Vehicle-mounted device interaction anomaly detection method based on tbox and scene perception

Abstract

The application provides a TBOX and scene perception based vehicle-mounted equipment interaction anomaly detection method, relates to the technical field of scene perception, and comprises the following steps: acquiring driving scene features of vehicle-mounted equipment, and constructing a scene constraint model; mapping driving state, environmental condition and safety rules into a semantic constraint space through semantic ontology reasoning, defining an interaction behavior legality interval based on fuzzy boundaries, and performing semantic analysis on operation behavior data to obtain preliminary semantic labels; dynamically associating the preliminary semantic labels with historical interaction sequences, calculating multi-dimensional similarity and fusing context information; adopting a knowledge driven hierarchical reasoning strategy to make a joint decision, evaluating intention rationality and scene adaptability of the operation behavior, and generating a semantic reasoning result with confidence; and controlling the vehicle-mounted equipment response according to the reasoning result, and adaptively adjusting the semantic constraint space by using feedback data. The application can accurately identify vehicle-mounted equipment interaction abnormal behaviors and improve vehicle use safety.

Description

TECHNICAL FIELD

[0001] The present application relates to scene perception technology, and in particular to a vehicle-mounted device interaction anomaly detection method based on TBOX and scene perception. BACKGROUND

[0002] With the rapid development of intelligent connected vehicle technology, the interaction between vehicle-mounted devices and users is increasingly diversified and complex. Modern vehicle-mounted systems not only need to handle various operation instructions of the driver, but also need to make real-time judgments on the rationality and safety of these operation behaviors in different driving scenarios. The TBOX of the vehicle-mounted device, as the core component of vehicle intelligent connection, carries multiple functions such as vehicle state monitoring, remote control, data communication, etc. In the actual driving process, the vehicle-mounted device needs to intelligently identify and respond to the user's operation behavior according to the current driving scene, vehicle state, environmental conditions and other multi-dimensional information. However, due to the dynamic changes of driving scenes, the individual differences of user operation habits, and potential malicious attacks, the vehicle-mounted device faces serious challenges in interaction anomaly detection. Traditional rule-based detection methods are difficult to adapt to complex and variable driving scenarios, while purely relying on machine learning methods has problems such as lack of explainability and dependence on abnormal samples. Therefore, how to combine the advantages of knowledge-driven and data-driven to build a vehicle-mounted device interaction anomaly detection method that can perceive scenes, understand semantics and adaptively optimize has become an important technical problem to be solved in the current intelligent connected vehicle field.

[0003] Existing vehicle-mounted device interaction anomaly detection methods generally lack deep perception ability of driving scenes, and only make judgments based on static rules or isolated operation behaviors, which cannot effectively combine real-time driving state, environmental conditions and safety level requirements of the vehicle for comprehensive analysis. This detection method lacking scene context perception leads to high false positive rate and false negative rate in complex dynamic scenes, making it difficult to accurately distinguish between normal operation and abnormal behavior, especially in the boundary fuzzy interaction scene.

[0004] Existing technologies lack deep reasoning ability at the semantic level in the anomaly detection process, mainly relying on surface data feature matching or simple threshold judgment, and cannot understand the real intention behind the user's operation behavior and the rationality of the behavior in the current scene. This shallow detection mechanism cannot handle operation behaviors with normal data features but abnormal intentions, and cannot effectively deal with disguised attacks by imitating normal behavior patterns, resulting in insufficient accuracy and robustness of detection.

[0005] Existing anomaly detection systems mostly use fixed detection models, lacking the ability to adaptively optimize according to actual operation feedback and scene changes. These systems are difficult to adapt to the evolution of user operation habits, the emergence of new attack patterns, and the diversification of driving scenarios after deployment, resulting in a decline in detection performance over time. At the same time, there is a lack of effective feedback mechanism to utilize misjudgment cases and environmental change information to continuously optimize the detection model, limiting the long-term usability and adaptability of the system. SUMMARY

[0006] The embodiment of the application provides a vehicle-mounted device interaction anomaly detection method based on TBOX and scene perception, which can solve the problems in the prior art.

[0007] In a first aspect, the embodiment of the application provides a vehicle-mounted device interaction anomaly detection method based on TBOX and scene perception, comprising:

[0008] Obtaining driving scene features of the vehicle-mounted device at the current time, constructing a scene constraint model according to the driving scene features; mapping the driving state, environmental conditions and hierarchical safety rules of the vehicle into a semantic constraint space based on semantic ontology reasoning, and defining an interaction behavior legality interval based on fuzzy boundaries in the semantic constraint space, performing semantic analysis and mapping on the operation behavior data to obtain preliminary semantic labels;

[0009] Dynamically associating the preliminary semantic labels with historical interaction sequences, calculating a multi-dimensional similarity measure between the current operation behavior and the historical interaction sequences within the interaction behavior legality interval, and fusing context information to obtain an association analysis result;

[0010] Designing a knowledge-driven hierarchical reasoning strategy, jointly deciding the association analysis result and the semantic constraint space, evaluating the intention rationality and scene adaptability of the operation behavior through the hierarchical reasoning strategy, and generating a semantic reasoning result with confidence based on the interaction behavior legality interval;

[0011] Controlling the response execution mode of the vehicle-mounted device according to the semantic reasoning result and its confidence, collecting user feedback information and environmental change information after execution, and adaptively adjusting the semantic constraint space using feedback data to realize dynamic optimization of the model.

[0012] Constructing a scene constraint model according to the driving scene features; mapping the driving state, environmental conditions and hierarchical safety rules of the vehicle into a semantic constraint space based on semantic ontology reasoning comprises:

[0013] According to the driving scene feature, the driving state information and the environmental condition information of the vehicle are extracted, a scene semantic ontology is constructed based on a semantic ontology reasoning framework, and a semantic association model is formed by establishing a semantic association between the driving state nodes and the environmental condition nodes in the scene semantic ontology through a weighted reasoning rule.

[0014] Based on the semantic association model, the risk characteristics of the current driving scene are analyzed, a set of applicable safety rules is determined, the safety rule set is graded according to the risk association degree in the semantic association model, and a graded safety rule with a risk level is generated. Meanwhile, the association and dependence between the graded safety rules are modeled by using the semantic association model, and a rule association network is constructed.

[0015] The driving state information, the environmental condition information and the graded safety rule are cross-ontology semantic reasoned by the semantic ontology reasoning framework, and the driving state, the environmental condition and the graded safety rule of the vehicle are mapped into a semantic constraint space based on the semantic association model and the rule association network.

[0016] According to the risk association degree in the semantic association model, the safety rule set is graded, and a graded safety rule with a risk level is generated. Meanwhile, the association and dependence between the graded safety rules are modeled by using the semantic association model, and a rule association network is constructed, including:

[0017] The risk association degree between the driving scene elements is extracted from the semantic association model, each safety rule in the safety rule set is risk evaluated according to the risk association degree, the risk level identifier corresponding to each safety rule is determined according to the risk association degree, the safety rule set is graded according to the risk level identifier, and a graded safety rule with a risk level is generated.

[0018] Based on the trigger condition and the constraint range of each safety rule in the graded safety rule, the pre-dependence relationship and the mutual exclusion conflict relationship between the graded safety rules are identified, and a rule association and dependence description is formed.

[0019] The rule association and dependence description is semantically parsed by using the semantic association model, the association and dependence strength between the graded safety rules is quantified by using the semantic reasoning mechanism in the semantic association model, the graded safety rules are taken as network nodes, and the association and dependence strength is taken as the connection relationship between the nodes, so as to construct a rule association network.

[0020] In the semantic constraint space, a fuzzy boundary-based interactive behavior legality interval is defined, and the operation behavior data is semantically parsed and mapped to obtain a preliminary semantic label, including:

[0021] obtain constraint boundary parameters corresponding to different risk levels from the semantic constraint space, construct a fuzzy boundary function in the semantic constraint space based on the constraint boundary parameters, define a transition area of a constraint boundary through the fuzzy boundary function, and define an interactive behavior legality interval based on the fuzzy boundary in the semantic constraint space.

[0022] Obtain operation behavior data, perform semantic analysis on operation type features and operation parameter features in the operation behavior data, map the operation type features and the operation parameter features to semantic dimension coordinates in the semantic constraint space, and obtain a preliminary semantic label according to the position of the semantic dimension coordinates in the interactive behavior legality interval.

[0023] Dynamically associate the preliminary semantic label with a historical interaction sequence, calculate a multi-dimensional similarity measure between the current operation behavior and the historical interaction sequence in the interactive behavior legality interval, and fuse context information to obtain an association analysis result including:

[0024] Map the preliminary semantic label and the historical semantic label in the historical interaction sequence for mapping analysis, construct an association structure between the semantic labels based on a dynamic association network, determine a hierarchical relationship between the labels in the dynamic association network, and form a label association model;

[0025] Extract semantic dimension features of the current operation behavior from the interactive behavior legality interval according to the label association model, and construct a semantic feature vector; model the semantic dimension features of the historical operation behavior in the historical interaction sequence based on the label association model, and generate a historical feature vector set;

[0026] In the interactive behavior legality interval, calculate a spatial distance measure using the semantic feature vector and the historical feature vector set, calculate a semantic association measure in combination with the hierarchical relationship in the dynamic association network; determine a weight distribution strategy according to the label association model, and perform weighted fusion on the spatial distance measure and the semantic association measure to obtain a multi-dimensional similarity measure between the current operation behavior and the historical interaction sequence;

[0027] Obtain context information, construct a calibration model using the dynamic association network and the multi-dimensional similarity measure, and dynamically calibrate the multi-dimensional similarity measure; perform feature fusion on the calibrated multi-dimensional similarity measure and the context information to generate an association analysis result.

