characteristics of the dynamic area, model fitting is used to optimize the parameters of the Gaussian mixture model, and probabilistic inference is used to calculate the conditional probability of a conflict based on a Bayesian framework.

[0033] The entire conflict prediction process is updated twice per second to ensure real-time perception of dynamic environments. Specifically, sensor data is refreshed every 500 milliseconds. Efficient parallel computing rapidly completes differential analysis and model fitting to generate an up-to-date spatiotemporal conflict map. Conflict probability and interference risk values ​​are then mapped to each node in the initial path sequence, outputting a conflict-annotated path sequence. This path sequence retains the node structure of the initial path but adds an interference risk value for each node. The risk value ranges from 0 to 1, with higher values ​​indicating a greater likelihood of interference from dynamic obstacles. During optimization, high-risk nodes can be prioritized for avoidance or path adjustments to reduce conflict probability. The aforementioned steps enable the conversion from the initial path sequence to the conflict-annotated path sequence. This not only captures the dynamic characteristics of the indoor environment but also quantifies the risk of path conflicts through Gaussian mixture models and probabilistic inference. The output conflict-annotated path sequence provides accurate data support for path reconstruction of floor scrubbers in complex dynamic environments. Furthermore, the high-frequency update mechanism ensures the algorithm's rapid response to environmental changes, ensuring real-time and reliable path planning.

[0034] S4: performing multi-objective optimization on the conflict-marked path sequence to obtain a reconstructed path sequence; In step S4, a reconstructed path sequence is generated through multi-objective optimization based on the conflict-annotated path sequence and interference risk values. The conflict-annotated path sequence contains node information for the initial path and the interference risk value for each node, reflecting the potential impact of dynamic obstacles (such as pedestrians or moving objects) on the path. Based on this data, a multi-objective genetic algorithm (NSGA-II or other improved versions) is used for path reconstruction, generating a set of candidate paths by simulating the biological evolution process. First, a population is initialized to randomly generate a set of candidate paths. Each path consists of a series of node coordinates and connection sequences. Subsequently, the candidate paths are updated through a group of crossover and mutation operations. The crossover operation swaps parts of the node sequences of two parent paths to generate a new path, while the mutation operation randomly adjusts the coordinates or connection sequence of certain nodes to increase path diversity.

[0035] During the optimization process, the objective function comprehensively considers three key indicators: clean coverage, path length, and obstacle avoidance safety. Clean coverage is quantified by evaluating whether the path covers high-soil areas (determined by the set of weight vectors in the hierarchical feature graph in step S1). Path length is represented by the sum of the Euclidean distances between nodes. Obstacle avoidance safety is based on the interference risk value in step S3, with nodes with lower risk values ​​being prioritized. The weight of the objective function is dynamically adjusted by the probability value generated in step S2. For example, when the probability field indicates that an area is densely polluted, the weight of the clean coverage will be increased accordingly. This dynamic weighting mechanism ensures that the optimization results adapt to the needs of different indoor environments. Each candidate path is scored through fitness evaluation. The fitness function combines the weighted sum of the three objectives, prioritizing those paths that perform well in covering high-soil areas, shortening path length, and reducing collision risk, while eliminating paths with low fitness to accelerate convergence.

[0036] The entire optimization process consists of three core stages: population initialization, fitness evaluation, and iterative optimization. It gradually converges to a Pareto-optimal solution set over approximately 100 iterations. To improve computational efficiency, parallel computing techniques are used to simultaneously evaluate the expected benefits of multiple candidate paths. The expected benefits are estimated by simulating each path's cleaning performance and conflict probability. For example, paths that cover more contaminated areas and have lower interference risk have higher expected benefits. In each iteration, the optimal path is selected based on its fitness ranking, and the next generation of the population is generated through selection, crossover, and mutation operations. After multiple iterations, a set of Pareto-optimal paths is output, from which the final solution that meets the cleaning task requirements is selected. The output is a reconstructed path sequence consisting of optimized node coordinates and priority labels. Node coordinates are adjusted to avoid high-risk areas, and priority labels reflect the cleaning order and importance of path nodes. This achieves the transition from a conflict-annotated path sequence to a reconstructed path sequence. Through global search and dynamic weight adjustment, the multi-objective genetic algorithm balances cleaning efficiency, path economy, and obstacle avoidance safety. The output reconstructed path sequence not only retains the cleaning objectives of the initial path but also significantly reduces the interference risk of dynamic obstacles.

[0037] S5: performing deep learning control on the reconstructed path sequence based on an experience replay mechanism to obtain a real-time path sequence including a dynamic offset vector; In step S5, based on the reconstructed path sequence and optimized node coordinates, a real-time path sequence containing dynamic offset vectors is generated through deep learning control based on an experience replay mechanism. A deep Q-network (DQN) is used to achieve adaptive path adjustment, ensuring that the floor scrubber can accurately and efficiently perform cleaning tasks in a dynamically changing indoor environment while also addressing real-time changes in stains and obstacles. First, the reconstructed path sequence is input. This sequence contains node coordinates and priority labels that have undergone multi-objective optimization. These nodes constitute the floor scrubber's basic driving route. Based on these nodes and combined with real-time environmental data collected by sensors (such as changes in stain distribution and the location of dynamic obstacles), path control is performed using a deep Q-network. The deep Q-network is a neural network model based on reinforcement learning. The purpose of the deep Q-network is to find the optimal strategy for adjusting the path in a complex environment through a state-action-reward cycle.

[0038] Specifically, the deep Q-network takes path node coordinates, collision probabilities (provided by the collision prediction process in step S3), and real-time soiling changes as state inputs. After state encoding, high-dimensional information such as node coordinates, probability values, and soiling distribution is converted into feature vectors that can be processed by the neural network. The network then predicts Q values ​​through multiple fully connected layers or convolutional layers, outputting the expected reward for each possible action. The action space includes the deviation direction of path nodes (such as left, right, forward, or backward) and local path replanning (such as bypassing temporary obstacles). To drive the network to learn the optimal strategy, a comprehensive reward function is used. This function simultaneously considers cleaning efficiency (i.e., the proportion of soiled areas covered), obstacle avoidance success rate (i.e., the ability to avoid collisions with dynamic obstacles), and energy consumption (i.e., the additional power required for path adjustments). For example, when a floor scrubber successfully bypasses an obstacle and cleans a highly soiled area, the reward value is high, while if a collision occurs or the target area is deviated, the reward value is reduced. Through trial and error and optimization, the network gradually learns to select the action with the highest reward under different states.

