Progressive privacy calculation method and system for virtual-real fusion scene of intelligent sweeper

Abstract

The invention provides a progressive privacy calculation method and system for a virtual-real fusion scene of an intelligent sweeper, and relates to the technical field of sweepers, and the method comprises the steps: collecting indoor space data through a multi-modal sensor, and constructing a scene knowledge graph through an entangled state registration network and a graph neural network; then, based on deep reinforcement learning and zero-knowledge proof, hierarchical encryption is carried out on the atlas, and scene data versions with different privacy protection intensities are generated through federated learning; and finally, performing path planning in a secure multi-party computing environment, and optimizing a cleaning strategy by using neural symbol reasoning and a graph attention network. According to the invention, scene understanding and path planning of the intelligent sweeper can be realized on the premise of guaranteeing the privacy of the user, the sweeping efficiency is improved, and the user experience is improved.

Description

Technical Field

[0001] The present invention relates to sweeping robot technology, and in particular to a progressive privacy computing method and system for a virtual-reality fusion scenario of an intelligent sweeping robot. Background Art

[0002] The rapid development of smart homes has spurred the demand for smart cleaning equipment, among which sweeping robots, as a typical representative, are gradually becoming a must-have for modern families. In order to improve cleaning efficiency and user experience, researchers are committed to developing smart sweeping robots that can understand and adapt to complex indoor environments. Existing sweeping robot technology mainly relies on SLAM (Simultaneous Localization and Mapping) technology, combined with sensors such as lidar and cameras for navigation and path planning. Some high-end products have also introduced AI object recognition technology to achieve smarter cleaning strategies, such as avoiding obstacles and identifying room types.

[0003] Insufficient privacy protection: Existing sweeping robots usually collect a large amount of indoor environment data, including room layout, furniture placement, etc., which may contain users' private information. Existing privacy protection mechanisms are usually relatively simple, such as data anonymization or encrypted storage, which cannot effectively prevent the risk of data leakage and abuse.

[0004] Limited scene understanding ability: Current sweeping robots mainly understand scenes based on two-dimensional maps and simple object recognition, and lack a deep understanding of scene semantic information. For example, it is difficult for robots to distinguish different functional areas in a room, identify the user's activity status, etc., which limits their intelligence and service capabilities.

[0005] Path planning is not flexible enough: Most existing path planning algorithms are based on pre-set rules or simple optimization strategies, which are difficult to adapt to complex and changing indoor environments. For example, when the environment changes, the robot may need to re-build the map and re-plan the path, which is inefficient. Moreover, it is difficult for existing algorithms to customize cleaning strategies according to the user's personalized needs. Summary of the invention

[0006] The embodiments of the present invention provide a progressive privacy computing method and system for the virtual-reality fusion scenario of an intelligent sweeping robot, which can solve the problems in the prior art.

[0007] According to a first aspect of the embodiments of the present invention, Provides a progressive privacy computing method for the virtual-reality fusion scenario of the smart sweeper, including: The multimodal sensor array of the intelligent sweeper is used to collect indoor space data, obtain point cloud data of micron-level laser radar, three-dimensional image data of depth perception camera, spatial response data of sonar sensor and reflection signal data of millimeter-wave radar, and input the point cloud data, the three-dimensional image data, the spatial response data and the reflection signal data into the entangled state registration network to generate a six-dimensional scene feature tensor in a unified coordinate system; the six-dimensional scene feature tensor is input into the graph neural network for semantic segmentation and time series modeling, and a scene knowledge graph containing the spatial distribution of scene objects, object attribute information and the spatiotemporal interaction relationship between objects is constructed; The scene knowledge graph is input into a deep reinforcement learning network, and a privacy sensitivity scoring matrix is ​​constructed according to the object attribute information in the scene knowledge graph; based on the privacy sensitivity scoring matrix, the nodes in the scene knowledge graph are hierarchically encrypted using a zero-knowledge proof mechanism to generate a multi-level encrypted graph; the multi-level encrypted graph is submitted to a blockchain network node group, and a plurality of scene data versions with different privacy protection strengths are calculated and output through a federated learning framework, each of which contains a corresponding privacy protection strength parameter and a scene information integrity parameter; According to the scenario data version and its corresponding parameters, a path planning optimization model is constructed in a secure multi-party computing environment; the path planning optimization model is input into a neural symbolic reasoning engine to generate an initial cleaning path solution set; a graph attention network is used to perform distributed optimization evaluation on the initial cleaning path solution set, and an optimal cleaning strategy with a credibility quantification indicator is output; based on the optimal cleaning strategy, an optimal cleaning path is selected, and the optimal cleaning path is sent to the smart sweeper for execution through an encrypted channel; real-time status data generated by the smart sweeper during execution is collected, and the real-time status data is transmitted back to the scenario knowledge graph for dynamic updating, so as to achieve continuous optimization of the scenario model.

[0008] Inputting the scene knowledge graph into a deep reinforcement learning network, constructing a privacy sensitivity scoring matrix according to the object attribute information in the scene knowledge graph; based on the privacy sensitivity scoring matrix, hierarchically encrypting the nodes in the scene knowledge graph using a zero-knowledge proof mechanism to generate a multi-level encrypted graph, including: The scene knowledge graph is input into a deep reinforcement learning network, wherein the deep reinforcement learning network comprises a state space module, an action space module and a value assessment module; the state space module receives spatial position information, object usage frequency information and interaction relationship information between objects in the scene knowledge graph; the action space module generates a privacy sensitivity score interval based on the output information of the state space module; the value assessment module calculates the privacy sensitivity score value according to the spatiotemporal correlation of object attributes and user privacy annotation information; The privacy sensitivity score value output by the deep reinforcement learning network is input into a scoring matrix construction module, and the scoring matrix construction module uses an adaptive weight fusion algorithm to normalize the privacy sensitivity score value to generate a privacy sensitivity scoring matrix that characterizes the degree of privacy association between object nodes in the scene knowledge graph; based on the privacy sensitivity scoring matrix, the object nodes are divided into sensitive protection levels higher than a preset sensitivity threshold and sensitive protection levels lower than the preset sensitivity threshold using a spectral clustering algorithm; A zero-knowledge proof mechanism is called to perform hierarchical encryption processing on object nodes with different sensitive protection levels, and the zero-knowledge proof mechanism includes a proof generation unit and a verification execution unit; the proof generation unit constructs an encryption proof using the Groth16 protocol, and adopts a fully homomorphic encryption scheme for object nodes with a sensitive protection level higher than a preset sensitivity threshold, and adopts a lightweight searchable encryption scheme for object nodes with a sensitive protection level lower than the preset sensitivity threshold; the verification execution unit verifies the legitimacy of the encrypted data through a pairing operation, outputs the encryption proof verification result, and reconstructs the encrypted object nodes that have passed the verification into a multi-level encryption graph.

[0009] The multi-level encrypted graph is submitted to the blockchain network node group, and multiple scene data versions with different privacy protection strengths are calculated and output through the federated learning framework. Each scene data version contains corresponding privacy protection strength parameters and scene information integrity parameters including: The multi-level encrypted graph is divided into data shards according to the spatial position relationship, and the data shards are distributed to multiple nodes in the blockchain network node group using a consistent hashing algorithm, wherein the data shards include node encryption attribute data, edge relationship encryption data, and shard index information; a practical Byzantine fault-tolerant consensus algorithm is used in the blockchain network node group to perform data consistency verification, calculation correctness verification, and version synchronization verification on the data shards; A federated learning framework is constructed based on the blockchain network node group, and the federated learning framework includes a feature extraction network, a feature fusion network and a decision network; the feature extraction network extracts features from encrypted data local to the node, the feature fusion network integrates feature information of multiple nodes, and the decision network generates scene data versions with different privacy protection strengths; an adaptive federated optimization algorithm is used in the federated learning framework for model training, and the adaptive federated optimization algorithm performs local model updates through nodes, reduces communication overhead through gradient compression, uses homomorphic encryption to aggregate model parameters, and introduces a differential privacy protection gradient update process; The output results of the federated learning framework are input into the scene data generation module, which generates privacy protection strength parameters according to the degree of data desensitization, encryption strength and noise level, and generates scene information integrity parameters according to topology preservation, attribute retention rate and relationship integrity.