[0028] Design a knowledge-driven hierarchical reasoning strategy, jointly decide the association analysis result and the semantic constraint space, evaluate the intention rationality and scenario adaptability of the operation behavior through the hierarchical reasoning strategy, and generate a semantic reasoning result with confidence based on the interactive behavior legality interval including:

[0029] constructing a knowledge base and designing a layered reasoning strategy based on the knowledge base, performing semantic decomposition and conversion of the correlation analysis result to an intention reasoning layer of the layered reasoning strategy, constructing a reasoning constraint network according to a comprehensive correlation degree in the correlation analysis result and driving intention knowledge, analyzing intention semantic features of a current operation behavior through the reasoning constraint network, and forming an intention rationality result;

[0030] performing dimensional decomposition on the semantic constraint space based on the reasoning constraint network, constructing a scene constraint mode corresponding to a current driving scene, performing semantic reconstruction and conversion of the intention rationality result and the scene constraint mode to a scene reasoning layer, establishing a scene adaptation relationship by using the layered reasoning strategy, and forming a scene adaptability result;

[0031] constructing a hierarchical progressive relationship between the intention rationality result and the scene adaptability result in the reasoning constraint network, determining a feature combination mode based on the correlation analysis result and the semantic constraint space, and generating a comprehensive reasoning result;

[0032] calculating a confidence parameter by using the reasoning constraint network according to a distribution feature of the comprehensive reasoning result in the interactive behavior legality interval, and generating a semantic reasoning result with confidence.

[0033] performing dimensional decomposition on the semantic constraint space based on the reasoning constraint network, constructing a scene constraint mode corresponding to a current driving scene, includes:

[0034] performing hierarchical deconstruction on constraint elements in the semantic constraint space according to driving scene semantics, decomposing each constraint element into corresponding semantic dimensional features, and establishing an association dependency between constraint elements in the reasoning constraint network;

[0035] performing hierarchical combination on the semantic dimensional features based on the association dependency, constructing a multi-dimensional constraint expression reflecting a feature of a current driving scene, and determining an action boundary of each dimensional constraint; generating a scene constraint mode according to the multi-dimensional constraint expression and the action boundary, and establishing a hierarchical correspondence relationship between the scene constraint mode and the reasoning constraint network.

[0036] A second aspect of the embodiment of the application provides an electronic device, including:

[0037] a processor;

[0038] a memory for storing processor-executable instructions;

[0039] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0040] In a third aspect, the present application provides a computer readable storage medium having stored thereon computer program instructions which, when executed by a processor, implement the method described above.

[0041] The present application has the following advantages:

[0042] The present application can accurately identify abnormal interaction behaviors of the vehicle-mounted device, effectively reduce the misjudgment rate, improve the accuracy and reliability of the abnormal detection, and enhance the safety protection capability of the vehicle-mounted system, by constructing a scene constraint model and a semantic constraint space, deeply semantically mapping the vehicle driving state, the environmental condition and the safety rules, and performing accurate semantic analysis and multi-dimensional similarity calculation on the operation behavior within the legal interaction behavior interval.

[0043] The present application can realize intelligent processing of the whole process from detection to response, improve the understanding and processing capability of the system for complex interaction scenes, and improve the user experience, by adopting a knowledge-driven hierarchical reasoning strategy, jointly deciding the correlation analysis result and the semantic constraint space, generating a reasoning result with confidence by evaluating the intention rationality and scene adaptability of the operation behavior, and dynamically controlling the response execution mode of the vehicle-mounted device according to the reasoning result and the confidence.

[0044] The present application can realize continuous learning and dynamic optimization of the model by collecting user feedback information and environmental change information after execution, and adaptively adjusting the semantic constraint space using the feedback data, so that the abnormal detection model can be continuously improved as the use scene changes, the adaptability and robustness of the system are enhanced, the effective use period of the model is prolonged, and the later maintenance cost is reduced. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 FIG. 1 is a flowchart of a vehicle-mounted device interaction abnormal detection method based on TBOX and scene perception according to an embodiment of the present application;

[0046] Figure 2 FIG. 4 is a flowchart of multi-dimensional similarity calculation and analysis of the semantic features of the operation behavior according to an embodiment of the present application. DETAILED DESCRIPTION

[0047] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described in detail below with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0048] The technical solutions of the present application are described in detail below with specific examples. The following specific examples can be combined with each other, and the same or similar concepts or processes may not be described in detail in some examples.

[0049] Figure 1 The flowchart of the method for detecting abnormal interaction of the vehicle-mounted device based on TBOX and scene perception according to an embodiment of the present application is shown in FIG. 1, which includes the following steps: Figure 1

[0050] Obtaining driving scene features of the vehicle-mounted device at the current time, constructing a scene constraint model according to the driving scene features, mapping the driving state and environmental conditions of the vehicle and the hierarchical safety rules into a semantic constraint space based on semantic ontology reasoning, defining a fuzzy boundary-based interaction behavior legality interval in the semantic constraint space, and performing semantic analysis and mapping on the operation behavior data to obtain a preliminary semantic label;

[0051] Dynamically associating the preliminary semantic label with a historical interaction sequence, calculating a multi-dimensional similarity measure between the current operation behavior and the historical interaction sequence within the interaction behavior legality interval, and fusing context information to obtain an association analysis result;

[0052] Designing a knowledge-driven hierarchical reasoning strategy, jointly deciding the association analysis result and the semantic constraint space, evaluating the intention rationality and scene adaptability of the operation behavior through the hierarchical reasoning strategy, and generating a semantic reasoning result with confidence based on the interaction behavior legality interval;

[0053] Controlling the response execution mode of the vehicle-mounted device according to the semantic reasoning result and its confidence, collecting user feedback information and environmental change information after execution, and using the feedback data to adaptively adjust the semantic constraint space to realize dynamic optimization of the model.

[0054] In an optional implementation, constructing a scene constraint model according to the driving scene features, and mapping the driving state and environmental conditions of the vehicle and the hierarchical safety rules into a semantic constraint space based on semantic ontology reasoning include:

[0055] Extracting driving state information and environmental condition information of the vehicle according to the driving scene features, constructing a scene semantic ontology based on a semantic ontology reasoning framework, establishing semantic association between driving state nodes and environmental condition nodes in the scene semantic ontology through weighted reasoning rules to form a semantic association model;

[0056] ​Based on the semantic association model, the risk characteristics of the current driving scene are analyzed, a set of applicable safety rules is determined, the safety rule set is graded according to the risk association degree in the semantic association model, and a graded safety rule with a risk level is generated; meanwhile, the semantic association model is used to model the association and dependence between the graded safety rules, and a rule association network is constructed;

[0057] The driving state information, the environmental condition information and the graded safety rules are cross-ontology semantic reasoning through the semantic ontology reasoning framework, a dynamic constraint boundary is constructed based on the semantic association model and the rule association network, and the driving state, the environmental condition and the graded safety rules of the vehicle are mapped into a semantic constraint space.

[0058] The driving scene feature extraction module collects multi-dimensional driving data through the vehicle-mounted sensor network, including real-time speed information obtained by a speed sensor, steering operation data collected by a steering wheel angle sensor, brake intensity monitored by a brake pedal pressure sensor, acceleration intention recorded by a throttle pedal position sensor, and other driving state information. The environmental condition information is obtained by a GPS navigation system to obtain road type and location data, a radar sensor to detect the distance and relative speed of the front obstacle, a camera to identify traffic sign and lane line information, a rain sensor to detect rainfall intensity, and a light sensor to determine day and night state. The sensor data sampling frequency is uniformly set to 50Hz to ensure data timing synchronization, the vehicle speed accuracy requirement is ±0.5km / h, the distance measurement accuracy is ±0.2m, and the angle measurement accuracy is ±1 degree. The original data is processed by median filtering and Gaussian smoothing, and the filtering window length is set to 5 sampling points, and the outliers are detected and removed by the quartile method.

[0059] The driving state information extraction process fuses multi-sensor data into a structured state description, defines the vehicle motion state as six basic categories of static, uniform speed, acceleration, deceleration, left turn and right turn, and each state category corresponds to a specific sensor data combination mode. The uniform speed state determination condition is that the vehicle speed change rate is less than 2km / h per second and the steering wheel angle is less than 5 degrees, the acceleration state determination condition is that the throttle pedal opening is greater than 30% and the vehicle speed change rate is greater than 5km / h per second, and the brake state determination condition is that the brake pedal pressure is greater than 20% and the vehicle speed change rate is less than negative 3km / h per second. The environmental condition information extraction covers three dimensions of weather condition, road type and traffic condition, the weather condition is divided into four categories of sunny day, cloudy day, rainy day and foggy day, the road type is divided into four categories of expressway, urban main road, secondary road and rural road, and the traffic condition is divided into three categories of smooth, slow and congestion.