[0039] The deep Q network updates its decisions at a high frequency, specifically selecting actions three times per second, to adapt to the rapid changes in the indoor environment. Each time a decision is made, the network predicts the Q value based on the current state and selects the action with the highest Q value to perform path adjustment. At the same time, the algorithm significantly improves learning efficiency through the experience replay mechanism. The experience replay mechanism stores the four-tuple of historical state-action-reward-next state in the experience pool, and randomly extracts high-value samples (such as experience that leads to high rewards or significant path improvements) during training, thereby preventing the model from falling into local optimality and accelerating convergence. It is particularly suitable for dynamic changes in indoor environments (such as pedestrian movement or stain diffusion). The indoor environment has a certain degree of randomness. Experience replay can make full use of the effective information in historical data and enhance the generalization ability of the model.

[0040] After being regulated by the deep Q network, a real-time path sequence is output, and a dynamic offset vector is attached. The dynamic offset vector describes the adjustment direction and amplitude of each node in the real-time environment. For example, when the sensor detects a temporary obstacle near a node, the vector may instruct the node to offset 0.5 meters to the left. These offset vectors ensure that the floor scrubber can flexibly respond to local interference without deviating from the overall cleaning goal. The output real-time path sequence provides precise guidance for the cleaning execution of step S6, and provides basic data for dynamic adjustment of the closed-loop self-learning mechanism. Through the above process, step S5 realizes the conversion from static reconstructed path to dynamic real-time path. The deep Q network combined with the experience replay mechanism not only improves the adaptability and robustness of path control, but also provides strong technical support for the efficient cleaning of floor scrubbers in complex indoor environments through efficient decision updates and reward-driven learning.

[0041] S6: Controlling the floor scrubber to perform cleaning and obtain execution data according to the real-time path sequence, performing closed-loop self-learning processing on the execution data based on a closed-loop mechanism, and outputting a cleaning effect report.

[0042] In step S6, the scrubber is controlled to perform cleaning tasks and collect execution data. A closed-loop mechanism is used for self-learning processing, ultimately outputting a cleaning effect report and updating the feature map. This forms an adaptive path planning and optimization system, which not only ensures the efficient completion of cleaning tasks but also continuously improves system performance through data-driven model updates. The scrubber is controlled to perform cleaning tasks according to a real-time path sequence. The path nodes and dynamic offset vectors regulated by the deep Q network provide the scrubber with precise navigation instructions, adjusting its driving direction or taking a detour when encountering temporary obstacles. During execution, the scrubber records execution data in real time through sensors, including actual cleaning coverage (i.e., the proportion of the cleaned area to the target area), collision events (e.g., minor contact with obstacles), and energy consumption feedback (i.e., the power consumption required to execute the path).

[0043] In the closed-loop mechanism, the collected execution data is systematically analyzed and the data is inferred and processed through a Bayesian network. The Bayesian network uses the coverage, collision events and energy consumption information in the execution data, combined with the hierarchical feature map, to update the weight vector set in the feature map. The adjustment of the weight vector reflects the dynamic changes in the stain distribution and resistance (such as ground friction or obstacle density) in the indoor environment. If the execution data shows that the stain concentration in a certain area is higher than expected, the Bayesian network will increase the weight of the area in the feature map, thereby giving priority to covering the area in the next path planning. At the same time, the deep Q network uses the execution data for online fine-tuning, and optimizes the action strategy by using feedback data from actual execution (such as collision events or cleaning efficiency) as new state input. If a path node repeatedly causes collisions, the deep Q network will adjust the offset vector of the node or replan the local path to improve the success rate of obstacle avoidance. By iterating once per task cycle, it ensures that the model can quickly adapt to environmental changes.

[0044] Feedback collection ensures the integrity and real-time nature of execution data. Bayesian reasoning integrates multi-source data through probabilistic models to generate more accurate feature maps. Model updates improve the accuracy of path control through parameter optimization of the deep Q network. Finally, the algorithm outputs a cleaning effect report, which includes detailed coverage statistics (such as the percentage distribution of clean areas), energy consumption analysis, and updated feature maps. It not only provides users with an evaluation of the execution effect of the cleaning task, but also forms an adaptive closed-loop system through feedback to the hierarchical feature map of step S1. The closed-loop mechanism enables the system to continuously learn from environmental changes. When the indoor layout changes or the type of stain changes, the updated feature map will guide subsequent path planning to better meet actual needs.

[0045] Through the above process, a complete flow from real-time path execution to closed-loop self-learning is achieved. When performing cleaning tasks, the floor scrubber can flexibly respond to environmental changes based on the dynamic offset vector. The collaborative work of the Bayesian network and the deep Q network ensures the continuous optimization of feature maps and action strategies. This adaptive system not only improves cleaning efficiency and obstacle avoidance capabilities, but also provides users with transparent task evaluation through cleaning effect reports, ensuring the robustness and sustainability of the path planning method.

[0046] In one embodiment, the steps of collecting multi-source data of an indoor environment, clustering the multi-source data, and performing hierarchical modeling on the clustering results to obtain a hierarchical feature map including a weight vector set include: Collecting data from the indoor environment based on a sensor array to obtain a multi-source original data set, wherein the multi-source original data set includes three-dimensional spatial coordinates, ground friction resistance, and stain density distribution; performing noise filtering on the multi-source original data set to obtain a high-confidence data set; Extracting multi-dimensional features of the high-confidence data set to obtain a feature vector set, and performing hierarchical clustering processing on the high-confidence data set based on the feature vector set to obtain a multi-scale grouped data set; Performing hierarchical modeling on the multi-scale grouped data set to obtain a preliminary feature map; A weight vector assignment process is performed on the multi-scale grouped data set according to the preliminary feature map to obtain a hierarchical feature map including a weight vector set.

[0047] In the above embodiment, a multi-source raw data set is generated by collecting and processing multi-sensor data from the indoor environment using a variety of sensors on a floor scrubber. The floor scrubber's sensors include a lidar, ultrasonic sensor, and infrared camera, which collect spatial geometric boundary data, floor material friction coefficient data, and thermal imaging data on stain distribution, respectively. The acquisition process utilizes a time synchronization mechanism to ensure the alignment of timestamps for each data type. A high-speed sampling module captures environmental information at a frequency of over a thousand times per second. The multi-source raw data set includes initial data points for 3D spatial coordinates, floor friction resistance, and stain density distribution.