[0010] According to the scenario data version and its corresponding parameters, a path planning optimization model is constructed in a secure multi-party computing environment; the path planning optimization model is input into a neural symbolic reasoning engine to generate an initial cleaning path solution set; a distributed optimization evaluation is performed on the initial cleaning path solution set using a graph attention network, and the output of the optimal cleaning strategy with a trustworthy quantitative indicator includes: According to the scenario data version and its corresponding parameters, a path planning optimization model is constructed in a secure multi-party computing environment; the path planning optimization model adopts a three-party computing protocol based on secret sharing to decompose the computing task into a spatial constraint computing task, a temporal constraint computing task and a result aggregation computing task; the path planning optimization model outputs an initial path planning solution that integrates physical constraints, environmental constraints and privacy constraints; The path planning initial solution is input into the neural symbolic reasoning engine for deep reasoning operation. The encoding layer of the neural symbolic reasoning engine converts the scene spatial relationship, action temporal relationship and constraint rule relationship in the path planning initial solution into a vector representation; the reasoning layer of the neural symbolic reasoning engine extracts implicit rules from historical data based on the vector representation and merges them with expert rules to generate a path planning rule set; the decoding layer of the neural symbolic reasoning engine restores the reasoning result to a symbolic path description based on the path planning rule set to generate an initial cleaning path solution set containing multiple candidate paths; Constructing a graph structure representation for the candidate paths in the initial cleaning path solution set, wherein nodes in the graph structure representation correspond to path points and edges correspond to path segments; inputting the graph structure representation into a graph attention network, and calculating node-level attention weights, edge-level attention weights, and global attention weights through distributed computing nodes; prioritizing the candidate paths based on the node-level attention weights, edge-level attention weights, and global attention weights, and generating a path solution sequence with evaluation scores; A credibility quantization index is calculated for each path in the path solution sequence, and the credibility quantization index is input into the graph attention network as evaluation feedback, the node-level attention weight, the edge-level attention weight and the global attention weight are updated, the path solution sequence is optimized and cleaned, and an optimal cleaning strategy is output.

[0011] Inputting the graph structure representation into the graph attention network, calculating the node-level attention weight, the edge-level attention weight and the global attention weight through the distributed computing nodes; prioritizing the candidate paths based on the node-level attention weight, the edge-level attention weight and the global attention weight, and generating a path solution sequence with evaluation scores includes: Acquire path features, wherein the path features include geometric features of path segment length, path segment curvature, and path segment slope, as well as performance features of energy consumption index and time index; construct the geometric features and the performance features into a node attribute vector, and construct the corner features and distance features between the path segments into an edge attribute vector; The graph structure representation is input into a graph attention network, and the node attribute vector and the edge attribute vector are processed in parallel by distributed computing nodes; the distributed computing nodes construct a query vector and a key vector based on the node attribute vector, and use a self-attention mechanism to calculate the correlation between the query vector and the key vector to obtain a node attention weight; The distributed computing node concatenates the node attribute vector and the edge attribute vector and inputs them into a trainable parameter matrix, calculates the constraint relationship between adjacent path segments to obtain edge-level attention weights; the distributed computing node constructs a global context vector, concatenates the node attribute vector and the global context vector and inputs them into a multi-layer perceptron, calculates the correlation between the overall features and local features of the path to obtain a global attention weight; The distributed computing node performs a weighted combination of the node attention weight, the edge-level attention weight and the global attention weight based on preset weights to obtain a comprehensive path score; the candidate paths are prioritized according to the comprehensive path score to generate a path plan sequence with evaluation scores.

[0012] The optimal cleaning path is selected based on the optimal cleaning strategy, and the optimal cleaning path is sent to the intelligent sweeper through an encrypted channel for execution; the real-time status data generated by the intelligent sweeper during the execution process is collected, and the real-time status data is sent back to the scene knowledge graph for dynamic update, so as to realize continuous optimization of the scene model, including: Based on the optimal cleaning strategy, the path plan is evaluated for execution feasibility, task completion and resource consumption, and the optimal cleaning path is selected from the path plan according to the evaluation results; a secure transmission channel with end-to-end encryption function is constructed, and the secure transmission channel includes a hybrid encryption module, an identity authentication module and a dynamic key update module; the optimal cleaning path is sent to the intelligent sweeper through the secure transmission channel, wherein the sending process adopts segment confirmation and packet loss retransmission mechanism to ensure transmission reliability; Collect real-time status data from the intelligent sweeper, pre-process the real-time status data, perform noise filtering through Kalman filtering, achieve multi-source data fusion through time series alignment, perform anomaly detection through statistical analysis, and generate pre-processed status data; transmit the pre-processed status data back to the scene knowledge graph; The scene knowledge graph is dynamically updated through temporal knowledge fusion and spatial knowledge fusion. The temporal knowledge fusion is processed based on historical data weights, new data credibility and conflicting data reconciliation, and the spatial knowledge fusion is processed based on local update, global consistency and topology optimization. Based on the fusion results, the motion model parameters, environmental model parameters and task model parameters are optimized to achieve continuous optimization of the scene model.

[0013] Based on the optimal cleaning strategy, the path plan is evaluated for execution feasibility, task completion and resource consumption, and the optimal cleaning path is selected from the path plan according to the evaluation results; and a secure transmission channel with end-to-end encryption function is constructed, including: Based on the optimal cleaning strategy, the cleaning area is rasterized and divided, and a path plan to be evaluated is generated; based on the optimal cleaning strategy, an execution feasibility assessment is performed, the path plan to be evaluated is fitted with a Bezier curve to obtain path curvature data, a kinematic model is established based on the path curvature data to calculate the maximum allowable speed of each path segment, the attitude stability margin is analyzed using the zero moment point theory according to the maximum allowable speed, a wheel-ground force model is established in combination with the attitude stability margin and ground features, the distance constraints of adjacent path points and the speed continuity are verified, and an execution feasibility assessment result is generated; Based on the optimal cleaning strategy, the task completion is evaluated, the coverage of the cleaner in the rasterized cleaning area is calculated according to the path scheme to be evaluated, the number of effectively covered grids is counted, the multiple covered areas and their time distribution are detected, the uncovered areas are identified using connected domain analysis and their accessibility is evaluated, a supplementary cleaning path is generated for the uncovered areas, and the task completion evaluation result is obtained based on the supplementary cleaning path; Perform resource consumption evaluation based on the optimal cleaning strategy, calculate dynamic power consumption according to the path curvature data and the maximum allowable speed, calculate operation time based on the coverage range, evaluate obstacle avoidance time in combination with the wheel-ground force model, analyze the complexity of the path planning algorithm and the amount of real-time control calculation, and generate a resource consumption evaluation result; According to the execution feasibility assessment result, the task completion assessment result and the resource consumption assessment result, an optimal cleaning path is selected from the path schemes to be evaluated; and a secure transmission channel with end-to-end encryption function is constructed based on the selected optimal cleaning path.

[0014] According to a second aspect of the embodiments of the present invention, Provides a progressive privacy computing system for the virtual-real fusion scenario of smart sweepers, including: The first unit is used to collect indoor space data through the multimodal sensor array of the intelligent sweeper, obtain point cloud data of the micron-level laser radar, three-dimensional image data of the depth perception camera, spatial response data of the sonar sensor and reflection signal data of the millimeter wave radar, and input the point cloud data, the three-dimensional image data, the spatial response data and the reflection signal data into the entangled state registration network to generate a six-dimensional scene feature tensor in a unified coordinate system; input the six-dimensional scene feature tensor into the graph neural network for semantic segmentation and time series modeling, and construct a scene knowledge graph containing the spatial distribution of scene objects, object attribute information and the spatiotemporal interaction relationship between objects; The second unit is used to input the scene knowledge graph into a deep reinforcement learning network, and construct a privacy sensitivity scoring matrix according to the object attribute information in the scene knowledge graph; based on the privacy sensitivity scoring matrix, the nodes in the scene knowledge graph are hierarchically encrypted using a zero-knowledge proof mechanism to generate a multi-level encrypted graph; the multi-level encrypted graph is submitted to a blockchain network node group, and a plurality of scene data versions with different privacy protection strengths are calculated and output through a federated learning framework, each of which contains a corresponding privacy protection strength parameter and a scene information integrity parameter; The third unit is used to build a path planning optimization model in a secure multi-party computing environment according to the scene data version and its corresponding parameters; input the path planning optimization model into the neural symbolic reasoning engine to generate an initial cleaning path solution set; use the graph attention network to perform distributed optimization evaluation on the initial cleaning path solution set, and output an optimal cleaning strategy with a credibility quantification indicator; select the optimal cleaning path based on the optimal cleaning strategy, and send the optimal cleaning path to the smart sweeper for execution through an encrypted channel; collect the real-time status data generated by the smart sweeper during the execution process, and transmit the real-time status data back to the scene knowledge graph for dynamic updating to achieve continuous optimization of the scene model.

[0015] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0016] A fourth aspect of the embodiments of the present invention is: A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0017] The beneficial effects of this application are as follows: 1. Improve scene understanding capabilities: By integrating multimodal sensor data and graph neural networks, a scene knowledge graph containing the spatial distribution of objects, attribute information, and temporal and spatial interaction relationships is constructed, enabling the sweeper to understand the indoor environment more comprehensively and improve cleaning efficiency and intelligence.

[0018] 2. Ensure user privacy and security: Use zero-knowledge proof and multi-level encryption technology to process scene data, and combine federated learning and secure multi-party computing to achieve scene data version output with different privacy protection strengths, effectively protecting user privacy while sharing and utilizing data.