[0060] The scene semantic ontology is constructed using the Web Ontology Language (OWL) as the knowledge representation framework, and the concept hierarchy and attribute relationship in the driving field are defined. The top-level concepts include three abstract classes: driving entities, environmental entities, and safety entities. The driving entities include two sub-classes: vehicle states and driving behaviors. The environmental entities include three sub-classes: weather conditions, road conditions, and traffic conditions. The safety entities include two sub-classes: safety rules and risk levels. Each concept class is defined by data attributes to quantify the characteristics, such as the vehicle state class containing speed attributes, position attributes, and direction attributes with value ranges of 0-200 km / h, latitude and longitude coordinates, and 0-360 degrees, respectively. Object attributes define the semantic relationships between concepts, such as the adaptation relationship between vehicle states and road conditions, and the compliance relationship between driving behaviors and safety rules.

[0061] In the process of establishing weighted reasoning rules, the semantic association between driving state nodes and environmental condition nodes is modeled through conditional dependency relationships. The rules are represented in the IF-THEN form, with the condition part describing the combination pattern of driving states and environmental conditions, and the conclusion part giving the corresponding safety constraints or risk assessment. A typical rule is "IF vehicle speed is greater than 80 km / h AND road type is urban road THEN risk level is high." The rule weight is determined based on historical accident data and expert experience, with a value range of 0.1-1.0. The weight calculation uses a combination of support and confidence. Support represents the frequency of the rule appearing in the data set, and confidence represents the probability of the conclusion being true when the rule conditions are true. The rule base contains 150 reasoning rules, each associated with 3-8 premise conditions. The weight distribution is normally distributed with a mean of 0.65 and a standard deviation of 0.15.

[0062] In the process of constructing the semantic association model, driving state nodes and environmental condition nodes are represented as a graph structure, with nodes representing specific state or condition instances and edges representing the semantic association relationship between nodes. The graph construction algorithm uses a similarity-based adjacency matrix generation method, with node similarity calculated by cosine distance of feature vectors. The similarity threshold is set to 0.6, and edges are established between node pairs with a similarity above the threshold. Edge weights represent the association strength, which is quantified by mutual information with a value range of 0-1. Edges with a weight greater than 0.5 are marked as strong associations. The graph structure is represented sparsely to improve storage and computation efficiency, with an average node degree controlled between 8-12. Graph connectivity is verified by a depth-first search algorithm.

[0063] The risk feature analysis calculates the comprehensive risk score of the current driving scene based on a semantic correlation model, and adopts a multi-factor weighted evaluation method. The risk factors include five dimensions of speed risk, distance risk, weather risk, road risk, and behavior risk, and the weights are set as 0.25, 0.2, 0.2, 0.2, and 0.15 respectively. The speed risk is calculated by the ratio of the current speed to the road speed limit, and the higher the overspeed degree, the higher the risk score. The distance risk is calculated by the ratio of the distance to the front vehicle and the safe distance, and the risk score increases when the distance is insufficient. The weather risk is determined according to the visibility and road conditions, and the higher the risk score corresponds to the worse weather. The comprehensive risk score is calculated by the weighted average method, and the score range is 0-100, less than 30 is low risk, 30-70 is medium risk, and higher than 70 is high risk.

[0064] The applicable safety rule set determination process filters the relevant rules from the rule library according to the risk characteristics of the current scene. The rule filtering adopts a fuzzy matching algorithm to calculate the matching degree of the scene feature vector and the rule condition vector, and the matching degree threshold is set to 0.7, and the rules higher than the threshold are included in the applicable set. The rule importance sorting is based on the product of the rule weight and the matching degree, and the top 20 rules constitute the final applicable rule set. The rule conflict detection is realized by condition overlap analysis, and the rules with an overlap degree greater than 0.8 are marked as potential conflicts, and the rule priority and timestamp are used for conflict resolution.

[0065] The hierarchical safety rule generation adopts a hierarchical classification method based on the risk correlation degree, and divides the applicable rule set into three levels of emergency, important, and general. The emergency level rule corresponds to the dangerous scene, including emergency braking, anti-collision warning, etc., the trigger condition is strict, and the response time requirement is within 200 milliseconds; the important level rule corresponds to the potential risk scene, including vehicle distance keeping, speed adjustment, etc., the trigger condition is moderate, and the response time requirement is within 1 second; the general level rule corresponds to the daily safety reminder, including turning prompt, light control, etc., the trigger condition is loose, and the response time requirement is within 3 seconds. The hierarchical process adopts the K-means clustering algorithm, taking the rule weight and the matching degree as the clustering features, the K-means++ method is used for clustering center initialization, and the iteration number is limited to 100 times.

[0066] The rule association network constructs the dependency relationship between hierarchical safety rules, using a directed graph to represent the trigger order and constraint relationship between rules. The nodes in the graph represent a single safety rule, and the edges represent the dependency relationship between rules, with edge weights representing the dependency strength. Dependency relationship identification is based on semantic analysis of rule conditions and conclusions, using natural language processing techniques to extract causal relationships and temporal relationships between rules. The network construction algorithm uses a greedy strategy to connect rule pairs with high dependency strength, and the network density is controlled between 0.3-0.5 to ensure that the network can reflect rule association while avoiding excessive complexity. The cycle detection algorithm identifies and eliminates cyclic dependencies between rules to ensure the directed acyclic nature of the network.

[0067] Cross-ontology semantic reasoning uses a forward reasoning algorithm, taking driving state information and environmental condition information as initial facts, and generating reasoning conclusions through the matching and triggering of hierarchical safety rules. The reasoning process uses a depth-first search strategy, with a reasoning depth limit of 6 layers to avoid reasoning explosion and infinite loops. Rule matching uses a pattern matching algorithm that supports variable binding and condition constraint checking, with a matching success rate requirement of greater than 0.8. The reasoning results include the triggered safety rule set, the corresponding constraint conditions, and the confidence score, which is calculated by the geometric mean of rule weight and matching degree.

[0068] Dynamic constraint boundary construction is based on the output of the semantic association model and the rule association network, defining the feasible region of safe driving in a multi-dimensional feature space. The constraint boundary uses a set of hyperplanes to represent each safety constraint, with the plane normal vector and intercept determined by linear programming. The boundary dynamic adjustment mechanism updates the constraint parameters in real time according to the scene changes, with an adjustment frequency of every 500 milliseconds and a parameter change amplitude limited to within 15% of the previous value. The constraint space dimension is set to 12 dimensions, including speed, acceleration, direction angle, vehicle distance, braking force, and other key driving parameters.

[0069] Semantic constraint space mapping unifies the driving state, environmental conditions, and hierarchical safety rules into a constraint set, using a combination of linear transformation and nonlinear mapping. The mapping function is implemented through a multilayer perceptron, with the network containing an input layer, 2 hidden layers, and an output layer, with hidden layer neuron numbers of 64 and 32, respectively. The activation function uses ReLU, and the output layer uses tanh activation function to ensure the boundedness of the output range. The training data contains 50,000 labeled samples, trained using the mini-batch stochastic gradient descent algorithm with a batch size of 32, a learning rate of 0.001, and 500 training rounds.

[0070] In an optional implementation, the set of safety rules is ranked according to the risk association degree in the semantic association model, and a ranked safety rule with a risk level is generated; meanwhile, the association dependency between the ranked safety rules is modeled by using the semantic association model, and a rule association network is constructed, including:

[0071] The risk association degree between driving scene elements is extracted from the semantic association model, each safety rule in the set of safety rules is risk evaluated according to the risk association degree, a risk level identifier corresponding to each safety rule is determined according to the risk association degree, the set of safety rules is ranked according to the risk level identifier, and a ranked safety rule with a risk level is generated;

[0072] Based on the trigger condition and constraint range of each safety rule in the ranked safety rule, a pre-dependence relationship and a mutual exclusion conflict relationship existing between the ranked safety rules are identified, and a rule association dependency description is formed;

[0073] The semantic association model is used for semantic analysis of the rule association dependency description, the association dependency strength between the ranked safety rules is quantified by a semantic reasoning mechanism in the semantic association model, the ranked safety rules are taken as network nodes, and the association dependency strength is taken as a connection relationship between nodes, and a rule association network is constructed.

[0074] In actual operation, the risk association degree between driving scene elements is extracted from the semantic association model, and the extraction process depends on the semantic knowledge graph structure established in the model. The system accesses the knowledge storage layer of the semantic association model, traverses the driving scene element nodes stored therein, and these element nodes include but are not limited to vehicle speed, road type, weather condition, obstacle distance, traffic signal state and other key elements. For the association edge between each pair of element nodes, the system reads the risk association value marked on the association edge, and the value is learned from a large amount of real driving data in the model training stage. For example, the risk association value of the association edge between the vehicle speed node and the obstacle distance node is 0.85, indicating that there is a high risk association between the two elements.