[0048] In response to noise interference in multi-source original data sets, a dynamic threshold generation module is used to calculate local statistical features based on the temporal and spatial distribution characteristics of data points, and automatically adjust the filtering threshold to eliminate outliers and low-quality data points, retaining reliable data related to floor cleaning tasks. Specifically, the parallel computing framework can be used to accelerate execution to ensure real-time performance. The resulting high-confidence data set contains the cleaned spatial coordinates, friction coefficient, and stain density data.

[0049] For the spatial coordinates, friction coefficient, and stain density data in the high-confidence dataset, eigendecomposition techniques are used to extract spatial geometric constraint features, ground material resistance features, and stain distribution intensity features, respectively. Weighted mapping is used to project the multidimensional data into a unified feature space, generating a multidimensional feature description for each data point. The feature vector set consists of the feature descriptions of all data points, preserving the spatial correlation of the environment and the attributes related to the cleaning task. Hierarchical clustering is performed on the data points based on the feature vector set to obtain a multi-scale grouped dataset. Specifically, based on the multi-dimensional feature descriptions in the feature vector set, a hierarchical clustering algorithm is used to group data points according to spatial proximity and cleaning attribute similarity. The clustering process involves calculating the Euclidean distance and attribute correlation coefficient between feature vectors. A clustering tree is then gradually constructed through a bottom-up merging strategy to generate multi-scale grouping results. The multi-scale grouped dataset contains spatial regions of different granularities, each consisting of data points with similar characteristics.

[0050] Hierarchical modeling technology is used to construct a multi-level environmental representation for multi-scale grouped datasets. The bottom-level model mainly analyzes the spatial geometry and stain distribution characteristics of the local area, the middle-level model integrates the connectivity of adjacent areas, and the high-level model abstracts the global spatial structure. The modeling process is to adjust the association weights between each level through iterative optimization. The preliminary feature map consists of a multi-level structure, which describes the feature distribution of the indoor environment from local to global, providing a basis for weight allocation.

[0051] Based on the preliminary feature map, weight vectors are assigned to data points, resulting in a hierarchical feature map containing a set of weight vectors. Specifically, a weight vector is calculated for each data point, taking into account factors such as spatial connectivity, frictional resistance, and stain density. Normalization techniques are used to ensure consistency across dimensions, and multi-threaded parallel computing is employed to accelerate weight assignment. This generates a set of weight vectors that incorporate spatial location, cleaning priority, and regional relevance. The hierarchical feature map, containing this set of weight vectors, ultimately integrates the multi-level structure and weight information.

[0052] In one example, the step of performing a probability field construction process on the hierarchical feature map and performing a path initialization process on the hierarchical feature map after the probability field construction process to obtain an initial path sequence includes: Performing region segmentation processing on the hierarchical feature map based on spatial clustering to obtain a partition feature matrix; Performing probability field construction processing on the partition feature matrix to obtain an initial probability field model; Performing node gridding processing on the initial probability field model, and adding travel cost and cleaning priority labels to the initial probability field model after the node gridding processing to obtain a probability grid map; Performing path search processing on the probability grid graph through dynamic programming to obtain a candidate path set; Path optimization processing is performed on the candidate path set to obtain a preferred path sequence, and priority marking is performed on the preferred path sequence based on node priority analysis to obtain an initial path sequence.

[0053] In the above embodiment, based on a hierarchical feature map, spatial clustering and semantic segmentation techniques are used to segment the feature map into regions, generating a partition feature matrix. By analyzing the weight vector set in the hierarchical feature map and combining the regional stain distribution, material properties, and obstacle density, the indoor environment is divided into multiple sub-regions with independent cleaning requirements. Each sub-region is assigned a feature vector containing quantized values ​​of stain intensity, floor friction coefficient, and traffic complexity. The segmentation process is implemented through iterative clustering and boundary optimization to ensure smooth partition boundaries and consistent regional features, and the partition feature matrix is ​​output.

[0054] The initial probability field model is constructed by weighted superposition and probability distribution calculation based on the partition feature matrix. Based on the eigenvectors in the partition feature matrix, a probability field representing cleaning priority and driving difficulty is constructed, integrating stain intensity, friction resistance, and traffic complexity. A Markov random field model is then used to iteratively optimize the probability values ​​for each sub-region. The field strength distribution is adjusted through multiple iterations to balance cleaning requirements and driving efficiency, ensuring that highly soiled areas have higher probability values ​​and low-resistance areas have higher traffic priority. This output is the initial probability field model.

[0055] The probability field is meshed using spatial discretization techniques to generate a probability grid map. Specifically, the initial probability field model is divided into uniform grid cells, each corresponding to a node. Nodes inherit the probability value and eigenvector of the corresponding position in the probability field. The meshing process uses adaptive resolution adjustment to ensure a denser grid distribution in high-priority areas and a sparser grid in low-priority areas to optimize computational efficiency. Each node is labeled with a travel cost and a cleaning priority.

[0056] Based on a probabilistic grid graph, a path search process is performed on the probabilistic grid graph using dynamic programming and heuristic search algorithms to generate a set of candidate paths. Using the nodes in the probabilistic grid graph as the search space, the A* algorithm and dynamic programming are combined to search for multiple low-cost, high-priority path sequences from the starting point to the end point, combining node probabilities, travel costs, and cleanliness priority labels. Taking into account path length, number of turns, and regional priority, the expansion cost of each node is evaluated, prioritizing nodes with high probability and low resistance. This generates a set of candidate paths, each with a score and node sequence.

[0057] A multi-objective optimization algorithm is used to optimize the candidate path set to obtain a preferred path sequence. A genetic algorithm is used to iteratively optimize the candidate path set, combining the path's cleaning coverage, driving distance, and time efficiency. Paths that meet the multi-objective constraints are screened through crossover, mutation, and selection operations. A weighted scoring mechanism is then used to balance cleaning effect and driving efficiency. Local smoothing is performed on nodes in the path to reduce sharp turns and redundant movement, resulting in higher cleaning coverage and lower driving costs. Each node in the preferred path sequence is then prioritized and a priority label is generated based on the node's corresponding probability value, pollution intensity, and traffic complexity. The labels are divided into three levels: high priority, medium priority, and low priority. Priority labeling is achieved through weighted calculation and threshold segmentation to ensure that nodes in highly polluted areas are marked as high priority. A consistency check is performed on the entire path to ensure its coherence and executability, resulting in an initial path sequence containing a complete node sequence and priority labels.