[0019] 3. Optimize cleaning path planning: Based on technologies such as deep reinforcement learning, neural symbolic reasoning, and graph attention network, a path planning optimization model is constructed, and the scene model is dynamically updated in combination with real-time status data to achieve more intelligent, efficient, and reliable cleaning path planning. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a flowchart of a progressive privacy computing method for a virtual-reality fusion scenario of a smart sweeper according to an embodiment of the present invention; Figure 2 This is a structural diagram of the progressive privacy computing system for the virtual-reality fusion scenario of the smart sweeper according to an embodiment of the present invention. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0023] Figure 1 : is a flowchart of a progressive privacy computing method for a virtual-real fusion scenario of an intelligent sweeping machine according to an embodiment of the present invention, such as Figure 1 As shown, the method includes: S11. Collect indoor space data through the multimodal sensor array of the intelligent sweeper, obtain point cloud data of micron-level laser radar, three-dimensional image data of depth perception camera, spatial response data of sonar sensor and reflection signal data of millimeter wave radar, and input the point cloud data, the three-dimensional image data, the spatial response data and the reflection signal data into the entangled state registration network to generate a six-dimensional scene feature tensor in a unified coordinate system; input the six-dimensional scene feature tensor into the graph neural network for semantic segmentation and time series modeling, and construct a scene knowledge graph containing the spatial distribution of scene objects, object attribute information and the spatiotemporal interaction relationship between objects; S12. Input the scene knowledge graph into the deep reinforcement learning network, and construct a privacy sensitivity scoring matrix based on the object attribute information in the scene knowledge graph; based on the privacy sensitivity scoring matrix, use the zero-knowledge proof mechanism to perform hierarchical encryption processing on the nodes in the scene knowledge graph to generate a multi-level encrypted graph; submit the multi-level encrypted graph to the blockchain network node group, and calculate and output multiple scene data versions with different privacy protection strengths through the federated learning framework, each of which contains corresponding privacy protection strength parameters and scene information integrity parameters; S13. According to the scenario data version and its corresponding parameters, a path planning optimization model is constructed in a secure multi-party computing environment; the path planning optimization model is input into a neural symbolic reasoning engine to generate an initial cleaning path solution set; a graph attention network is used to perform distributed optimization evaluation on the initial cleaning path solution set, and an optimal cleaning strategy with a credibility quantification indicator is output; based on the optimal cleaning strategy, an optimal cleaning path is selected, and the optimal cleaning path is sent to the smart sweeper for execution through an encrypted channel; real-time status data generated by the smart sweeper during execution is collected, and the real-time status data is transmitted back to the scenario knowledge graph for dynamic updating, so as to achieve continuous optimization of the scenario model.

[0024] In an optional implementation, the scene knowledge graph is input into a deep reinforcement learning network, and a privacy sensitivity scoring matrix is ​​constructed according to the object attribute information in the scene knowledge graph; based on the privacy sensitivity scoring matrix, the nodes in the scene knowledge graph are hierarchically encrypted using a zero-knowledge proof mechanism to generate a multi-level encrypted graph, including: The scene knowledge graph is input into a deep reinforcement learning network, wherein the deep reinforcement learning network comprises a state space module, an action space module and a value assessment module; the state space module receives spatial position information, object usage frequency information and interaction relationship information between objects in the scene knowledge graph; the action space module generates a privacy sensitivity score interval based on the output information of the state space module; the value assessment module calculates the privacy sensitivity score value according to the spatiotemporal correlation of object attributes and user privacy annotation information; The privacy sensitivity score value output by the deep reinforcement learning network is input into a scoring matrix construction module, and the scoring matrix construction module uses an adaptive weight fusion algorithm to normalize the privacy sensitivity score value to generate a privacy sensitivity scoring matrix that characterizes the degree of privacy association between object nodes in the scene knowledge graph; based on the privacy sensitivity scoring matrix, the object nodes are divided into sensitive protection levels higher than a preset sensitivity threshold and sensitive protection levels lower than the preset sensitivity threshold using a spectral clustering algorithm; A zero-knowledge proof mechanism is called to perform hierarchical encryption processing on object nodes with different sensitive protection levels, and the zero-knowledge proof mechanism includes a proof generation unit and a verification execution unit; the proof generation unit constructs an encryption proof using the Groth16 protocol, and adopts a fully homomorphic encryption scheme for object nodes with a sensitive protection level higher than a preset sensitivity threshold, and adopts a lightweight searchable encryption scheme for object nodes with a sensitive protection level lower than the preset sensitivity threshold; the verification execution unit verifies the legitimacy of the encrypted data through a pairing operation, outputs the encryption proof verification result, and reconstructs the encrypted object nodes that have passed the verification into a multi-level encryption graph.

[0025] First, build a scene knowledge graph. The scene knowledge graph is a graph structure consisting of nodes and edges. Nodes represent objects in the scene, and edges represent the relationship between objects. Each node contains attribute information of the object, such as name, location, frequency of use, etc. For example, in a knowledge graph of a smart home scene, nodes can be "smart speakers", "cameras", "TVs", etc., and edges can be "control", "connection", "located in the same room", etc.

[0026] Next, the scene knowledge graph is input into the deep reinforcement learning network. The network includes a state space module, an action space module, and a value assessment module. The state space module receives the spatial position information, usage frequency information, and interaction relationship information of objects in the scene knowledge graph. For example, the location, usage frequency, and connection relationship of the smart speaker with other devices. The action space module generates a privacy sensitivity score interval based on the output information of the state space module, for example, between 0 and 1, divided into three intervals of low, medium, and high. The value assessment module calculates the privacy sensitivity score value based on the spatiotemporal correlation of the object attributes and the user's privacy annotation information. For example, if the camera is labeled as "highly sensitive" by the user, its privacy sensitivity score value will be higher. Assuming that the smart speaker is used frequently and has a connection relationship with the camera, its privacy sensitivity score value will also increase accordingly. Assuming that the final calculated score value is 0.8.

[0027] Then, a privacy sensitivity scoring matrix is ​​constructed. The privacy sensitivity scoring values ​​output by the deep reinforcement learning network are input into the scoring matrix construction module. This module uses an adaptive weight fusion algorithm to normalize the privacy sensitivity scoring values ​​and generate a privacy sensitivity scoring matrix that represents the degree of privacy association between object nodes in the scene knowledge graph. For example, if the privacy sensitivity scores of the smart speaker and the camera are 0.8 and 0.9 respectively, the corresponding element values ​​in the matrix will be higher, indicating that there is a strong privacy association between them.

[0028] Subsequently, the object nodes are classified. Based on the privacy sensitivity scoring matrix, the spectral clustering algorithm is used to divide the object nodes into sensitive protection levels higher than the preset sensitivity threshold and sensitive protection levels lower than the preset sensitivity threshold. Assuming that the preset sensitivity threshold is 0.7, both the smart speaker and the camera belong to the sensitive protection level higher than the preset sensitivity threshold.

[0029] Finally, the nodes are encrypted in different levels. The zero-knowledge proof mechanism is called for hierarchical encryption of object nodes with different sensitivity protection levels. The zero-knowledge proof mechanism includes a proof generation unit and a verification execution unit. The proof generation unit constructs an encryption proof using the Groth16 protocol, and adopts a fully homomorphic encryption scheme for object nodes with a sensitivity protection level higher than the preset sensitivity threshold, and a lightweight searchable encryption scheme for object nodes with a sensitivity protection level lower than the preset sensitivity threshold. For example, smart speakers and cameras are encrypted using a fully homomorphic encryption scheme, while other nodes with lower sensitivity are encrypted using a lightweight searchable encryption scheme. The verification execution unit verifies the legitimacy of the encrypted data through a pairing operation, outputs the encryption proof verification result, and reconstructs the verified encrypted object nodes into a multi-level encryption graph.

[0030] The solution of this application can: Enhanced privacy protection capabilities: Through deep reinforcement learning and zero-knowledge proof technology, fine-grained hierarchical encryption of sensitive information in the scene knowledge graph is achieved, effectively preventing privacy leakage and ensuring data security. Improved data availability: The hierarchical encryption strategy allows access and use of low-sensitivity data while protecting sensitive information, ensuring data availability and avoiding data access restrictions caused by "one-size-fits-all" encryption methods. Improved system efficiency: The application of lightweight searchable encryption schemes reduces the computational overhead of encryption and decryption, improves the overall efficiency of the system, and makes the scheme more practical.