[0075] When performing risk assessment on each safety rule in the safety rule set, all driving scene elements involved in the safety rule are parsed. Take the safety rule "When the distance to the front obstacle is less than 10 meters and the vehicle speed is greater than 40 km / h, emergency braking must be performed immediately" as an example, the rule involves two core elements, obstacle distance and vehicle speed. The system queries the risk association degree between the two elements from the semantic association model and obtains a value of 0.85. At the same time, the system also queries the risk association degree between each of the two elements and other related elements, including the association degree between vehicle speed and road friction coefficient 0.72, and the association degree between obstacle distance and driver reaction time 0.68. The system comprehensively calculates these risk association values in a weighted cumulative manner, with a weight of 0.6 for the association between main elements and a weight of 0.4 for the association between secondary elements, and finally obtains a comprehensive risk assessment value of 0.79 for the safety rule.

[0076] According to the comprehensive risk assessment value, determine the risk level identifier corresponding to each safety rule, preset the risk level threshold interval, and identify the assessment value of 0.75 to 1.0 as a high risk level, the assessment value of 0.5 to 0.75 as a medium risk level, and the assessment value of 0 to 0.5 as a low risk level. The assessment value 0.79 of the aforementioned emergency braking rule falls into the high risk level interval, and the system adds the identifier "RISK_LEVEL_HIGH" to the rule. Another rule "When the vehicle is in the parking lot and the speed is less than 10 km / h, low-speed driving is allowed" goes through the same evaluation process, which involves scene elements including vehicle location and vehicle speed, and the risk association degree between these elements is low, the comprehensive assessment value is 0.32, and the system adds the identifier "RISK_LEVEL_LOW" to it.

[0077] According to the risk level identifier, the safety rule set is classified, and rules with the same risk level identifier are classified into the same classification group. After classification, a classified safety rule structure including a high-risk classification rule group, a medium-risk classification rule group and a low-risk classification rule group is generated. The high-risk classification rule group contains 15 rules, covering key safety scenes such as emergency braking, collision warning and out-of-control protection; the medium-risk classification rule group contains 32 rules, covering regular driving scenes such as lane keeping, speed adjustment and following distance control; and the low-risk classification rule group contains 28 rules, covering auxiliary scenes such as parking assistance, low-speed cruise and comfort adjustment.

[0078] Based on the trigger condition and constraint range of each safety rule in the hierarchical safety rule, the associated dependent relationship between the rules is identified. The system parses the structured description of each rule, extracts the trigger condition field and the constraint range field. The trigger condition of the high-risk rule "When the distance of the detected front obstacle is less than 10 meters and the vehicle speed is greater than 40 kilometers per hour, the emergency braking must be performed immediately" is that the distance of the detected front obstacle is less than 10 meters and the vehicle speed is greater than 40 kilometers per hour, and the constraint range is to perform the emergency braking operation. The trigger condition of another medium-risk rule "When the vehicle performs emergency braking, the hazard warning light must be activated synchronously" is that the vehicle performs emergency braking, and the constraint range is to activate the hazard warning light. The system identifies through the semantic matching mechanism that the trigger condition of the second rule is directly associated with the constraint range of the first rule, determines that the second rule depends on the execution result of the first rule, and establishes the pre-dependent relationship.

[0079] When identifying the mutual exclusion conflict relationship between the rules, the constraint range fields of the rules are compared. The constraint range of the high-risk rule "When the approaching side vehicle is detected, the lane changing operation is prohibited" is to prohibit lane changing, and the constraint range of the medium-risk rule "When the navigation instruction needs to change lanes, the lane changing preparation is performed" is to perform lane changing. The system detects that the constraint ranges of the two rules are contradictory in the driving behavior level, and lane changing is mutually exclusive, so it is determined that there is a mutual exclusion conflict relationship between the two rules. The system records the pre-dependent relationship and the mutual exclusion conflict relationship identified as the inter-rule associated dependent description, stores it in a structured data format, and includes the source rule identifier, the target rule identifier, the association type, and the association description field.

[0080] The semantic association model is used for semantic analysis of the inter-rule associated dependent description. The associated dependent description is input into the reasoning layer of the semantic association model. The reasoning layer of the model includes a semantic understanding module and a relationship quantification module. The semantic understanding module analyzes the natural language text in the associated dependent description and extracts key semantic features. For the pre-dependent relationship "Rule B depends on the execution result of Rule A", the semantic understanding module identifies the dependent keyword and the execution result keyword, and determines that it is a strong dependent relationship. For the mutual exclusion conflict relationship "The constraint ranges of Rule C and Rule D are mutually exclusive", the semantic understanding module identifies the exclusion keyword and determines that it is a conflict relationship.

[0081] The semantic reasoning mechanism in the semantic association model quantifies the correlation dependency strength between hierarchical safety rules, which is based on the comprehensive calculation of semantic features and rule risk levels. For the precedent dependency relationship, if both the source rule and the target rule are high-risk levels, the system determines the strength value of this dependency relationship as 0.95; if the source rule is a high-risk level and the target rule is a medium-risk level, the dependency strength value is adjusted to 0.78; if both are low-risk levels, the dependency strength value is 0.45. For the mutual exclusion conflict relationship, the system quantifies the correlation strength according to the severity of the conflict, and assigns a strength value of 0.92 to a conflict involving safety-critical operations and a strength value of 0.63 to a conflict involving general driving operations.

[0082] The hierarchical safety rules are used as network nodes to build a rule association network, with each rule corresponding to a unique node in the network. The node contains attribute information such as rule identification, risk level identification, trigger condition, and constraint range. The system uses the quantified correlation dependency strength as the connection relationship between nodes, and establishes a directed connection edge between rule nodes with precedent dependency or mutual exclusion conflict. The connection edge carries the association type attribute and the association strength value, and the connection edge direction of the precedent dependency type points to the dependent target node, and the connection edge of the mutual exclusion conflict type is a bidirectional connection. The rule association network finally forms a network topology structure containing 75 nodes and 132 connection edges, which completely describes the correlation dependency relationship and its strength distribution between all hierarchical safety rules.

[0083] In an optional implementation, an interaction behavior legality interval based on a fuzzy boundary is defined in the semantic constraint space, and operation behavior data is semantically analyzed and mapped to obtain a preliminary semantic label, including:

[0084] Constraint boundary parameters corresponding to different risk levels are obtained from the semantic constraint space, a fuzzy boundary function is constructed in the semantic constraint space based on the constraint boundary parameters, a transition region of the constraint boundary is defined by the fuzzy boundary function, and an interaction behavior legality interval based on a fuzzy boundary is defined in the semantic constraint space;

[0085] Obtain operation behavior data, semantically analyze the operation type features and operation parameter features in the operation behavior data, map the operation type features and the operation parameter features to the semantic dimension coordinates in the semantic constraint space, and obtain a preliminary semantic label according to the position of the semantic dimension coordinates in the interaction behavior legality interval.

[0086] The process of constructing the legal behavior interval in the semantic constraint space needs to start from the extraction of constraint boundary parameters. The system reads the constraint boundary parameters corresponding to different risk levels from the pre-established semantic constraint space, which includes the strict threshold of high-risk operations, the moderate threshold of medium-risk operations, and the loose threshold of low-risk operations. Taking data access behavior as an example, the constraint boundary parameters of high-risk operations are set as access frequency threshold of no more than 5 times per minute, data volume threshold of no more than 1000 records per time, and sensitive level threshold of level 3. The constraint boundary parameters of medium-risk operations are set as access frequency threshold of no more than 20 times per minute, data volume threshold of no more than 5000 records per time, and sensitive level threshold of level 2. The constraint boundary parameters of low-risk operations are set as access frequency threshold of no more than 50 times per minute, data volume threshold of no more than 10000 records per time, and sensitive level threshold of level 1.

[0087] Based on the obtained constraint boundary parameters, a fuzzy boundary function is constructed in the semantic constraint space. The core function of the fuzzy boundary function is to establish a continuous transition relationship between the constraint boundaries of different risk levels, rather than a simple discrete boundary. Taking the access frequency dimension as an example, between the high-risk threshold of 5 times and the medium-risk threshold of 20 times, the fuzzy boundary function defines a transition region, and the width of the region is set as 20% of the difference between the adjacent thresholds, i.e. 3 units wide. In this transition region, the system uses membership degree calculation to determine the risk attribution degree of the operation behavior. When the access frequency of a certain operation behavior is 7 times per minute, the value falls within the transition region of the high-risk boundary, and the system calculates the membership degree of the high-risk interval as 0.67 and the membership degree of the medium-risk interval as 0.33. This membership degree allocation is based on the linear interpolation calculation of the relative position of the value in the transition region.

[0088] The definition of the transition region fully considers the multi-dimensional characteristics of the semantic constraint space. In the 3-dimensional semantic space, the access frequency, data volume, and sensitive level dimensions each have independent transition region settings. The transition region width of the access frequency dimension is set as 20% of the threshold difference, the transition region width of the data volume dimension is set as 15% of the threshold difference, and the transition region of the sensitive level dimension is realized through a weight decay coefficient, which is set as 0.8 between adjacent levels. When the operation behavior is in the transition region in multiple dimensions at the same time, the system comprehensively evaluates the membership degrees of each dimension and calculates the overall membership degree using a weighted average method, where the weight of the access frequency dimension is 0.4, the weight of the data volume dimension is 0.4, and the weight of the sensitive level dimension is 0.2.