[0058] In one example, the step of acquiring the dynamic change region of the initial path sequence in real time and performing conflict prediction processing on the initial path sequence based on the dynamic change region to obtain the conflict-marked path sequence includes: Performing dynamic environment fusion on the initial path sequence to obtain an environment perception data set; Performing incremental differential analysis on the environmental perception data set to obtain a dynamically changing region set; Constructing a spatiotemporal conflict graph for the initial path sequence according to the dynamically changing region set to obtain a spatiotemporal conflict probability graph; Setting a dynamic conflict probability threshold, and performing conflict threshold segmentation processing on the spatiotemporal conflict probability graph based on the dynamic conflict probability threshold to obtain a conflict risk node set; Performing conflict marking processing on the initial path sequence according to the conflict risk node set to obtain a preliminary conflict path sequence; The preliminary conflict path sequence is smoothed and consistency-checked to obtain a conflict-marked path sequence.

[0059] In the above embodiment, by dynamically fusion the initial path sequence, an environmental perception data set is generated, laying the foundation for the subsequent analysis of dynamically changing areas. This process first relies on the various sensors carried by the floor scrubber, including lidar, ultrasonic sensors, and visual cameras. These sensors collect multi-source data of the indoor environment in real time. For example, lidar provides high-precision distance measurement and generates point cloud data, ultrasonic sensors detect close-range obstacles, and visual cameras capture texture and color information in the environment. These multi-source data have different sampling frequencies and formats, so they need to be aligned through time series synchronization technology to ensure data consistency in the time dimension. Specifically, the algorithm uses the timestamp alignment method to map the data of each sensor to a unified time axis, which is expressed as t sync =argmin t ∑ i ∣t i -t|, where t i is the timestamp of each sensor data, t sync The Kalman filter method is further used to smooth and reduce noise on the sensor data, and the 3D point cloud reconstruction technology is combined to generate an environmental perception data set containing obstacle location, speed, and type information.

[0060] Based on the environmental perception data set, incremental differential analysis is performed on environmental data from consecutive time frames to generate a set of dynamically changing regions. By comparing the differences in point cloud data between adjacent time frames, the spatial position changes of moving obstacles or newly appearing objects in the environment are identified. Region segmentation technology is used to divide the difference regions into independent change units. The instantaneous speed and direction of each change unit are calculated using a motion trajectory tracking algorithm, generating a set of dynamically changing regions containing position, speed, and trajectory characteristics.

[0061] Based on the set of dynamically changing regions, a spatiotemporal conflict graph is constructed for the path nodes in the initial path sequence, generating a spatiotemporal conflict probability map. A Gaussian mixture model is used to probabilistically model the trajectory data for the dynamically changing regions, predicting the potential movement range of each changing region within the next 10 seconds. A three-dimensional spatiotemporal conflict graph is constructed by combining the coordinates and timestamps of the nodes in the path sequence. Each node is assigned a conflict probability value based on distance and time, generating a spatiotemporal conflict probability map that includes the conflict probabilities of all nodes.

[0062] The spatiotemporal conflict probability graph is segmented using conflict thresholds to generate a set of conflict risk nodes. A dynamic conflict probability threshold is set to filter out path nodes in the spatiotemporal conflict probability graph with conflict probabilities above the preset threshold. Cluster analysis is then used to group high-risk nodes, generating a set of conflict risk nodes containing their locations, times, and probability values. This process also retains the background information of low-risk nodes, providing conflict area demarcation for path sequence annotation.

[0063] Conflict annotation is performed on the nodes in the initial path sequence based on the conflict risk node set to obtain a preliminary conflict-annotated path sequence. The node information in the conflict risk node set is mapped to the initial path sequence, and the conflict status, conflict time window, and conflict probability value of each node are marked. Time series correlation technology is used to ensure the continuity and consistency of the annotations, and a preliminary conflict-annotated path sequence containing all node conflict annotations is generated.

[0064] The initial conflict-annotated path sequence is smoothed and consistency-checked to obtain the final conflict-annotated path sequence. The conflict probabilities in the annotation sequence are smoothed using a sliding window technique to eliminate annotation mutations caused by sensor noise or transient interference. A Markov random field model is used to perform consistency checks on the annotation sequence, correcting annotation errors that do not conform to path continuity or temporal logic. This results in a smooth and logically consistent conflict-annotated path sequence that includes the conflict status, probability, and time information for each node, providing reliable path data.

[0065] In one example, the step of performing multi-objective optimization on the conflict-annotated path sequence to obtain a reconstructed path sequence includes: Extracting spatiotemporal distribution features of the conflicting annotated path sequence, and performing segmentation decomposition processing on the conflicting annotated path sequence based on the spatiotemporal distribution features to obtain a path segment set; Constructing a multi-objective evaluation function for the path segment set to obtain a segment objective weight matrix; Performing genetic optimization processing on the path segment set based on the segment target weight moment to obtain a candidate path segment set; Performing dynamic risk assessment on the candidate path segment set based on a Bayesian probability model to obtain a risk path segment set; Performing path integration processing on the risk path segment set to obtain a preliminary reconstructed path sequence; The preliminary reconstructed path sequence is smoothed and optimized by Bezier curve fitting to obtain a reconstructed path sequence.

[0066] In the above embodiment, by analyzing the positions of conflict annotation points and the boundaries of dynamically changing areas, the path sequence is divided into multiple continuous path segments, each segment contains a local conflict annotation and a corresponding environmental feature vector, and the spatiotemporal distribution characteristics of the conflict annotation points are extracted. Then, a dynamic time segmentation algorithm is used to determine the segmentation points based on the conflict density and regional complexity. The segmentation is iteratively performed according to the geometric continuity of the path and the heterogeneity of the conflict distribution to ensure that the length and conflict complexity of each path segment are balanced. The output path segment set is a structured data set containing the segment start point, end point and local conflict annotation.