[0031] In an optional implementation, the multi-level encrypted graph is submitted to the blockchain network node group, and multiple scene data versions with different privacy protection strengths are calculated and output through the federated learning framework, each of which contains corresponding privacy protection strength parameters and scene information integrity parameters including: The multi-level encrypted graph is divided into data shards according to the spatial position relationship, and the data shards are distributed to multiple nodes in the blockchain network node group using a consistent hashing algorithm, wherein the data shards include node encryption attribute data, edge relationship encryption data, and shard index information; a practical Byzantine fault-tolerant consensus algorithm is used in the blockchain network node group to perform data consistency verification, calculation correctness verification, and version synchronization verification on the data shards; A federated learning framework is constructed based on the blockchain network node group, and the federated learning framework includes a feature extraction network, a feature fusion network and a decision network; the feature extraction network extracts features from encrypted data local to the node, the feature fusion network integrates feature information of multiple nodes, and the decision network generates scene data versions with different privacy protection strengths; an adaptive federated optimization algorithm is used in the federated learning framework for model training, and the adaptive federated optimization algorithm performs local model updates through nodes, reduces communication overhead through gradient compression, uses homomorphic encryption to aggregate model parameters, and introduces a differential privacy protection gradient update process; The output results of the federated learning framework are input into the scene data generation module, which generates privacy protection strength parameters according to the degree of data desensitization, encryption strength and noise level, and generates scene information integrity parameters according to topology preservation, attribute retention rate and relationship integrity.

[0032] First, prepare multi-level encrypted graph data. The graph contains nodes and edges. Nodes have attribute information, and edges represent the relationship between nodes. All sensitive data is encrypted. For example, node attributes can be encrypted using attributes, and edge relationships can be encrypted using homomorphic encryption. Graph data is sharded according to spatial location, and each shard contains node encrypted attribute data, edge relationship encrypted data, and shard index information. For example, a city traffic map can be divided into multiple shards according to administrative districts. Each shard contains information such as roads and intersections within the administrative district, as well as corresponding encrypted attributes and relationship data.

[0033] Then, the data shards are distributed to the blockchain network node group. The consistent hashing algorithm is used to assign each data shard to the appropriate node to ensure that the data is evenly distributed and easy to find. For example, the SHA256 algorithm is used to calculate the hash value of each shard, and then the shard is mapped to different blockchain nodes based on the hash value. The blockchain network uses a practical Byzantine fault-tolerant consensus algorithm to verify the data consistency, calculation correctness, and version synchronization of the data shards to ensure that the data of all nodes remains consistent and safe and reliable. For example, when a node receives a new data shard, other nodes will jointly verify the integrity and correctness of the shard to prevent malicious nodes from tampering with the data.

[0034] Next, a federated learning framework based on a blockchain network node group is constructed. The framework includes a feature extraction network, a feature fusion network, and a decision network. The feature extraction network extracts features from encrypted data locally at each node. For example, a convolutional neural network is used to extract road traffic flow features. The feature fusion network integrates feature information extracted by multiple nodes. For example, a secure multi-party computing protocol is used to weighted average the feature vectors of each node without leaking the original data. The decision network generates scene data versions with different privacy protection strengths based on the fused feature information. For example, by adjusting the output layer parameters of the decision network, the degree of desensitization of the generated data can be controlled. In the federated learning framework, an adaptive federated optimization algorithm is used for model training. The algorithm performs local model updates through nodes and uses gradient compression to reduce communication overhead. At the same time, homomorphic encryption is used to aggregate model parameters, and a differential privacy mechanism is introduced to protect the gradient update process, further enhancing privacy protection capabilities. For example, after each node trains the model locally, it only uploads the encrypted gradient information to the central server, and the server aggregates the encrypted gradient and distributes it to each node.

[0035] Finally, the output results of the federated learning framework are input into the scenario data generation module. This module generates privacy protection strength parameters based on the degree of data desensitization, encryption strength, and noise level, and generates scenario information integrity parameters based on topology preservation, attribute retention rate, and relationship integrity. For example, if the degree of data desensitization is high, the privacy protection strength parameter is also correspondingly high; if the topology structure remains intact, the topology preservation parameter is high. Finally, multiple scenario data versions with different privacy protection strengths are output, each version contains the corresponding privacy protection strength parameters and scenario information integrity parameters. For example, a highly desensitized data version can be generated, which is suitable for public release; at the same time, a low-desensitized data version can be generated, which is suitable for internal research.

[0036] The solution of this application can: Enhanced data privacy protection: Through the combination of multi-level encryption, blockchain technology, federated learning and differential privacy, all-round protection of sensitive data is achieved, effectively preventing data leakage and abuse. Improved data availability: While protecting data privacy, the solution can generate multiple scene data versions with different privacy protection strengths to meet data usage requirements in different scenarios, improving data availability and value. Ensure data integrity and consistency: Using the distributed ledger and consensus mechanism of blockchain, the integrity and consistency of data are ensured, preventing data from being tampered with and forged, and improving the credibility of data.

[0037] In an optional implementation, a path planning optimization model is constructed in a secure multi-party computing environment according to the scenario data version and its corresponding parameters; the path planning optimization model is input into a neural symbolic reasoning engine to generate an initial cleaning path solution set; a distributed optimization evaluation is performed on the initial cleaning path solution set using a graph attention network, and the output of the optimal cleaning strategy with a credibility quantification indicator includes: According to the scenario data version and its corresponding parameters, a path planning optimization model is constructed in a secure multi-party computing environment; the path planning optimization model adopts a three-party computing protocol based on secret sharing to decompose the computing task into a spatial constraint computing task, a temporal constraint computing task and a result aggregation computing task; the path planning optimization model outputs an initial path planning solution that integrates physical constraints, environmental constraints and privacy constraints; The path planning initial solution is input into the neural symbolic reasoning engine for deep reasoning operation. The encoding layer of the neural symbolic reasoning engine converts the scene spatial relationship, action temporal relationship and constraint rule relationship in the path planning initial solution into a vector representation; the reasoning layer of the neural symbolic reasoning engine extracts implicit rules from historical data based on the vector representation and merges them with expert rules to generate a path planning rule set; the decoding layer of the neural symbolic reasoning engine restores the reasoning result to a symbolic path description based on the path planning rule set to generate an initial cleaning path solution set containing multiple candidate paths; Constructing a graph structure representation for the candidate paths in the initial cleaning path solution set, wherein nodes in the graph structure representation correspond to path points and edges correspond to path segments; inputting the graph structure representation into a graph attention network, and calculating node-level attention weights, edge-level attention weights, and global attention weights through distributed computing nodes; prioritizing the candidate paths based on the node-level attention weights, edge-level attention weights, and global attention weights, and generating a path solution sequence with evaluation scores; A credibility quantization index is calculated for each path in the path solution sequence, and the credibility quantization index is input into the graph attention network as evaluation feedback, the node-level attention weight, the edge-level attention weight and the global attention weight are updated, the path solution sequence is optimized and cleaned, and an optimal cleaning strategy is output.

[0038] First, prepare the scene data. The scene data includes static data and dynamic data. Static data includes room layout, furniture location, etc. Taking version v1.0 as an example, the parameters include room size (length 10 meters, width 8 meters), obstacle location (coordinate (2,3), size 1x1 meter, coordinate (5,6), size 2x2 meters), etc. Dynamic data includes real-time personnel location, pet location, etc. Taking version v2.0 as an example, the parameters include personnel location (coordinate (3,4)), pet location (coordinate (7,2)), etc. These data will be used to build the path planning model.

[0039] Then, a path planning optimization model is constructed in a secure multi-party computing environment. A three-party computing protocol based on secret sharing is used to decompose the computing task into a spatial constraint computing task, a temporal constraint computing task, and a result aggregation computing task. Taking three computing nodes A, B, and C as an example, the room size information is secretly shared with the three nodes. Node A holds the value of the first half of the room length, which is 5, and node B holds the value of the second half of the room length, which is 5. Node C holds the key for decryption. The three nodes calculate the relationship between their respective data and the obstacle position, and secretly share and aggregate the intermediate results. Finally, the initial solution of the path planning is obtained, which integrates physical constraints (such as room boundaries and obstacle positions), environmental constraints (such as the need to avoid personnel positions), and privacy constraints (such as personnel position information is not leaked). Assume that the initial solution is a series of discrete coordinate points [(1,1), (2,2), (3,3), (4,4)].

[0040] Next, the initial solution of the path planning is input into the neural symbolic reasoning engine. The encoding layer of the neural symbolic reasoning engine converts the scene spatial relationship (for example, the coordinate points (1,1) and (2,2) are adjacent), the action temporal relationship (for example, the order from (1,1) to (2,2) and then to (3,3)) and the constraint rule relationship (for example, avoiding obstacles) in the initial solution of the path planning into vector representations. For example, the coordinate point (1,1) is represented as the vector [0.1, 0.1], (2,2) is represented as the vector [0.2, 0.2], the adjacent relationship is represented as the vector [0.9], and the obstacle avoidance rule is represented as the vector [0.8]. Based on these vector representations, the reasoning layer of the neural symbolic reasoning engine extracts implicit rules from the historical cleaning data (for example, the corners of the room are usually cleaned first), and merges them with expert rules (for example, the high-frequency activity areas are cleaned first) to generate a path planning rule set. For example, the rule set can be expressed as "clean the corners first, then the high-frequency areas, and finally the other areas". The decoding layer of the neural symbolic reasoning engine restores the reasoning results to a symbolic path description based on the path planning rule set, such as "from (1,1) to (4,1), and then from (4,1) to (4,4)", and generates an initial sweeping path plan set containing multiple candidate paths, such as {Path 1: [(1,1), (2,1), (3,1), (4,1), (4,2), (4,3), (4,4)], Path 2: [(1,1), (1,2), (1,3), (1,4), (2,4),(3,4), (4,4)]}.