[0089] After the construction of the fuzzy boundary function is completed, the legality interval of the interactive behavior in the semantic constraint space is defined. The legality interval is divided into four levels: completely legal region, conditional legal region, suspected illegal region and explicit illegal region. The completely legal region corresponds to the case where all dimension parameters of the operation behavior are below the low risk threshold. The conditional legal region corresponds to the transition region where at least one dimension parameter is between the low risk threshold and the medium risk threshold. The suspected illegal region corresponds to the transition region where at least one dimension parameter is between the medium risk threshold and the high risk threshold. The explicit illegal region corresponds to the case where at least one dimension parameter exceeds the high risk threshold. To illustrate with a specific case, the access frequency of an operation behavior is 15 times per minute, the data volume is 3000 records, and the sensitive level is level 2. The access frequency falls between the low risk threshold and the medium risk threshold, the data volume also falls between the low risk threshold and the medium risk threshold, and the sensitive level is equal to the medium risk threshold. The operation behavior is defined in the conditional legal region.

[0090] After obtaining the operation behavior data, the data is structured and parsed to extract operation type features and operation parameter features. The extraction of operation type features is achieved by identifying the action identifier of the operation behavior, which includes types such as query, insert, update, delete, export, etc. Taking a database operation log as an example, the log record shows that the operation type is query operation, the target table of the query is the user information table, the query time is a specific timestamp, and the number of records returned by the query is 4200. The system extracts the operation type feature as query from the log, and extracts operation parameter features including target table name, operation time, and return record number, etc.

[0091] The semantic analysis of operation parameter features requires converting the original parameters into standardized expressions recognizable by the semantic constraint space. The parameter of return record number 4200 is converted to the value 4200 of the data volume dimension. The operation time parameter is calculated by comparing with the historical operation records to obtain the value of the access frequency dimension. Assuming that the user performed 18 similar queries in the previous minute of the operation time, the value of the access frequency dimension is 18 times per minute. The parameter of the target table being the user information table is determined by querying the data sensitivity classification table to determine that the sensitive level of the user information table is level 2, and thus the value of the sensitive level dimension is 2.

[0092] The extracted and parsed operation type features and operation parameter features are mapped to the semantic dimension coordinates in the semantic constraint space. The access frequency of 18 times per minute is mapped to the coordinate position 18 of the access frequency dimension, the data volume of 4200 is mapped to the coordinate position 4200 of the data volume dimension, and the sensitive level of 2 is mapped to the coordinate position 2 of the sensitive level dimension. These three coordinate values form a three-dimensional coordinate point of the operation behavior in the semantic constraint space, recorded as coordinate point position 18, 4200, 2.

[0093] According to the position of the semantic dimension coordinates in the interaction behavior legality interval, it is determined that the operation behavior belongs to a legality region. The value of the coordinate point in the access frequency dimension is between the low risk threshold 50 and the medium risk threshold 20, and falls within the conditional legality region range. The value of the coordinate point in the data volume dimension is between the low risk threshold 10000 and the medium risk threshold 5000, and also falls within the conditional legality region range. The value of the coordinate point in the sensitive level dimension is equal to the medium risk threshold 2, which reaches the medium risk boundary. The system determines that the operation behavior is located in the conditional legality region by comprehensively determining the position of the three dimensions, and generates a preliminary semantic label as a conditional legal operation, and an additional detailed label as a combination label of medium access frequency, medium data volume and medium sensitive level, which is used for subsequent deep analysis and risk assessment process.

[0094] In an optional implementation, the preliminary semantic label is dynamically associated with a historical interaction sequence, a multi-dimensional similarity measure between the current operation behavior and the historical interaction sequence is calculated in the interaction behavior legality interval, and context information is fused to obtain an association analysis result, including:

[0095] The preliminary semantic label is mapped and analyzed with a historical semantic label in a historical interaction sequence, an association structure between semantic labels is constructed based on a dynamic association network, a hierarchical relationship between the labels is determined in the dynamic association network, and a label association model is formed;

[0096] According to the label association model, semantic dimension features of the current operation behavior are extracted from the interaction behavior legality interval to construct a semantic feature vector; historical operation behaviors in the historical interaction sequence are modeled in semantic dimension features based on the label association model to generate a historical feature vector set;

[0097] In the interaction behavior legality interval, a spatial distance measure is calculated using the semantic feature vector and the historical feature vector set, a semantic association measure is calculated in combination with the hierarchical relationship in the dynamic association network; a multi-dimensional similarity measure between the current operation behavior and the historical interaction sequence is obtained by weighting and fusing the spatial distance measure and the semantic association measure according to the weight distribution strategy determined by the label association model;

[0098] Context information is obtained, a calibration model is constructed using the dynamic association network and the multi-dimensional similarity measure to dynamically calibrate the multi-dimensional similarity measure; and the calibrated multi-dimensional similarity measure is feature-fused with the context information to generate an association analysis result.

[0099] As shown in Figure 2 , the method comprises:

[0100] In the process of dynamically associating the preliminary semantic label with the historical interaction sequence, the preliminary semantic label and the historical semantic label stored in the historical interaction sequence are extracted, and these labels are compared one by one through the label mapping analysis module. Specifically, the system establishes a label mapping table, which records that the label code corresponding to the preliminary semantic label "document viewing" is L001, the historical semantic label "document editing" contained in the historical interaction sequence corresponds to the code L002, the label code corresponding to "document download" is L003, and the label code corresponding to "document sharing" is L004. The system identifies the semantic association between the labels through the semantic analysis module, for example, "document viewing" has a before-and-after operation association with "document editing", has a similar operation attribute with "document download", and has an extended operation association with "document sharing". On this basis, a dynamic association network is constructed, which uses the structure of nodes and edges to represent, each semantic label as a node, and the association strength between nodes is reflected by the weight value of the edge. Take a specific case as an example, the edge weight value between node L001 and node L002 is 0.85, indicating that they have a strong operation sequence association; the edge weight value between node L001 and node L003 is 0.72, indicating a medium intensity of functional similarity; and the edge weight value between node L001 and node L004 is 0.56, indicating a weak association.

[0101] The dynamic association network determines the hierarchical relationship between the labels through hierarchical analysis, and divides the labels into a basic operation layer, an extended operation layer, and a high-level operation layer according to the operation level. The basic operation layer includes basic function labels such as "document viewing", the extended operation layer includes enhanced function labels such as "document editing", and the high-level operation layer includes collaboration function labels such as "document sharing". The system assigns a level identifier to each label, the level value of the basic operation layer label is set to 1, the level value of the extended operation layer label is set to 2, and the level value of the high-level operation layer label is set to 3. By traversing all nodes and their connection relationships in the dynamic association network, a label association model containing semantic features, association strength, and level attributes is formed. The model is stored in the form of a data structure, including a label identifier field, a semantic description field, an association weight field, a level identifier field, and a timestamp field.

[0102] When the semantic dimension features of the current operation behavior are extracted from the interaction behavior legality interval based on the label association model, the legality interval analysis module is called to read the preset legality interval range. Assuming that the current operation behavior is accessing the business system during the working hours, the interaction behavior legality interval is defined as the time dimension being between 8 am and 6 pm, the operation frequency dimension being no more than 100 times per hour, and the access permission dimension being within the 2-level permission range. The system extracts the operation type feature, operation object feature, operation time feature, operation frequency feature, permission level feature, and the like from the current operation behavior, and constructs a semantic feature vector containing these features. The first dimension of the vector records the operation type as "document access", the second dimension records the operation object as "contract file", the third dimension records the operation time as 10:30 am, the fourth dimension records the operation frequency in the last 1 hour as 15 times, and the fifth dimension records the current user permission level as 2 levels.

[0103] When the semantic dimension features of the historical operation behaviors in the historical interaction sequence are modeled, all the historical operation records stored in the historical interaction sequence are traversed. The historical interaction sequence contains 500 operation records in the last 30 days, and each record contains operation timestamp, operation type, operation object, operation result, permission level, and the like. The system reads each historical record in turn, extracts the semantic features of each record according to the label association model, and generates the corresponding historical feature vector. For example, a historical record shows that the user performed a document editing operation at 2 pm three days ago, the system extracts the semantic features of the operation to generate a historical feature vector, the first dimension of the vector is "document editing", the second dimension is "report file", the third dimension is 2 pm, the fourth dimension is the operation frequency in the last 1 hour as 8 times, and the fifth dimension is 2-level permission. By extracting the features of all the historical records, a historical feature vector set containing 500 historical feature vectors is formed.

[0104] When the spatial distance metric is calculated within the interaction behavior legality interval, the vector distance calculation module is used to calculate the distance between the semantic feature vector and each vector in the historical feature vector set. Specifically, the system compares the semantic feature vector with the first historical feature vector, and calculates the cumulative value of the difference in each dimension as the spatial distance. For example, the difference value of the operation type dimension is 0.15, the difference value of the operation object dimension is 0.23, the difference value of the operation time dimension is 0.08, the difference value of the operation frequency dimension is 0.12, and the difference value of the permission level dimension is 0.00. The cumulative value of these difference values is 0.58, which is the spatial distance metric value of the historical feature vector. The system sequentially calculates the spatial distance metric values of the semantic feature vector and all 500 vectors in the historical feature vector set, and generates a spatial distance metric set.