[0067] A multi-objective evaluation function is constructed for each path segment, generating a segment-by-segment objective weight matrix. This integrates the target requirements for cleaning coverage, path length, and obstacle avoidance safety. A multi-objective evaluation function is constructed for each path segment. The environmental characteristics of each segment are first extracted, including stain distribution density, obstacle concentration, and dynamic change frequency. The weights of each objective are then dynamically calculated using an entropy method. These weights are adaptively adjusted based on the environmental differences within the segment. The evaluation function values ​​for each segment are then calculated in parallel, generating a comprehensive weight matrix encompassing cleaning efficiency, path cost, and safety constraints. The segment-by-segment objective weight matrix is ​​then output as two-dimensional structured data.

[0068] According to the segmented objective weight matrix, the path segment set is genetically optimized to obtain a set of candidate path segment sets. A multi-objective genetic algorithm is used to generate a set of candidate path segment combinations by initializing the population. This includes crossover and mutation operations for each segment. The crossover operation generates new combinations by exchanging local path fragments of different segments. The mutation operation introduces diversity by randomly adjusting the path node coordinates. The population is iteratively optimized according to the evaluation function of the segmented objective weight matrix. Convergence is achieved after about 100 iterations. Candidate solutions that meet high clean coverage, short path length, and low obstacle avoidance risk are preferentially retained. The output candidate path segment set is a set containing multiple optimized path segment combinations, with each segment accompanied by optimized node coordinates and priority. The form of the genetic optimization process is: ; in, is the overall fitness of the i-th candidate segment combination; K is the number of objective functions; is the weight of the i-th path segment to the k-th target; is the score of the i-th individual in the k-th objective function; The average score of all individuals under this goal; is a nonlinear adjustment parameter used for smoothing the score; The power exponent for adjusting the nonlinearity of the target impact; is the diversity penalty coefficient; is the genetic difference measure of individual i (such as Hamming distance or segmental difference), and e is the base of the natural logarithm.

[0069] Perform dynamic risk assessment on the candidate path segment set to obtain a risk-labeled path segment set. By acquiring dynamic change data of the indoor environment in real time, including the speed and trajectory prediction of moving obstacles, risk assessment is performed on each candidate path segment. The Bayesian probability model is used to estimate the conflict probability of the path segment in the dynamic environment. The sensitivity of the risk assessment is adjusted in combination with the step-segment target weight matrix. By processing the expected risk value of each segment in parallel, path segments with risks below the set threshold are screened out, and the risk level and key conflict nodes are labeled for each segment. The output risk-labeled path segment set is a structured set containing risk assessment results and optimized node coordinates. The form of dynamic risk assessment is: ; in, The posterior probability that path segment i has high risk; is the risk status of path segment i; is the environmental feature vector of path segment i; Given the environmental data under the condition that path segment i has conflict risk Likelihood function of ; is the prior risk probability of path segment i; is a normalization factor used to ensure that the sum of the posterior probabilities is 1, which is the probability of path segment i in the conflict (r=1) and non-conflict (r=0) states, given the environmental data The joint probability of and, The risk status of the path segment, which can be 0 (no conflict) or 1 (conflict); is the prior probability of path segment i in risk state r.

[0070] A path integration process is performed on the risk-labeled path segment set to obtain a preliminary reconstructed path sequence. The segments in the risk-labeled path segment set are reconnected using a global optimization strategy. A path connection graph is constructed based on the geometric continuity and risk level between segments. An improved A* algorithm is then used to search for the globally optimal connection solution. The search process comprehensively considers the clean coverage between segments, path length, and obstacle avoidance safety. The coordinates and order of the connection points are iteratively adjusted to generate a smooth and efficient path sequence. The resulting preliminary reconstructed path sequence contains the coordinates of the complete path nodes and segment priorities.

[0071] The preliminary reconstructed path sequence is smoothed and optimized to obtain a reconstructed path sequence. The path sequence is smoothed using Bezier curve fitting and a local adjustment strategy. High-curvature nodes and potential oscillation regions in the preliminary reconstructed path sequence are first identified. A smooth Bezier curve is then generated based on curvature and path length constraints. The smoothing error for each path segment is calculated in parallel, and the control point positions are iteratively adjusted to minimize path deviation while retaining the risk annotation information from step 4. This ensures that the smoothed path meets obstacle avoidance safety and clean coverage requirements. The reconstructed path sequence is then output as the final path data, which includes smoothed node coordinates, optimization priorities, and dynamic constraints.

[0072] In one example, the step of performing deep learning control on the reconstructed path sequence based on the experience replay mechanism to obtain a real-time path sequence including a dynamic offset vector includes: Performing state feature extraction processing on the reconstructed path sequence according to the dynamically changing region to obtain a path state feature set; Constructing an experience replay library, and performing sample screening processing on the experience replay library according to the path state feature set to obtain a high-value experience sample set; Training a deep reinforcement learning network based on the high-value experience samples to obtain a path control strategy model; Performing action prediction processing on the reconstructed path sequence according to the path control strategy model to obtain an initial adjusted path sequence; Performing dynamic offset vector generation processing on the initial adjustment path sequence to obtain a dynamic offset vector set; Path reconstruction processing is performed on the initial adjusted path sequence according to the dynamic offset vector set to obtain a real-time path sequence.

[0073] In the above embodiment, based on the reconstructed path sequence, combined with the dynamically changing areas of the indoor environment and the real-time positioning data of the floor-cleaning vehicle, state feature extraction and processing are performed on the node coordinates, path curvature, and conflict annotation information in the path sequence to obtain a path state feature set. A multidimensional feature encoding algorithm is used to convert the geometric information of the path nodes, the position distribution of dynamic obstacles, and the stain density in the cleaning area into a high-dimensional state vector. A convolutional neural network is used to extract features from the spatial relationships between nodes, and time series analysis is used to capture the dynamic changing trends of the path sequence. A unique state identifier is assigned to each node, and feature sampling is performed according to the time step, with feature updates performed once per second, ensuring that the state feature set fully represents the real-time dynamic characteristics of the path sequence.

[0074] Based on the path state feature set, the system filters and processes historical path data and control records in the pre-built experience replay library to generate a high-value experience sample set. Using a priority sampling algorithm, it extracts historical samples with the highest similarity to the current path state feature set from the experience replay library. The similarity between state vectors is calculated based on cosine distance, and the samples are sorted based on the reward values ​​of the historical samples, prioritizing samples with high cleaning efficiency, high obstacle avoidance success rate, and low energy consumption. The experience replay library is sharded for storage and management, grouping samples by timestamp and scene complexity. Sample priorities are updated every second to filter out a fixed number of high-value sample sets.