[0041] Subsequently, a graph structure representation is constructed for the candidate paths in the initial set of sweeping path solutions. The nodes in the graph structure representation correspond to path points, and the edges correspond to path segments. The graph structure representation is input into the graph attention network, and the node-level attention weights, edge-level attention weights, and global attention weights are calculated through distributed computing nodes. For example, the node-level attention weight of path point (1,1) is 0.2, the edge-level attention weight of edge (1,1)-(2,1) is 0.3, and the global attention weight of path 1 is 0.5. Based on these weights, the candidate paths are prioritized and a sequence of path solutions with evaluation scores is generated. For example, the evaluation score of path 1 is 0.8, and the evaluation score of path 2 is 0.7, then the path solution sequence is {path 1, path 2}.

[0042] Finally, the credibility quantification index is calculated for each path in the path solution sequence. For example, the credibility of path 1 is 0.9, and the credibility of path 2 is 0.8. The credibility quantification index is used as evaluation feedback to input the graph attention network, update the node-level attention weight, edge-level attention weight, and global attention weight, optimize the path solution sequence and output the optimal sweeping strategy, for example, select path 1 as the optimal sweeping strategy.

[0043] The solution of this application can: Privacy protection: Path planning is performed in a secure multi-party computing environment to protect privacy information in scene data, such as personnel locations. Intelligent optimization: Using the neural symbolic reasoning engine and graph attention network, the cleaning path can be intelligently generated and optimized according to scene data and constraints to improve cleaning efficiency. Credibility quantification: By calculating the credibility quantification index, the reliability of the path plan can be evaluated, and the optimal cleaning strategy can be selected to improve the success rate of cleaning.

[0044] In an optional implementation, the graph structure representation is input into a graph attention network, and node-level attention weights, edge-level attention weights, and global attention weights are calculated through distributed computing nodes; candidate paths are prioritized based on the node-level attention weights, the edge-level attention weights, and the global attention weights, and a path solution sequence with evaluation scores is generated, including: Acquire path features, wherein the path features include geometric features of path segment length, path segment curvature, and path segment slope, as well as performance features of energy consumption index and time index; construct the geometric features and the performance features into a node attribute vector, and construct the corner features and distance features between the path segments into an edge attribute vector; The graph structure representation is input into a graph attention network, and the node attribute vector and the edge attribute vector are processed in parallel by distributed computing nodes; the distributed computing nodes construct a query vector and a key vector based on the node attribute vector, and use a self-attention mechanism to calculate the correlation between the query vector and the key vector to obtain a node attention weight; The distributed computing node concatenates the node attribute vector and the edge attribute vector and inputs them into a trainable parameter matrix, calculates the constraint relationship between adjacent path segments to obtain edge-level attention weights; the distributed computing node constructs a global context vector, concatenates the node attribute vector and the global context vector and inputs them into a multi-layer perceptron, calculates the correlation between the overall features and local features of the path to obtain a global attention weight; The distributed computing node performs a weighted combination of the node attention weight, the edge-level attention weight and the global attention weight based on preset weights to obtain a comprehensive path score; the candidate paths are prioritized according to the comprehensive path score to generate a path plan sequence with evaluation scores.

[0045] First, the features of each candidate path are obtained, including the path segment length, path segment curvature, path segment slope that describe the path geometry, and energy consumption index and time index that reflect the path performance.

[0046] Next, the extracted path features are constructed as the input of the graph structure representation. Each path segment is regarded as a node in the graph, and the geometric features and performance features of the path segment constitute the attribute vector of the node. The angle features and distance features between the path segments constitute the attribute vector of the edge connecting these nodes. For example, the angle between path segment 1 and path segment 2 is 30 degrees and the distance is 50 meters, then the edge attribute vector between them is (30, 50).

[0047] Then, the constructed graph structure representation is input into the graph attention network for processing. The distributed computing nodes in the network process the node attribute vector and the edge attribute vector in parallel. Each distributed computing node first constructs a query vector and a key vector based on the node attribute vector, and obtains the node attention weight by calculating the correlation between the query vector and the key vector, which reflects the importance of each path segment itself.

[0048] At the same time, each distributed computing node concatenates the node attribute vector and the edge attribute vector, and then inputs them into a trainable parameter matrix to calculate the constraint relationship between adjacent path segments and obtain the edge-level attention weight to reflect the mutual influence between path segments.

[0049] In addition, each distributed computing node also constructs a global context vector, concatenates the node attribute vector and the global context vector and inputs them into the multi-layer perceptron to calculate the correlation between the overall and local features of the path and obtain the global attention weight, which reflects the impact of each path segment on the overall path.

[0050] After that, each distributed computing node will weight the node attention weight, edge attention weight, and global attention weight according to the preset weights to obtain the comprehensive score of each path. For example, if the preset weights are 0.4, 0.3, 0.3, then the comprehensive score of a path is: 0.4 node attention weight + 0.3 edge attention weight + 0.3 global attention weight.

[0051] Finally, the candidate paths are prioritized according to the calculated comprehensive scores, and a path solution sequence with evaluation scores is generated. For example, the comprehensive scores of the three candidate paths are 85, 92, and 78, respectively, and the generated path solution sequence is: path 2 (92 points), path 1 (85 points), and path 3 (78 points).

[0052] The solution of this application can: By comprehensively considering the geometric characteristics of the path, the performance characteristics, and the relationship between the path segments, the pros and cons of the path can be more comprehensively evaluated, thereby selecting a more appropriate path. The use of distributed computing nodes to process information in parallel improves the efficiency of path planning, which is especially suitable for processing large-scale road networks and complex scenarios. By using the learning ability of the graph attention network, the model parameters can be adjusted according to different application scenarios and requirements, thereby achieving more accurate path planning.

[0053] In an optional implementation, an optimal cleaning path is selected based on the optimal cleaning strategy, and the optimal cleaning path is sent to the smart sweeper through an encrypted channel for execution; real-time status data generated by the smart sweeper during execution is collected, and the real-time status data is sent back to the scene knowledge graph for dynamic update, so as to realize continuous optimization of the scene model, including: Based on the optimal cleaning strategy, the path plan is evaluated for execution feasibility, task completion and resource consumption, and the optimal cleaning path is selected from the path plan according to the evaluation results; a secure transmission channel with end-to-end encryption function is constructed, and the secure transmission channel includes a hybrid encryption module, an identity authentication module and a dynamic key update module; the optimal cleaning path is sent to the intelligent sweeper through the secure transmission channel, wherein the sending process adopts segment confirmation and packet loss retransmission mechanism to ensure transmission reliability; Collect real-time status data from the intelligent sweeper, pre-process the real-time status data, perform noise filtering through Kalman filtering, achieve multi-source data fusion through time series alignment, perform anomaly detection through statistical analysis, and generate pre-processed status data; transmit the pre-processed status data back to the scene knowledge graph; The scene knowledge graph is dynamically updated through temporal knowledge fusion and spatial knowledge fusion. The temporal knowledge fusion is processed based on historical data weights, new data credibility and conflicting data reconciliation, and the spatial knowledge fusion is processed based on local update, global consistency and topology optimization. Based on the fusion results, the motion model parameters, environmental model parameters and task model parameters are optimized to achieve continuous optimization of the scene model.

[0054] First, build a scene knowledge graph. This graph contains static data such as the room's geometric structure, furniture location, obstacle information, and dynamic data such as dust distribution and ground material. For example, a room can be represented as a two-dimensional plane, with furniture and obstacles represented by polygons, dust distribution represented by grayscale values, and ground materials represented by different labels. The initial graph can be constructed through manual user input or sensor scanning.

[0055] Next, several candidate cleaning path solutions are generated based on the optimal cleaning strategy. The optimal cleaning strategy can be set according to user needs, for example, giving priority to cleaning areas with more dust, or giving priority to cleaning areas with high frequency of activity. Taking "bow-shaped" cleaning as an example, multiple different "bow-shaped" paths can be generated based on the geometric shape of the room and the layout of furniture. Assuming that the room is a rectangle of 5 meters and 4 meters, two candidate paths can be generated: one from left to right and one from top to bottom.