[0105] In the calculation of the semantic correlation metric in the hierarchical relationship in the dynamic association network, the hierarchical relationship analysis module is called to read the hierarchical information in the label association model. The system identifies that the label "document viewing" corresponding to the current operation behavior is located in the basic operation layer, and the hierarchical value is 1. The hierarchical value of the label corresponding to the historical operation behavior is found. For the historical operation behavior "document editing", the hierarchical value is 2. The system calculates the absolute value of the hierarchical difference value as 1, and converts the hierarchical difference value into an adjustment coefficient of the semantic correlation metric. When the hierarchical difference value is 0, the adjustment coefficient is 1.0, when the hierarchical difference value is 1, the adjustment coefficient is 0.8, and when the hierarchical difference value is 2, the adjustment coefficient is 0.6. The system calculates the semantic correlation metric according to the edge weight value stored in the dynamic association network and the hierarchical adjustment coefficient. The semantic correlation metric value 0.68 is obtained by multiplying the edge weight value 0.85 and the adjustment coefficient 0.8. The same calculation process is performed on each vector in the historical feature vector set to generate a semantic correlation metric set.

[0106] In the determination of the weight allocation strategy according to the label association model, the weight configuration parameters stored in the label association model are read. The weight allocation strategy sets different weight proportions according to different operation scenarios. For a scenario with high security requirements, the weight of the spatial distance metric is set to 0.6, and the weight of the semantic correlation metric is set to 0.4. For a scenario with strong business process association, the weight of the spatial distance metric is set to 0.4, and the weight of the semantic correlation metric is set to 0.6. The system selects the corresponding weight configuration according to the type of the current operation behavior. Assuming that the current scenario is a document access scenario with high security requirements, the system selects the weight proportions 0.6 and 0.4. The system multiplies the spatial distance metric value 0.58 by the weight 0.6 to obtain the weighted value 0.348, and multiplies the semantic correlation metric value 0.68 by the weight 0.4 to obtain the weighted value 0.272. The two weighted values are added to obtain the multi-dimensional similarity metric value 0.62. The same weighted fusion calculation is performed on all vectors in the historical feature vector set to generate a multi-dimensional similarity metric set.

[0107] In the acquisition of context information, the context acquisition module is called to read various context data of the current operation environment. The context information includes that the type of the device currently logged in by the user is a desktop computer, the network connection state is an enterprise intranet, the geographic location is an office area, the current time period is working hours, the number of online users in the same period is 85, the system load state is normal, and the number of recent abnormal operation records is 0. The system performs structured processing on these context information to generate a context feature vector, which includes device type encoding, network type encoding, location type encoding, time period encoding, concurrent quantity value, load state identifier, and abnormal times value.

[0108] When the calibration model is constructed by using the dynamic association network and the multi-dimensional similarity metric, a calibration parameter mapping table is created, which determines the calibration coefficient according to the label association strength in the dynamic association network and the multi-dimensional similarity metric value. When the multi-dimensional similarity metric value is above 0.8 and the label association strength is above 0.7, the calibration coefficient is set to 1.05, indicating that the similarity is enhanced. When the multi-dimensional similarity metric value is between 0.5 and 0.8, the calibration coefficient is set to 1.0, keeping the original value. When the multi-dimensional similarity metric value is below 0.5 and the label association strength is below 0.4, the calibration coefficient is set to 0.9, indicating that the similarity is reduced. The system multiplies the calculated multi-dimensional similarity metric value 0.62 by the corresponding calibration coefficient 1.0 to obtain the calibrated multi-dimensional similarity metric value 0.62.

[0109] When the calibrated multi-dimensional similarity metric is feature fused with the context information, a feature fusion module is used to integrate the two types of data. The system takes the calibrated multi-dimensional similarity metric value as the core feature and the dimensions in the context feature vector as the auxiliary feature to construct a comprehensive feature matrix. The first column of the matrix is the calibrated similarity value 0.62, and the subsequent columns are context feature values such as device type encoding and network type encoding. The system performs feature weighting processing on the comprehensive feature matrix, assigning a weight of 0.7 to the similarity feature and a weight of 0.3 to the context feature. The system calculates the association analysis result, which includes a comprehensive similarity score of 0.65, a risk level identification of low risk, a recommended operation of allowing access, and a confidence value of 0.88. The association analysis result is passed to the decision execution module for subsequent access control decision.

[0110] In an alternative embodiment, a knowledge-driven hierarchical reasoning strategy is designed to jointly decide the association analysis result and the semantic constraint space. The hierarchical reasoning strategy evaluates the intention rationality and scenario adaptability of the operation behavior, and generates a semantic reasoning result with confidence based on the legality interval of the interaction behavior, including:

[0111] A knowledge base is constructed and a hierarchical reasoning strategy is designed based on the knowledge base. The association analysis result is semantically decomposed and converted to the intention reasoning layer of the hierarchical reasoning strategy. The reasoning constraint network is constructed according to the comprehensive association degree in the association analysis result and the driving intention knowledge. The intention semantic features of the current operation behavior are analyzed through the reasoning constraint network to form an intention rationality result.

[0112] The semantic constraint space is dimensionally disassembled based on the reasoning constraint network, and the scenario constraint mode corresponding to the current driving scenario is constructed. The intention rationality result and the scenario constraint mode are semantically reconstructed and converted to the scenario reasoning layer. The scenario adaptability relationship is established using the hierarchical reasoning strategy to form a scenario adaptability result.

[0113] The intention rationality result and the scene adaptability result are constructed in a hierarchical progressive relationship in the reasoning constraint network, a feature combination mode is determined based on the correlation analysis result and the semantic constraint space, and a comprehensive reasoning result is generated;

[0114] According to the distribution characteristics of the comprehensive reasoning result in the interaction behavior legality interval, a confidence parameter is calculated using the reasoning constraint network, and a semantic reasoning result with confidence is generated.

[0115] The knowledge base construction link starts from multi-dimensional driving behavior data, and maps the steering angle, brake pressure, throttle opening and other physical quantities in the vehicle operation trajectory to the driving scene information. Driving intention knowledge is organized into a hierarchical structure, including basic intention types such as acceleration overtaking, deceleration avoidance, lane change, and parking on the side. Each intention type is associated with several feature patterns, for example, the acceleration overtaking intention corresponds to the feature combination of continuous increase in throttle opening and lateral deviation in steering angle. The knowledge base stores 320 historical driving cases, each case containing operation sequence, scene parameters, and intention label. These cases are manually annotated and verified to ensure their accuracy. The layered reasoning strategy is designed as a three-layer architecture, with the bottom layer being the feature extraction layer responsible for processing raw data, the middle layer being the intention reasoning layer responsible for analyzing operation intention, and the top layer being the scene reasoning layer responsible for evaluating scene adaptability.

[0116] After the correlation analysis result is input, it is processed by semantic decomposition, in which the comprehensive correlation degree value is split into operation intensity dimension and operation coordination dimension. In a specific case, the comprehensive correlation degree of a driving operation is 0.78, of which the operation intensity contributes 0.42 and the operation coordination contributes 0.36. The semantic decomposition process maps these values to the driving intention knowledge space, with operation intensity corresponding to the intensity features of throttle and brake, and operation coordination corresponding to the timing coordination features between multiple operations. The intention reasoning layer receives the decomposed semantic features, retrieves matching intention types from the knowledge base, and in a specific case, the throttle opening increases from 30% to 65% within 2 seconds, while the steering wheel angle deflects 15 degrees to the left. The feature combination has a similarity of 0.83 with the standard pattern of acceleration lane change intention.

[0117] The reasoning constraint network is instantiated at the intent reasoning layer, which contains 32 constraint nodes, each representing a specific intent decision rule. There are dependencies between constraint nodes, for example, the acceleration intent node requires the simultaneous activation of the throttle persistence node and the vehicle speed increase node. The feature vector of the current operational behavior is sequentially matched through each constraint node, and the node activation state is recorded as a Boolean value sequence. In a practical case, 26 of the 32 nodes were activated, and all the core nodes related to acceleration and lane changing were activated, while all the nodes related to emergency braking were not activated. The intent semantic features are clustered through the activation pattern, and the activation pattern of this case is compared with the standard pattern of the acceleration and lane changing intent in the knowledge base. The similarity is calculated using the Hamming distance measurement method, and the calculation result shows that the difference number is 6, the total number is 32, and the similarity is 0.8125. The intent rationality result is directly represented by the similarity value, and when the value exceeds the preset threshold of 0.75, the intent is determined to be reasonable.

[0118] The semantic constraint space dimension decomposition process identifies three core dimensions: space, time, and traffic rules. The space dimension includes parameters such as lane width, front vehicle distance, and lateral distance. In a certain scenario, the lane width is 3.5 meters, the front vehicle distance is 28 meters, and there is a vehicle in the left lane at a distance of 12 meters. The time dimension includes parameters such as response time and duration. In this case, the response time from detecting the overtaking opportunity to starting to turn is 1.2 seconds, and the entire lane changing process lasts 4.8 seconds. The traffic rules dimension includes constraints such as speed limit requirements and lane usage specifications. The current road segment has a speed limit of 80 kilometers per hour, allows overtaking, and has no solid line separation. The scene constraint pattern is constructed based on these parameters, with spatial constraints requiring a lateral distance greater than the safety threshold of 8 meters, temporal constraints requiring a lane changing duration within the 3 to 6 second interval, and traffic rule constraints requiring a current speed lower than the speed limit and being in an overtaking allowed section.