[0075] The process of developing a path control strategy model involves deploying a deep Q-network, which takes a set of path state features as input and outputs path adjustment actions, including node offset direction, local path replanning, and speed adjustment. Q-values ​​are predicted for these state features using a multi-layer, fully connected neural network. Actions are selected using an ε-greedy strategy, with the reward function taking into account cleaning coverage, obstacle avoidance success rate, and the vehicle's energy consumption. Forward propagation is performed at a fixed frequency to calculate Q-values, and network weights are updated through backpropagation. Training is performed three times per second, accelerating model convergence based on a high-value set of empirical samples, resulting in a path control strategy model that can dynamically adapt to environmental changes.

[0076] The path control strategy model predicts the actions of the reconstructed path sequence to obtain the initial adjusted path sequence. Specifically, the path control strategy model performs forward reasoning on the current path state feature set to predict the adjustment action for each path node, including the offset vector, path curvature correction, and local replanning instructions. The optimal action combination is selected using the Q-value distribution output by the model. This involves reasoning about the path sequence node by node, verifying the action prediction results with real-time sensor data, and updating the prediction results once a second to ensure that the initial adjusted path sequence retains the structure of the reconstructed path sequence.

[0077] Based on the initial adjusted path sequence, the motion prediction results of the path nodes are processed for dynamic offset vector generation to obtain a dynamic offset vector set. A vector decomposition algorithm is used to convert the motion instructions in the initial adjusted path sequence into three-dimensional offset vectors, including horizontal offset, vertical offset, and angular adjustment. The algorithm combines the kinematic constraints of the floor scrubber and the spatial distribution of environmental obstacles to perform constrained optimization of the offset vector for each node, ensuring that the offset path meets passability requirements. The execution logic includes offset calculation of path nodes according to time steps, fine-tuning the vector based on real-time environmental feedback, and performing two vector updates per second to generate a dynamic offset vector set.

[0078] Based on the dynamic offset vector set, the initial adjusted path sequence is reconstructed to produce a real-time path sequence that includes the dynamic offset vectors. The dynamic offset vector set is applied to the initial adjusted path sequence based on a path smoothing algorithm, updating the coordinates of the path nodes to generate a smooth and continuous path trajectory. Bezier curve interpolation technology is used to smooth the path, while a collision detection algorithm verifies the path feasibility. Offset vectors are applied sequentially to the nodes, and distance constraints between the path and obstacles are checked in real time. Path reconstruction is performed once per second to generate a real-time path sequence that retains the structural characteristics of the initial adjusted path sequence and adds dynamic offset vectors, ensuring that the floor scrubber can efficiently perform cleaning tasks according to the real-time path sequence.

[0079] In one embodiment, the steps of controlling the floor scrubber to perform cleaning and obtain execution data according to the real-time path sequence, performing closed-loop self-learning processing on the execution data based on a closed-loop mechanism, and outputting a cleaning effect report include: Controlling the floor scrubber to perform a cleaning task according to the real-time path sequence to obtain an execution data set; Performing dynamic feature extraction processing on the execution data set based on a temporal convolutional network to obtain a dynamic feature vector group; performing update correction processing on the hierarchical feature map according to the dynamic feature vector group to obtain an updated feature map; Performing closed-loop feedback analysis on the updated feature map based on Markov decision making to obtain a feedback control parameter set for controlling the path and action strategy; Dynamically adjusting the real-time path sequence according to the feedback control parameter set to obtain an optimized path sequence; The execution effect of the optimized path sequence is evaluated to obtain a cleaning effect report including an updated self-learning model, and the updated self-learning model is fed back to the hierarchical modeling process to form a closed-loop adaptive system.

[0080] In the above embodiment, based on the real-time path sequence, combined with the sensor data and motion control module of the floor scrubber, the floor scrubber is controlled and processed to perform the cleaning task, and an execution data set including the actual operation trajectory, cleaning coverage area, collision event records and energy consumption data is obtained.

[0081] Dynamic feature extraction is performed on the execution data set to generate a dynamic feature vector group reflecting the execution status of the cleaning task. Based on the actual operation trajectory, coverage area, collision events, and energy consumption data in the execution data set, principal component analysis and temporal convolutional networks are used to extract key features, including the spatial distribution characteristics of cleaning coverage, the spatiotemporal correlation characteristics of collision events, and the time series characteristics of energy consumption. Multidimensional decomposition and reconstruction of the execution data generate a dynamic feature vector group containing coverage deviation, collision probability, and energy consumption fluctuation rate, ensuring that the feature vector can represent the dynamic changes of the cleaning task.

[0082] Based on the dynamic feature vector group and the hierarchical feature map, the initial feature map is updated and corrected to produce an updated feature map that reflects dynamic environmental changes. Using a Bayesian inference mechanism, the coverage deviation, collision probability, and energy consumption fluctuation rate in the dynamic feature vector group are compared and analyzed with the weight vector set of the initial feature map. The probability of environmental change in each region is calculated, and the weight values ​​of each grid in the feature map are updated. The updated feature map is generated through probabilistic mapping of feature vectors, iterative adjustment of weight vectors, and spatial reconstruction of the feature map.

[0083] Closed-loop feedback analysis is performed on the updated feature graph to generate a set of feedback control parameters for regulating the path and action strategy. A deep reinforcement learning framework is used to analyze the performance of cleaning tasks, combining the weight distribution and dynamic feature vector group in the updated feature graph. The system calculates the priority of areas with insufficient coverage, the avoidance coefficient for areas with high collision risk, and the control parameters for energy optimization. This implementation involves feedback analysis and iterative parameter optimization based on a Markov decision process, generating a set of feedback control parameters that includes path adjustment coefficients, action priorities, and energy constraints.

[0084] The real-time path sequence is dynamically adjusted based on the feedback control parameter set to generate an optimized path sequence. Based on the path adjustment coefficients and action priorities in the feedback control parameter set, combined with the scrubber's real-time position and environmental perception data, the real-time path sequence is locally replanned. Node coordinates and offset vectors in the path are adjusted to avoid collision-prone areas and prioritize areas with low cleaning rates. Path length and number of turns are optimized based on energy consumption constraints. Dynamic interpolation of path nodes and trajectory smoothing are applied to generate the optimized path sequence.