[0056] Then, the candidate path solutions are evaluated. The evaluation indicators include execution feasibility, task completion and resource consumption. Execution feasibility refers to whether the path will collide with obstacles; task completion refers to the proportion of the cleaning area covered by the path; resource consumption refers to the estimated time and power required to complete the path. For example, assuming that the path from left to right will collide with the table, its execution feasibility is low; the path from top to bottom can cover 90% of the area, is expected to take 20 minutes, and consume 20% of the power, so its task completion is high and resource consumption is moderate.

[0057] Filter the optimal cleaning path based on the evaluation results. Select the path with the highest comprehensive score of the evaluation indicators as the optimal path. For example, if the path from top to bottom has the highest comprehensive score in all indicators, then select this path as the optimal cleaning path.

[0058] A secure communication channel is established to send the optimal cleaning path to the smart sweeper. The channel adopts a hybrid encryption mechanism, combining symmetric encryption and asymmetric encryption to ensure the security of data transmission. First, the sweeper and the control center exchange symmetric keys through an asymmetric encryption algorithm; then, the path data is encrypted using the symmetric key for transmission. At the same time, the channel also includes an identity authentication module and a dynamic key update module to further enhance security. For example, the sweeper and the control center can pre-share a public key and private key pair for identity authentication and key exchange; at regular intervals, the symmetric key is dynamically updated to prevent key leakage. During the transmission process, segmented confirmation and packet loss retransmission mechanisms are used to ensure that the path data is transmitted to the sweeper completely and reliably. For example, the path data is divided into multiple data packets. After each data packet is sent, the sweeper needs to return confirmation information; if the control center does not receive the confirmation information, the data packet is resent.

[0059] The smart sweeper starts to perform the cleaning task and collects status data in real time, including location coordinates, power, sensor data, etc. For example, the sweeper uploads the current location coordinates, remaining power, and dust sensor readings once a second.

[0060] Preprocess the collected status data. First, use the Kalman filter algorithm to filter the data for noise, for example, remove the jitter noise in the position coordinates. Then, use the timing alignment technology to fuse the data from different sensors, for example, associate the position coordinates and dust sensor readings to the same time point. Finally, use statistical analysis methods to detect anomalies, for example, determine whether a sudden drop in power is an abnormal situation.

[0061] The pre-processed status data is sent back to the scene knowledge graph for dynamic updates. For example, the area cleaned by the sweeper is marked as cleaned, the dust distribution information is updated, and the location information of obstacles is recorded.

[0062] The update of the scene knowledge graph includes temporal knowledge fusion and spatial knowledge fusion. Temporal knowledge fusion considers the weight of historical data, the credibility of new data, and the reconciliation of conflicting data. For example, if the new dust distribution information conflicts with historical data, a trade-off is made based on the reliability of the data source. Spatial knowledge fusion considers local updates, global consistency, and topological optimization. For example, after updating the dust distribution information of a certain area, it is necessary to check whether it is consistent with the information of the surrounding areas, and perform topological optimization on the map to maintain the connectivity and integrity of the map.

[0063] Finally, based on the fused scene knowledge graph, the parameters of the scene model are optimized. For example, the weight parameters of the cleaning strategy are adjusted according to the updated dust distribution information; the parameters of the path planning algorithm are optimized according to the location information of the obstacles.

[0064] The solution of this application can: Improve cleaning efficiency: Through dynamic path planning and scene model self-learning, the sweeper can cover the cleaning area more effectively, reduce repeated cleaning and missed cleaning, and thus improve cleaning efficiency. Enhance environmental adaptability: By continuously updating the scene knowledge graph, the sweeper can better adapt to environmental changes, such as changes in the position of furniture, the emergence of new obstacles, etc., so as to maintain the stability of the cleaning effect. Improve the level of intelligence: Through automated path planning and scene model self-learning, the sweeper can complete the cleaning task autonomously without human intervention, which improves the level of intelligence and enhances the user experience.

[0065] In an optional implementation, based on the optimal cleaning strategy, the path plan is evaluated for execution feasibility, task completion, and resource consumption, and the optimal cleaning path is selected from the path plan according to the evaluation results; building a secure transmission channel with end-to-end encryption function includes: Based on the optimal cleaning strategy, the cleaning area is rasterized and divided, and a path plan to be evaluated is generated; based on the optimal cleaning strategy, an execution feasibility assessment is performed, the path plan to be evaluated is fitted with a Bezier curve to obtain path curvature data, a kinematic model is established based on the path curvature data to calculate the maximum allowable speed of each path segment, the attitude stability margin is analyzed using the zero moment point theory according to the maximum allowable speed, a wheel-ground force model is established in combination with the attitude stability margin and ground features, the distance constraints of adjacent path points and the speed continuity are verified, and an execution feasibility assessment result is generated; Based on the optimal cleaning strategy, the task completion is evaluated, the coverage of the cleaner in the rasterized cleaning area is calculated according to the path scheme to be evaluated, the number of effectively covered grids is counted, the multiple covered areas and their time distribution are detected, the uncovered areas are identified using connected domain analysis and their accessibility is evaluated, a supplementary cleaning path is generated for the uncovered areas, and the task completion evaluation result is obtained based on the supplementary cleaning path; Perform resource consumption evaluation based on the optimal cleaning strategy, calculate dynamic power consumption according to the path curvature data and the maximum allowable speed, calculate operation time based on the coverage range, evaluate obstacle avoidance time in combination with the wheel-ground force model, analyze the complexity of the path planning algorithm and the amount of real-time control calculation, and generate a resource consumption evaluation result; According to the execution feasibility assessment result, the task completion assessment result and the resource consumption assessment result, an optimal cleaning path is selected from the path schemes to be evaluated; and a secure transmission channel with end-to-end encryption function is constructed based on the selected optimal cleaning path.

[0066] First, determine the optimal cleaning strategy and perform regional gridding and path plan generation. According to the characteristics of the cleaning environment (such as obstacle distribution, regional shape, etc.) and the requirements of the cleaning task (such as coverage requirements, time limits, etc.), select appropriate cleaning strategies, such as bow-shaped cleaning, spiral cleaning, etc. Then, rasterize the cleaning area and discretize the continuous cleaning area into several grids of equal size. Based on the selected cleaning strategy and the results of rasterization, generate multiple path plans to be evaluated, each of which consists of a series of path points connecting the centers of adjacent grids. For example, in a 10m x 10m area, using a 0.5m x 0.5m grid for division, multiple bow-shaped or spiral path plans can be generated.

[0067] Next, the generated path plan is evaluated for feasibility. Bezier curve fitting is performed on each path plan to be evaluated to obtain a smooth path curve and path curvature data. Based on the path curvature data, a kinematic model is established to calculate the maximum allowable speed of each path segment. For example, at a curve with a large curvature, the maximum allowable speed will be reduced to ensure the stability of the cleaning robot. Then, based on the maximum allowable speed, the zero moment point theory is used to analyze the attitude stability margin. For example, by analyzing the center of gravity offset and tilt angle of the cleaning robot at different speeds, it is determined whether it will roll over. Combining the attitude stability margin with ground characteristics (such as friction coefficient, ground slope, etc.), a wheel-ground force model is established to verify the distance constraints and speed continuity of adjacent path points. For example, ensure that the distance between adjacent path points does not exceed the range of motion of the cleaning robot and that the speed change is not too drastic. Finally, the feasibility evaluation results are generated, such as feasible / infeasible, or a scoring system.

[0068] Then, the task completion of the path plan is evaluated. According to each path plan to be evaluated, the coverage of the sweeper in the rasterized cleaning area is calculated, and the number of effectively covered grids is counted. For example, the number of grids completely covered by the sweeper and the number of grids partially covered are counted. Detect multiple coverage areas and their time distribution, for example, identify which areas are repeatedly cleaned and the number and time of repeated cleaning. Use connected domain analysis to identify uncovered areas and evaluate their accessibility, for example, determine whether the uncovered areas are inaccessible due to obstacles or other reasons. For uncovered areas, generate supplementary cleaning paths. For example, if there is a reachable uncovered area, generate a new path segment to supplement the cleaning. Based on the supplementary cleaning path, update the coverage and obtain the task completion evaluation results, such as the coverage percentage. Assuming that a path plan covers 200 grids and the total number of grids is 250, the coverage rate is 80%.

[0069] Subsequently, the resource consumption of the path plan is evaluated. Dynamic power consumption is calculated based on the path curvature data and the maximum allowed speed. For example, power consumption will be higher in the case of high-speed movement or frequent turns. Operation time is calculated based on the coverage area. For example, the total time required to complete the cleaning task is calculated based on the cleaning speed and coverage area of ​​the sweeper. Obstacle avoidance time is evaluated in combination with the wheel-ground force model. For example, the additional time required for obstacle avoidance is estimated based on the number of obstacles and the obstacle avoidance strategy. The complexity of the path planning algorithm and the amount of real-time control calculations are analyzed. For example, the computational efficiency and real-time performance of different path planning algorithms are evaluated. Finally, resource consumption evaluation results are generated, such as total power consumption, total time, and amount of calculations. For example, the total power consumption of a path plan is 1000J and the total time is 30 minutes.