[0119] The intent type label and feature vector contained in the intent rationality result are input into the scene reasoning layer for semantic reconstruction. The semantic reconstruction process converts abstract intent semantics into specific requirements that can be directly compared with scene parameters, accelerating the conversion of the lane change intent into specific conditions such as a demand lateral space greater than the vehicle width plus 2 meters of buffer, a demand longitudinal acceleration capability supporting speed increase within a limited time, and a demand traffic environment allowing lane changing operations. The scene reasoning layer establishes a scene adaptation relationship and compares the reconstructed intent requirements with scene constraint patterns item by item. In this case, the lateral space requirement is 4.5 meters, the actual available space is 4 meters obtained by subtracting the safety distance of 8 meters from the distance of 12 meters to the left vehicle, and there is a gap of 0.5 meters. The longitudinal acceleration requirement calculates that the time required to increase the current speed of 60 kilometers per hour to 75 kilometers per hour is 5.3 seconds, and the actual available time is estimated to be 7.2 seconds based on the distance and relative speed of the preceding vehicle, meeting the requirement. The traffic rule constraints are all met. The scene adaptability result synthesizes the three comparison results using a weighted summation method, with a spatial adaptation degree of 0.89, a temporal adaptation degree of 0.95, and a rule adaptation degree of 1.00, and the comprehensive scene adaptability is 0.913.

[0120] The hierarchical progressive relationship is realized in the reasoning constraint network through directed edges between nodes, with the intent rationality result corresponding node as the predecessor node and the scene adaptability result corresponding node as the successor node. The progressive relationship is defined as only when the output result of the predecessor node exceeds the threshold value, the successor node performs reasoning calculation. This mechanism ensures the hierarchical nature of the reasoning logic. The feature combination method is determined according to the dimension weight of the correlation analysis result and the constraint strength of the semantic constraint space, with the operation intensity dimension weight being 0.6, the operation coordination dimension weight being 0.4, the spatial constraint strength being 0.7, the temporal constraint strength being 0.5, and the rule constraint strength being 0.9. The comprehensive reasoning result is generated by fusing the intent rationality value and the scene adaptability value according to a preset proportion, with the intent rationality weight being set to 0.45 and the scene adaptability weight being set to 0.55. The comprehensive reasoning result value of this case is 0.896.

[0121] The distribution characteristics of the comprehensive reasoning result in the interaction behavior legality interval are determined by an interval mapping method. The legality interval is divided into five sub-intervals, 0 to 0.2 is a serious violation interval, 0.2 to 0.45 is a mild violation interval, 0.45 to 0.7 is a boundary interval, 0.7 to 0.9 is a compliance interval, and 0.9 to 1.0 is an excellent interval. The case comprehensive reasoning result 0.896 falls in the compliance interval, the distance from the lower limit of the interval 0.7 is 0.196, and the distance from the upper limit of the interval 0.9 is 0.004. The confidence parameter calculation considers the position stability of the result value in the corresponding interval. The case is at the top of the compliance interval, the position stability is high, and the confidence base value is set to 0.85. The proportion of activated constraint nodes in the reasoning process is 0.8125, which is used as a confidence correction factor, and the final confidence parameter is 0.873. The semantic reasoning result with confidence includes three elements: the comprehensive reasoning result value 0.896, the corresponding interval label compliance interval, and the confidence parameter 0.873. The result indicates that the current driving operation behavior performs well in the intention and scene level and the judgment result has high credibility.

[0122] In an optional implementation, the dimension disassembly of the semantic constraint space based on the reasoning constraint network includes:

[0123] The constraint elements in the semantic constraint space are hierarchically deconstructed according to the driving scene semantics, each constraint element is disassembled into corresponding semantic dimension features, and the association and dependence between the constraint elements are established in the reasoning constraint network.

[0124] Based on the association and dependence, the semantic dimension features are hierarchically combined to construct a multi-dimensional constraint expression reflecting the characteristics of the current driving scene, and the action boundary of each dimension constraint is determined. The scene constraint mode is generated according to the multi-dimensional constraint expression and the action boundary, and the hierarchical correspondence relationship between the scene constraint mode and the reasoning constraint network is established.

[0125] In the current driving scene, the constraint elements in the semantic constraint space are processed by hierarchical decomposition. Specifically, the system identifies all constraint elements contained in the semantic constraint space, including lane boundary constraints, traffic signal constraints, pedestrian dynamic constraints, vehicle spacing constraints, and speed limit constraints. The system disassembles the lane boundary constraint element into three semantic dimension characteristics: lane width dimension, lane line type dimension, and lane curvature dimension. The lane width dimension is recorded as 3.5 meters, the lane line type dimension is identified as a solid line type, and the lane curvature dimension is quantified as a 15-degree turn per 100 meters. For the traffic signal constraint element, the system disassembles the signal state dimension, the remaining time dimension, and the signal position dimension. The current signal state dimension is recorded as a red light state, the remaining time dimension is 23 seconds, and the signal position dimension is recorded as 48 meters from the vehicle in front.

[0126] The correlation dependency relationship between constraint elements is established in the reasoning constraint network. There is a direct correlation dependency between lane boundary constraints and vehicle spacing constraints, because the value of the lane width dimension directly affects the safety threshold of the lateral spacing. When the lane width dimension is 3.5 meters, the system determines that the lateral safety spacing should be no less than 0.8 meters. The traffic signal constraint and the speed limit constraint establish a time sequence correlation dependency, the remaining time of 23 seconds in the red light state combined with the current vehicle speed of 15 meters per second determines the braking distance requirement of 35 meters. The pedestrian dynamic constraint and the lane boundary constraint constitute a spatial correlation dependency, when a pedestrian is detected to be located within 0.5 meters outside the lane boundary, the system reduces the effective width of the lane boundary to 3 meters, reserving a larger safety buffer space.

[0127] The reasoning constraint network combines the semantic dimension characteristics based on the established correlation dependency relationship. The system combines the lane width dimension of 3.5 meters, the lane line type dimension of solid line type, and the lateral safety spacing of 0.8 meters into a lateral space constraint layer. This constraint layer determines that the vehicle's available lateral activity range is from 1.35 meters to the left of the lane center line to 1.35 meters to the right. The system combines the signal state dimension of red light, the remaining time of 23 seconds, the current vehicle speed of 15 meters per second, and the braking distance of 35 meters into a longitudinal dynamic constraint layer. This constraint layer calculates that the vehicle needs to start braking action 13 meters before the stop line. The system combines the pedestrian position of 0.5 meters from the lane boundary, the pedestrian moving speed of 1.2 meters per second, and the current vehicle speed of 15 meters per second into a dynamic obstacle avoidance constraint layer. This constraint layer requires the vehicle to maintain a safety distance of at least 2 meters from the predicted path of the pedestrian.

[0128] A multi-dimensional constraint expression reflecting the characteristics of the current driving scene is constructed, and a lateral spatial constraint layer, a longitudinal dynamic constraint layer and a dynamic obstacle avoidance constraint layer jointly form a 3-dimensional constraint expression structure. In the lateral dimension, the constraint expression explicitly defines that the vehicle center point offset range is -1.35 meters to 1.35 meters. In the longitudinal dimension, the constraint expression stipulates that the vehicle deceleration curve needs to complete the speed change from 15 meters per second to 0 within a distance of 35 meters. In the time dimension, the constraint expression limits that the entire braking process needs to be completed within 23 seconds, leaving 3 seconds as a system response margin.

[0129] The action boundary of each dimension constraint is determined. The action boundary of the lateral spatial constraint is set as the road area within 80 meters in front of the vehicle, and the lane curvature change beyond the range causes the constraint parameters to be recalculated. The action boundary of the longitudinal dynamic constraint extends from the current position of the vehicle to the stop line position, and the action distance is 48 meters. The action boundary of the dynamic obstacle avoidance constraint establishes a circular area with a radius of 5 meters centered on the pedestrian, and the area is dynamically updated as the pedestrian moves. When the distance between the vehicle and the edge of the circular area is less than 10 meters, the dynamic obstacle avoidance constraint starts to take effect and affects path planning.

[0130] A scene constraint mode is generated according to the multi-dimensional constraint expression and the action boundary. The generated scene constraint mode includes a constraint type identifier, a constraint parameter set and a constraint priority configuration. The constraint type identifier marks the current scene as a red light stop scene, and the corresponding code is scene type 003. The constraint parameter set records that the lateral offset limit is -1.35 meters to 1.35 meters, the braking distance requirement is 35 meters, the obstacle avoidance safety distance is 2 meters, and the speed change rate is not more than 3 meters per second acceleration. The constraint priority configuration sets the longitudinal dynamic constraint as the highest priority level 9, the dynamic obstacle avoidance constraint as the second highest priority level 8, and the lateral spatial constraint as the third priority level 7.