[0085] The optimized path sequence is evaluated for execution effectiveness, resulting in a cleaning performance report that includes cleaning coverage statistics, energy consumption analysis, and collision avoidance effectiveness, as well as updates to the self-learning model. Based on the optimized path sequence, the floor scrubber is instructed to perform the cleaning task again, collecting new execution data. By comparing the consistency of the new execution data with the updated feature map, the effectiveness of the path optimization is evaluated, and the coverage improvement ratio, collision reduction rate, and energy consumption reduction rate are calculated. Simultaneously, a deep Q-network is used to fine-tune the self-learning model's policy network online, updating model parameters to adapt to environmental changes. A cleaning performance report is output, and the updated self-learning model is fed back into the feature map modeling process in step one, forming a closed-loop adaptive system.

[0086] Reference Figure 2 , a path planning device for an indoor floor washing vehicle, comprising: The acquisition module 100 is used to collect multi-source data of the indoor environment, perform clustering processing on the multi-source data, and perform hierarchical modeling on the clustering results to obtain a hierarchical feature map including a weight vector set; A construction module 200 is configured to perform a probability field construction process on the hierarchical feature map, and perform a path initialization process on the hierarchical feature map after the probability field construction process to obtain an initial path sequence; The prediction module 300 is configured to obtain a dynamic change region of the initial path sequence in real time, and perform conflict prediction processing on the initial path sequence based on the dynamic change region to obtain a conflict-marked path sequence; An optimization module 400 is configured to perform multi-objective optimization on the conflict-marked path sequence to obtain a reconstructed path sequence; A control module 500 is configured to perform deep learning control on the reconstructed path sequence based on an experience replay mechanism to obtain a real-time path sequence including a dynamic offset vector; The control module 600 is used to control the floor scrubber to perform cleaning and obtain execution data according to the real-time path sequence, perform closed-loop self-learning processing on the execution data based on a closed-loop mechanism, and output a cleaning effect report.

[0087] In the above embodiment, by collecting multi-source data of the indoor environment and clustering the collected data, and then performing hierarchical modeling on the clustering results, a hierarchical feature map containing a weight vector set is constructed to achieve a high-dimensional abstract expression of the indoor environment; on this basis, through probability field construction processing, the path pass probability in different areas is quantitatively modeled, and the path initialization processing is performed in combination with the model to generate an initial path sequence with global feasibility and regional adaptability, thereby improving the basic accuracy and environmental adaptability of path planning.

[0088] By perceiving the dynamic changes in the areas involved in the initial path sequence in real time, it is possible to identify the impact of potential changes on path traffic, predict conflicts based on this, and generate an information-based path sequence with conflict annotations. Based on a multi-objective optimization algorithm, it takes into account path length, obstacle avoidance efficiency, and execution continuity, and obtains a path result with global optimization characteristics. The reconstructed path is dynamically regulated through a deep learning model, giving the path the ability to offset in real time and ensuring the path's robustness in complex dynamic scenarios. Furthermore, through a closed-loop mechanism, execution data feedback is used for self-learning, achieving the continuous evolution of the path planning strategy, and ultimately generating a data report reflecting the cleaning effect to continuously ensure the quality of the cleaning operation.

[0089] Based on the same idea as the method in the above embodiment, the indoor floor washing vehicle path planning device provided by the present application can implement the method in the above embodiment. For the convenience of explanation, the structural schematic diagram of the device embodiment only shows the parts related to the embodiment of the present application. Those skilled in the art can understand that the illustrated structure does not constitute a limitation on the device, and it may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.

[0090] Reference Figure 3 In the embodiment of the present application, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a network interface and a database connected via a system bus. The processor of the computer design is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store data such as a path planning method for an indoor floor-washing vehicle. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a path planning method for an indoor floor-washing vehicle is implemented.

[0091] An embodiment of the present application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a path planning method for an indoor floor-washing vehicle, characterized in that the method comprises: collecting multi-source data of an indoor environment, clustering the multi-source data, and performing hierarchical modeling on the clustering results to obtain a hierarchical feature map including a weight vector set; performing probability field construction processing on the hierarchical feature map, and performing path initialization processing on the hierarchical feature map after the probability field construction processing to obtain an initial path sequence; obtaining a dynamic change area of ​​the initial path sequence in real time, and performing conflict prediction processing on the initial path sequence based on the dynamic change area to obtain a conflict-labeled path sequence; performing multi-objective optimization on the conflict-labeled path sequence to obtain a reconstructed path sequence; performing deep learning control on the reconstructed path sequence based on an experience replay mechanism to obtain a real-time path sequence including a dynamic offset vector; controlling the floor-washing vehicle to perform cleaning and obtain execution data according to the real-time path sequence, performing closed-loop self-learning processing on the execution data based on a closed-loop mechanism, and outputting a cleaning effect report.

[0092] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When executed, the computer program can include the processes of the above-described method embodiments. Any reference to memory, storage, database, or other media provided herein and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct RAMbus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM).

[0093] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A path planning method for an indoor floor-cleaning vehicle, characterized in that: include: Collecting multi-source data of the indoor environment, clustering the multi-source data, and performing hierarchical modeling on the clustering results to obtain a hierarchical feature map including a weight vector set; Performing a probability field construction process on the hierarchical feature map, and performing a path initialization process on the hierarchical feature map after the probability field construction process to obtain an initial path sequence; Acquire the dynamic change area of ​​the initial path sequence in real time, and perform conflict prediction processing on the initial path sequence based on the dynamic change area to obtain a conflict marked path sequence; Performing multi-objective optimization on the conflict-marked path sequence to obtain a reconstructed path sequence; Performing deep learning control on the reconstructed path sequence based on an experience replay mechanism to obtain a real-time path sequence including a dynamic offset vector; The floor scrubber is controlled to perform cleaning and obtain execution data according to the real-time path sequence, and closed-loop self-learning processing is performed on the execution data based on a closed-loop mechanism to output a cleaning effect report.

2. A path planning method for an indoor floor-cleaning vehicle according to claim 1, characterized in that: The step of collecting multi-source data of the indoor environment, clustering the multi-source data, and performing hierarchical modeling on the clustering results to obtain a hierarchical feature map including a weight vector set includes: Collecting data from the indoor environment based on a sensor array to obtain a multi-source original data set, wherein the multi-source original data set includes three-dimensional spatial coordinates, ground friction resistance, and stain density distribution; performing noise filtering on the multi-source original data set to obtain a high-confidence data set; Extracting multi-dimensional features of the high-confidence data set to obtain a feature vector set, and performing hierarchical clustering processing on the high-confidence data set based on the feature vector set to obtain a multi-scale grouped data set; Performing hierarchical modeling on the multi-scale grouped data set to obtain a preliminary feature map; A weight vector assignment process is performed on the multi-scale grouped data set according to the preliminary feature map to obtain a hierarchical feature map including a weight vector set.