[0070] Finally, the optimal cleaning path is selected based on the evaluation results. The optimal cleaning path is selected from the evaluated path solutions based on the execution feasibility evaluation results, task completion evaluation results, and resource consumption evaluation results. For example, a path solution with high coverage, low power consumption, short time, and executable is selected. A secure transmission channel with end-to-end encryption function is constructed based on the selected optimal cleaning path to securely transmit the path data to the cleaning robot.

[0071] The solution of this application can: Improve cleaning efficiency: Through the optimal cleaning strategy and path planning algorithm, repeated cleaning and ineffective cleaning can be minimized, cleaning efficiency can be improved, and cleaning time can be shortened. Reduce resource consumption: By optimizing path planning, the power consumption and loss of the cleaning robot can be reduced, its service life can be extended, and energy costs can be reduced. Enhance security: Building a secure transmission channel with end-to-end encryption function can effectively protect the security of path data and prevent data leakage and tampering.

[0072] Figure 2 Schematic diagram of the structure of the progressive privacy computing system for the virtual-real fusion scenario of the smart sweeper according to an embodiment of the present invention. Figure 2 As shown, the system comprises: The first unit is used to collect indoor space data through the multimodal sensor array of the intelligent sweeper, obtain point cloud data of the micron-level laser radar, three-dimensional image data of the depth perception camera, spatial response data of the sonar sensor and reflection signal data of the millimeter wave radar, and input the point cloud data, the three-dimensional image data, the spatial response data and the reflection signal data into the entangled state registration network to generate a six-dimensional scene feature tensor in a unified coordinate system; input the six-dimensional scene feature tensor into the graph neural network for semantic segmentation and time series modeling, and construct a scene knowledge graph containing the spatial distribution of scene objects, object attribute information and the spatiotemporal interaction relationship between objects; The second unit is used to input the scene knowledge graph into a deep reinforcement learning network, and construct a privacy sensitivity scoring matrix according to the object attribute information in the scene knowledge graph; based on the privacy sensitivity scoring matrix, the nodes in the scene knowledge graph are hierarchically encrypted using a zero-knowledge proof mechanism to generate a multi-level encrypted graph; the multi-level encrypted graph is submitted to a blockchain network node group, and a plurality of scene data versions with different privacy protection strengths are calculated and output through a federated learning framework, each of which contains a corresponding privacy protection strength parameter and a scene information integrity parameter; The third unit is used to build a path planning optimization model in a secure multi-party computing environment according to the scene data version and its corresponding parameters; input the path planning optimization model into the neural symbolic reasoning engine to generate an initial cleaning path solution set; use the graph attention network to perform distributed optimization evaluation on the initial cleaning path solution set, and output an optimal cleaning strategy with a credibility quantification indicator; select the optimal cleaning path based on the optimal cleaning strategy, and send the optimal cleaning path to the smart sweeper for execution through an encrypted channel; collect the real-time status data generated by the smart sweeper during the execution process, and transmit the real-time status data back to the scene knowledge graph for dynamic updating to achieve continuous optimization of the scene model.

[0073] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0074] A fourth aspect of the embodiments of the present invention is: A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0075] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0076] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A progressive privacy computing method for the virtual-real fusion scenario of an intelligent sweeping robot, characterized in that: include: The multimodal sensor array of the intelligent sweeper is used to collect indoor space data, obtain point cloud data of micron-level laser radar, three-dimensional image data of depth perception camera, spatial response data of sonar sensor and reflection signal data of millimeter-wave radar, and input the point cloud data, the three-dimensional image data, the spatial response data and the reflection signal data into the entangled state registration network to generate a six-dimensional scene feature tensor in a unified coordinate system; the six-dimensional scene feature tensor is input into the graph neural network for semantic segmentation and time series modeling, and a scene knowledge graph containing the spatial distribution of scene objects, object attribute information and the spatiotemporal interaction relationship between objects is constructed; Inputting the scene knowledge graph into a deep reinforcement learning network, and constructing a privacy sensitivity scoring matrix according to the object attribute information in the scene knowledge graph; Based on the privacy sensitivity scoring matrix, the nodes in the scenario knowledge graph are hierarchically encrypted using a zero-knowledge proof mechanism to generate a multi-level encrypted graph; the multi-level encrypted graph is submitted to the blockchain network node group, and multiple scenario data versions with different privacy protection strengths are calculated and output through a federated learning framework, each of which contains corresponding privacy protection strength parameters and scenario information integrity parameters; According to the scenario data version and its corresponding parameters, a path planning optimization model is constructed in a secure multi-party computing environment; the path planning optimization model is input into a neural symbolic reasoning engine to generate an initial cleaning path solution set; a graph attention network is used to perform distributed optimization evaluation on the initial cleaning path solution set, and an optimal cleaning strategy with a quantified trustworthiness index is output; an optimal cleaning path is selected based on the optimal cleaning strategy, and the optimal cleaning path is sent to the intelligent sweeper through an encrypted channel for execution; The real-time status data generated by the intelligent sweeper during execution is collected, and the real-time status data is transmitted back to the scene knowledge graph for dynamic updating to achieve continuous optimization of the scene model.

2. The method according to claim 1, characterized in that Inputting the scene knowledge graph into a deep reinforcement learning network, and constructing a privacy sensitivity scoring matrix according to the object attribute information in the scene knowledge graph; Based on the privacy sensitivity scoring matrix, the nodes in the scenario knowledge graph are hierarchically encrypted using a zero-knowledge proof mechanism to generate a multi-level encrypted graph, including: The scene knowledge graph is input into a deep reinforcement learning network, wherein the deep reinforcement learning network comprises a state space module, an action space module and a value assessment module; the state space module receives spatial position information, object usage frequency information and interaction relationship information between objects in the scene knowledge graph; the action space module generates a privacy sensitivity score interval based on the output information of the state space module; the value assessment module calculates the privacy sensitivity score value according to the spatiotemporal correlation of object attributes and user privacy annotation information; The privacy sensitivity score value output by the deep reinforcement learning network is input into a scoring matrix construction module, and the scoring matrix construction module uses an adaptive weight fusion algorithm to normalize the privacy sensitivity score value to generate a privacy sensitivity scoring matrix that characterizes the degree of privacy association between object nodes in the scene knowledge graph; based on the privacy sensitivity scoring matrix, the object nodes are divided into sensitive protection levels higher than a preset sensitivity threshold and sensitive protection levels lower than the preset sensitivity threshold using a spectral clustering algorithm; A zero-knowledge proof mechanism is called to perform hierarchical encryption processing on object nodes with different sensitive protection levels, and the zero-knowledge proof mechanism includes a proof generation unit and a verification execution unit; the proof generation unit constructs an encryption proof using the Groth16 protocol, and adopts a fully homomorphic encryption scheme for object nodes with a sensitive protection level higher than a preset sensitivity threshold, and adopts a lightweight searchable encryption scheme for object nodes with a sensitive protection level lower than the preset sensitivity threshold; the verification execution unit verifies the legitimacy of the encrypted data through a pairing operation, outputs the encryption proof verification result, and reconstructs the encrypted object nodes that have passed the verification into a multi-level encryption graph.

3. The method according to claim 1, characterized in that The multi-level encrypted graph is submitted to the blockchain network node group, and multiple scene data versions with different privacy protection strengths are calculated and output through the federated learning framework. Each scene data version contains corresponding privacy protection strength parameters and scene information integrity parameters including: The multi-level encrypted graph is divided into data shards according to the spatial position relationship, and the data shards are distributed to multiple nodes in the blockchain network node group using a consistent hashing algorithm, wherein the data shards include node encryption attribute data, edge relationship encryption data, and shard index information; a practical Byzantine fault-tolerant consensus algorithm is used in the blockchain network node group to perform data consistency verification, calculation correctness verification, and version synchronization verification on the data shards; A federated learning framework is constructed based on the blockchain network node group, and the federated learning framework includes a feature extraction network, a feature fusion network and a decision network; the feature extraction network extracts features from encrypted data local to the node, the feature fusion network integrates feature information of multiple nodes, and the decision network generates scene data versions with different privacy protection strengths; an adaptive federated optimization algorithm is used in the federated learning framework for model training, and the adaptive federated optimization algorithm performs local model updates through nodes, reduces communication overhead through gradient compression, uses homomorphic encryption to aggregate model parameters, and introduces a differential privacy protection gradient update process; The output results of the federated learning framework are input into the scene data generation module, which generates privacy protection strength parameters according to the degree of data desensitization, encryption strength and noise level, and generates scene information integrity parameters according to topology preservation, attribute retention rate and relationship integrity.