3. The indoor floor-cleaning vehicle path planning method according to claim 1, characterized in that: The step of performing a probability field construction process on the hierarchical feature map and performing a path initialization process on the hierarchical feature map after the probability field construction process to obtain an initial path sequence includes: Performing region segmentation processing on the hierarchical feature map based on spatial clustering to obtain a partition feature matrix; Performing probability field construction processing on the partition feature matrix to obtain an initial probability field model; Performing node gridding processing on the initial probability field model, and adding travel cost and cleaning priority labels to the initial probability field model after the node gridding processing to obtain a probability grid map; Performing path search processing on the probability grid graph through dynamic programming to obtain a candidate path set; Path optimization processing is performed on the candidate path set to obtain a preferred path sequence, and priority marking is performed on the preferred path sequence based on node priority analysis to obtain an initial path sequence.

4. The indoor floor-cleaning vehicle path planning method according to claim 1, characterized in that: The step of acquiring the dynamic change area of ​​the initial path sequence in real time and performing conflict prediction processing on the initial path sequence based on the dynamic change area to obtain a conflict marked path sequence includes: Performing dynamic environment fusion on the initial path sequence to obtain an environment perception data set; Performing incremental differential analysis on the environmental perception data set to obtain a dynamically changing region set; Constructing a spatiotemporal conflict graph for the initial path sequence according to the dynamically changing region set to obtain a spatiotemporal conflict probability graph; Setting a dynamic conflict probability threshold, and performing conflict threshold segmentation processing on the spatiotemporal conflict probability graph based on the dynamic conflict probability threshold to obtain a conflict risk node set; Performing conflict marking processing on the initial path sequence according to the conflict risk node set to obtain a preliminary conflict path sequence; The preliminary conflict path sequence is smoothed and consistency-checked to obtain a conflict-marked path sequence.

5. The indoor floor-cleaning vehicle path planning method according to claim 1, characterized in that: The step of performing multi-objective optimization on the conflict-marked path sequence to obtain a reconstructed path sequence includes: Extracting spatiotemporal distribution features of the conflicting annotated path sequence, and performing segmentation decomposition processing on the conflicting annotated path sequence based on the spatiotemporal distribution features to obtain a path segment set; Constructing a multi-objective evaluation function for the path segment set to obtain a segment objective weight matrix; Performing genetic optimization processing on the path segment set based on the segment target weight moment to obtain a candidate path segment set; Performing dynamic risk assessment on the candidate path segment set based on a Bayesian probability model to obtain a risk path segment set; Performing path integration processing on the risk path segment set to obtain a preliminary reconstructed path sequence; The preliminary reconstructed path sequence is smoothed and optimized by Bezier curve fitting to obtain a reconstructed path sequence.

6. The indoor floor-cleaning vehicle path planning method according to claim 1, characterized in that: The step of performing deep learning control on the reconstructed path sequence based on the experience replay mechanism to obtain a real-time path sequence including a dynamic offset vector includes: Performing state feature extraction processing on the reconstructed path sequence according to the dynamically changing region to obtain a path state feature set; Constructing an experience replay library, and performing sample screening processing on the experience replay library according to the path state feature set to obtain a high-value experience sample set; Training a deep reinforcement learning network based on the high-value experience samples to obtain a path control strategy model; Performing action prediction processing on the reconstructed path sequence according to the path control strategy model to obtain an initial adjusted path sequence; Performing dynamic offset vector generation processing on the initial adjustment path sequence to obtain a dynamic offset vector set; Path reconstruction processing is performed on the initial adjusted path sequence according to the dynamic offset vector set to obtain a real-time path sequence.

7. The indoor floor-cleaning vehicle path planning method according to claim 1, characterized in that: The steps of controlling the floor scrubber to perform cleaning and obtain execution data according to the real-time path sequence, performing closed-loop self-learning processing on the execution data based on a closed-loop mechanism, and outputting a cleaning effect report include: Controlling the floor scrubber to perform a cleaning task according to the real-time path sequence to obtain an execution data set; Performing dynamic feature extraction processing on the execution data set based on a temporal convolutional network to obtain a dynamic feature vector group; performing update correction processing on the hierarchical feature map according to the dynamic feature vector group to obtain an updated feature map; Performing closed-loop feedback analysis on the updated feature map based on Markov decision making to obtain a feedback control parameter set for controlling the path and action strategy; Dynamically adjusting the real-time path sequence according to the feedback control parameter set to obtain an optimized path sequence; The execution effect of the optimized path sequence is evaluated to obtain a cleaning effect report including an updated self-learning model, and the updated self-learning model is fed back to the hierarchical modeling process to form a closed-loop adaptive system.

8. A path planning device for an indoor floor-washing vehicle, characterized in that: include: An acquisition module is used to collect multi-source data of the indoor environment, perform clustering on the multi-source data, and perform hierarchical modeling on the clustering results to obtain a hierarchical feature map including a weight vector set; A construction module is used to perform a probability field construction process on the hierarchical feature map, and perform a path initialization process on the hierarchical feature map after the probability field construction process to obtain an initial path sequence; A prediction module is used to obtain the dynamic change area of ​​the initial path sequence in real time, and perform conflict prediction processing on the initial path sequence based on the dynamic change area to obtain a conflict marked path sequence; An optimization module, configured to perform multi-objective optimization on the conflict-marked path sequence to obtain a reconstructed path sequence; A control module, configured to perform deep learning control on the reconstructed path sequence based on an experience replay mechanism to obtain a real-time path sequence including a dynamic offset vector; The control module is used to control the floor scrubber to perform cleaning and obtain execution data according to the real-time path sequence, perform closed-loop self-learning processing on the execution data based on a closed-loop mechanism, and output a cleaning effect report.

9. A computer device, characterized in that: It includes a processor, a memory, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the indoor floor washing vehicle path planning method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the indoor floor-washing vehicle path planning method according to any one of claims 1 to 7 is implemented.

Images