4. The method according to claim 1, characterized in that: According to the scenario data version and its corresponding parameters, a path planning optimization model is constructed in a secure multi-party computing environment; the path planning optimization model is input into a neural symbolic reasoning engine to generate an initial cleaning path solution set; The graph attention network is used to perform distributed optimization evaluation on the initial cleaning path solution set, and the optimal cleaning strategy with credibility quantification indicators is output, including: According to the scene data version and its corresponding privacy protection strength parameter and scene information integrity parameter, a path planning optimization model is constructed in a secure multi-party computing environment; the path planning optimization model adopts a three-party computing protocol based on secret sharing to decompose the computing task into a spatial constraint computing task, a temporal constraint computing task and a result aggregation computing task; the path planning optimization model outputs an initial path planning solution that integrates physical constraints, environmental constraints and privacy constraints; The path planning initial solution is input into the neural symbolic reasoning engine for deep reasoning operation. The encoding layer of the neural symbolic reasoning engine converts the scene spatial relationship, action temporal relationship and constraint rule relationship in the path planning initial solution into a vector representation; the reasoning layer of the neural symbolic reasoning engine extracts implicit rules from historical data based on the vector representation and merges them with expert rules to generate a path planning rule set; the decoding layer of the neural symbolic reasoning engine restores the reasoning result to a symbolic path description based on the path planning rule set to generate an initial cleaning path solution set containing multiple candidate paths; Constructing a graph structure representation for the candidate paths in the initial cleaning path solution set, wherein nodes in the graph structure representation correspond to path points and edges correspond to path segments; inputting the graph structure representation into a graph attention network, and calculating node-level attention weights, edge-level attention weights, and global attention weights through distributed computing nodes; prioritizing the candidate paths based on the node-level attention weights, edge-level attention weights, and global attention weights, and generating a path solution sequence with evaluation scores; A credibility quantization index is calculated for each path in the path plan sequence, and the credibility quantization index is input into the graph attention network as evaluation feedback, the node-level attention weight, the edge-level attention weight and the global attention weight are updated, the cleaning path plan sequence is optimized and the optimal cleaning strategy is output.

5. The method according to claim 4, characterized in that Input the graph structure representation into the graph attention network, and calculate the node-level attention weight, edge-level attention weight and global attention weight through distributed computing nodes; Prioritizing candidate paths based on the node-level attention weight, the edge-level attention weight, and the global attention weight to generate a path solution sequence with evaluation scores includes: Acquire path features, wherein the path features include geometric features of path segment length, path segment curvature, and path segment slope, as well as performance features of energy consumption index and time index; construct the geometric features and the performance features into a node attribute vector, and construct the corner features and distance features between the path segments into an edge attribute vector; The graph structure is input into a graph attention network, and the node attribute vector and the edge attribute vector are processed in parallel by distributed computing nodes; the distributed computing nodes construct a query vector and a key vector based on the node attribute vector, and a self-attention mechanism is used to calculate the correlation between the query vector and the key vector to obtain a node attention weight; The distributed computing node concatenates the node attribute vector and the edge attribute vector and inputs them into a trainable parameter matrix, calculates the constraint relationship between adjacent path segments to obtain edge-level attention weights; the distributed computing node constructs a global context vector, concatenates the node attribute vector and the global context vector and inputs them into a multi-layer perceptron, calculates the correlation between the overall features and local features of the path to obtain a global attention weight; The distributed computing node performs a weighted combination of the node attention weight, the edge-level attention weight and the global attention weight based on preset weights to obtain a comprehensive score of the path; the candidate paths are prioritized according to the comprehensive score of the path to generate a path solution sequence with evaluation scores.

6. The method according to claim 1, characterized in that The optimal cleaning path is selected based on the optimal cleaning strategy, and the optimal cleaning path is sent to the intelligent sweeper through an encrypted channel for execution; the real-time status data generated by the intelligent sweeper during the execution process is collected, and the real-time status data is sent back to the scene knowledge graph for dynamic update, so as to realize continuous optimization of the scene model, including: Based on the optimal cleaning strategy, the path plan is evaluated for execution feasibility, task completion and resource consumption, and the optimal cleaning path is selected from the path plan according to the evaluation results; a secure transmission channel with end-to-end encryption function is constructed, and the secure transmission channel includes a hybrid encryption module, an identity authentication module and a dynamic key update module; the optimal cleaning path is sent to the intelligent sweeper through the secure transmission channel, wherein the sending process adopts segment confirmation and packet loss retransmission mechanism to ensure transmission reliability; Collect real-time status data from the intelligent sweeper, pre-process the real-time status data, perform noise filtering through Kalman filtering, achieve multi-source data fusion through time series alignment, perform anomaly detection through statistical analysis, and generate pre-processed status data; transmit the pre-processed status data back to the scene knowledge graph; The scene knowledge graph is dynamically updated based on a periodic update trigger mechanism and an event-driven update trigger mechanism; the scene model is updated through temporal knowledge fusion and spatial knowledge fusion, the temporal knowledge fusion is processed based on historical data weights, new data credibility and conflicting data reconciliation, and the spatial knowledge fusion is processed based on local updates, global consistency and topology optimization; based on the fusion results, the motion model parameters, environmental model parameters and task model parameters are optimized to achieve continuous optimization of the scene model.

7. The method according to claim 6, characterized in that Based on the optimal cleaning strategy, the path plan is evaluated for execution feasibility, task completion and resource consumption, and the optimal cleaning path is selected from the path plan according to the evaluation results; and a secure transmission channel with end-to-end encryption function is constructed, including: Based on the optimal cleaning strategy, the cleaning area is rasterized and divided, and multiple path plans to be evaluated are generated; based on the optimal cleaning strategy, an execution feasibility assessment is performed, Bezier curve fitting is performed on the path plan to be evaluated to obtain path curvature data, a kinematic model is established based on the path curvature data to calculate the maximum allowable speed of each path segment, the attitude stability margin is analyzed using the zero moment point theory according to the maximum allowable speed, a wheel-ground force model is established in combination with the attitude stability margin and ground features, the distance constraints of adjacent path points and the speed continuity are verified, and an execution feasibility assessment result is generated; Based on the optimal cleaning strategy, the task completion is evaluated, the coverage of the cleaner in the rasterized cleaning area is calculated according to the path plan to be evaluated, the number of effective coverage grids is counted, the multiple coverage areas and their time distribution are detected, the uncovered areas are identified by connected domain analysis and their accessibility is evaluated, and a supplementary cleaning path is generated for the uncovered areas to obtain the task completion evaluation result; Perform resource consumption evaluation based on the optimal cleaning strategy, calculate dynamic power consumption according to the path curvature data and the maximum allowable speed, calculate operation time based on the coverage range, evaluate obstacle avoidance time in combination with the wheel-ground force model, analyze the complexity of the path planning algorithm and the amount of real-time control calculation, and generate a resource consumption evaluation result; According to the execution feasibility assessment result, the task completion assessment result and the resource consumption assessment result, an optimal cleaning path is selected from the path schemes to be evaluated; and a secure transmission channel with end-to-end encryption function is constructed based on the optimal cleaning path.

8. A progressive privacy computing system for a virtual-reality fusion scene of an intelligent sweeping machine, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to collect indoor space data through the multimodal sensor array of the intelligent sweeper, obtain point cloud data of the micron-level laser radar, three-dimensional image data of the depth perception camera, spatial response data of the sonar sensor and reflection signal data of the millimeter wave radar, and input the point cloud data, the three-dimensional image data, the spatial response data and the reflection signal data into the entangled state registration network to generate a six-dimensional scene feature tensor in a unified coordinate system; input the six-dimensional scene feature tensor into the graph neural network for semantic segmentation and time series modeling, and construct a scene knowledge graph containing the spatial distribution of scene objects, object attribute information and the spatiotemporal interaction relationship between objects; The second unit is used to input the scene knowledge graph into a deep reinforcement learning network, and construct a privacy sensitivity scoring matrix according to the object attribute information in the scene knowledge graph; Based on the privacy sensitivity scoring matrix, the nodes in the scenario knowledge graph are hierarchically encrypted using a zero-knowledge proof mechanism to generate a multi-level encrypted graph; the multi-level encrypted graph is submitted to the blockchain network node group, and multiple scenario data versions with different privacy protection strengths are calculated and output through a federated learning framework, each of which contains corresponding privacy protection strength parameters and scenario information integrity parameters; The third unit is used to build a path planning optimization model in a secure multi-party computing environment according to the scenario data version and its corresponding parameters; input the path planning optimization model into a neural symbolic reasoning engine to generate an initial cleaning path solution set; use a graph attention network to perform distributed optimization evaluation on the initial cleaning path solution set, and output an optimal cleaning strategy with a trustworthy quantitative indicator; select an optimal cleaning path based on the optimal cleaning strategy, and send the optimal cleaning path to the intelligent sweeper through an encrypted channel for execution; The real-time status data generated by the intelligent sweeper during execution is collected, and the real-time status data is transmitted back to the scene knowledge graph for dynamic updating to achieve continuous optimization of the scene model.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

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

Images