[0131] A hierarchical correspondence relationship between the scene constraint mode and the reasoning constraint network is established. The constraint type identifier in the scene constraint mode is mapped to the scene identification layer of the reasoning constraint network, and the layer stores 230 pre-defined scene types. The constraint parameter set is mapped to the parameter calculation layer of the reasoning constraint network, and the layer includes 60 calculation nodes for processing constraint parameters in different dimensions. The constraint priority configuration is mapped to the decision arbitration layer of the reasoning constraint network, and the layer solves the conflict between multiple constraints through a 9-level priority system. When the lateral spatial constraint requires the vehicle to keep driving in the lane center, and the dynamic obstacle avoidance constraint requires the vehicle to offset 0.6 meters to the left, the decision arbitration layer selects the instruction of the dynamic obstacle avoidance constraint according to the rule that the priority level 8 is higher than the priority level 7, and the vehicle finally executes the lateral adjustment action of offsetting 0.6 meters to the left.

[0132] In a second aspect, the embodiment of the present application provides an electronic device, comprising:

[0133] a processor;

[0134] a memory for storing processor-executable instructions;

[0135] wherein the processor is configured to invoke the instructions stored by the memory to perform the method as described above.

[0136] In a third aspect, the present application provides a computer readable storage medium, having stored thereon computer program instructions, which when executed by a processor implement the method as described above.

[0137] The present application can be a method, an apparatus, a system, and / or a computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions stored therein, which, when executed by a processor, perform various aspects of the present application.

[0138] It should be noted that the above-mentioned embodiments are merely used to illustrate the technical solutions of the present application, rather than limiting the present application; although the present application has been described in detail with reference to the above-mentioned embodiments, those skilled in the art should understand that the technical solutions recorded in the above-mentioned embodiments can be modified or equivalent replacements can be made to some or all of the technical features; and the 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 application.

Claims

1. A method for detecting interaction anomalies in in-vehicle devices based on TBOX and scene awareness, characterized in that, include: Obtain the driving scenario features of the in-vehicle device at the current moment, and construct a scenario constraint model based on the driving scenario features; Based on semantic ontology reasoning, the vehicle's driving state, environmental conditions and hierarchical safety rules are mapped into a semantic constraint space, and a legality interval for interactive behavior based on fuzzy boundaries is defined in the semantic constraint space. The operation behavior data is semantically parsed and mapped to obtain preliminary semantic labels. The initial semantic tags are dynamically associated with historical interaction sequences. Within the legality range of the interaction behavior, a multidimensional similarity measure between the current operation behavior and the historical interaction sequence is calculated. Contextual information is then fused to obtain the association analysis results. Design a knowledge-driven hierarchical reasoning strategy, which jointly decides on the association analysis results and the semantic constraint space, evaluates the intention rationality and scenario adaptability of the operation behavior through the hierarchical reasoning strategy, and generates a semantic reasoning result with confidence based on the legality interval of the interaction behavior. Based on the semantic reasoning results and their confidence levels, the response execution mode of the in-vehicle device is controlled, and user feedback information and environmental change information after execution are collected. The feedback data is used to adaptively adjust the semantic constraint space to achieve dynamic optimization of the model. Construct a scenario constraint model based on the driving scenario characteristics; Based on semantic ontology reasoning, the vehicle's driving state, environmental conditions, and graded safety rules are mapped into a semantic constraint space, including: Based on the characteristics of the driving scenario, the driving state information and environmental condition information of the vehicle are extracted. A scene semantic ontology is constructed based on the semantic ontology reasoning framework. In the scene semantic ontology, the semantic association between the driving state node and the environmental condition node is established through weighted reasoning rules to form a semantic association model. Based on the semantic association model, the risk characteristics of the current driving scenario are analyzed to determine the applicable set of safety rules. The set of safety rules is then classified according to the degree of risk association in the semantic association model to generate graded safety rules with risk levels. At the same time, the semantic association model is used to model the association dependencies between the graded safety rules to construct a rule association network. The semantic ontology reasoning framework performs cross-ontology semantic reasoning on the driving state information, environmental condition information, and hierarchical safety rules. Based on the semantic association model and the rule association network, a dynamic constraint boundary is constructed, mapping the vehicle's driving state, environmental conditions, and hierarchical safety rules into a semantic constraint space.

2. The method according to claim 1, characterized in that, The security rule set is classified according to the degree of risk association in the semantic association model to generate hierarchical security rules with risk levels; simultaneously, the semantic association model is used to model the association dependencies between the hierarchical security rules to construct a rule association network, including: The risk correlation degree between driving scenario elements is extracted from the semantic association model. Based on the risk correlation degree, each safety rule in the safety rule set is risk-assessed. The risk level identifier corresponding to each safety rule is determined according to the risk correlation degree. The safety rule set is classified according to the risk level identifier to generate graded safety rules with risk levels. Based on the triggering conditions and constraint scope of each security rule in the hierarchical security rules, the pre-dependencies and mutual exclusion conflicts between the hierarchical security rules are identified, and a description of the inter-rule dependencies is formed. The semantic association model is used to perform semantic parsing on the relationship dependency description between the rules. The semantic reasoning mechanism in the semantic association model is used to quantify the relationship dependency strength between the hierarchical security rules. The hierarchical security rules are used as network nodes, and the relationship dependency strength is used as the connection relationship between the nodes to construct a rule association network.

3. The method according to claim 1, characterized in that, In the semantic constraint space, a legality interval for interactive behaviors based on fuzzy boundaries is defined. Semantic parsing and mapping are performed on the operational behavior data to obtain preliminary semantic labels, including: Obtain constraint boundary parameters corresponding to different risk levels from the semantic constraint space, construct fuzzy boundary functions in the semantic constraint space based on the constraint boundary parameters, define the transition region of the constraint boundary through the fuzzy boundary functions, and define the legality interval of interactive behavior based on the fuzzy boundary in the semantic constraint space. Obtain operation behavior data, perform semantic parsing on the operation type features and operation parameter features in the operation behavior data, map the operation type features and operation parameter features to semantic dimension coordinates in the semantic constraint space, and obtain preliminary semantic labels based on the position of the semantic dimension coordinates in the legality interval of the interaction behavior.

4. The method according to claim 1, characterized in that, The initial semantic tags are dynamically associated with historical interaction sequences. Within the legality range of the interaction behavior, a multidimensional similarity measure between the current operation behavior and the historical interaction sequence is calculated. The association analysis results are obtained by fusing contextual information, including: The initial semantic tags are mapped and analyzed with the historical semantic tags in the historical interaction sequence. The association structure between semantic tags is constructed based on the dynamic association network, and the hierarchical relationship between tags is determined in the dynamic association network to form a tag association model. Based on the tag association model, semantic dimension features of the current operation behavior are extracted from the legality range of the interaction behavior, and a semantic feature vector is constructed; based on the tag association model, semantic dimension features of historical operation behaviors in the historical interaction sequence are modeled to generate a set of historical feature vectors. Within the legality range of the interaction behavior, a spatial distance metric is calculated using the semantic feature vector and the set of historical feature vectors, and a semantic association metric is calculated by combining the hierarchical relationship in the dynamic association network; a weight allocation strategy is determined according to the label association model, and the spatial distance metric and the semantic association metric are weighted and fused to obtain a multidimensional similarity metric between the current operation behavior and the historical interaction sequence. Obtain contextual information, construct a calibration model using the dynamic association network and the multidimensional similarity metric, and dynamically calibrate the multidimensional similarity metric; then fuse the calibrated multidimensional similarity metric with the contextual information to generate association analysis results.

5. The method according to claim 1, characterized in that, Design a knowledge-driven hierarchical reasoning strategy, jointly decide with the association analysis results and the semantic constraint space, evaluate the intention rationality and scenario adaptability of the operation behavior through the hierarchical reasoning strategy, and generate a semantic reasoning result with confidence based on the legality interval of the interaction behavior, including: A knowledge base is constructed and a hierarchical reasoning strategy is designed based on the knowledge base. The association analysis results are semantically decomposed and transformed to the intent reasoning layer of the hierarchical reasoning strategy. A reasoning constraint network is constructed based on the comprehensive association degree in the association analysis results and the driving intent knowledge. The intention semantic features of the current operation behavior are analyzed through the reasoning constraint network to form an intent rationality result. Based on the inference constraint network, the semantic constraint space is dimensionally decomposed to construct the scene constraint pattern corresponding to the current driving scenario. The intention rationality result and the scene constraint pattern are semantically reconstructed and transformed to the scene inference layer. The hierarchical inference strategy is used to establish the scene adaptation relationship and form the scene adaptation result. The intention rationality result and the scenario adaptability result are constructed in a hierarchical relationship in the reasoning constraint network. Based on the association analysis result and the semantic constraint space, the feature combination method is determined to generate a comprehensive reasoning result. Based on the distribution characteristics of the comprehensive reasoning results within the legality interval of the interactive behavior, the confidence parameter is calculated using the reasoning constraint network to generate a semantic reasoning result with confidence.

6. The method according to claim 5, characterized in that, Based on the inference constraint network, the semantic constraint space is decomposed dimensionally to construct the scene constraint pattern corresponding to the current driving scenario, including: The constraint elements in the semantic constraint space are deconstructed hierarchically according to the semantics of the driving scenario. Each constraint element is decomposed into corresponding semantic dimension features, and the correlation and dependency between constraint elements are established in the inference constraint network. Based on the aforementioned association dependencies, the semantic dimension features are hierarchically combined to construct a multidimensional constraint expression reflecting the characteristics of the current driving scenario, and the action boundaries of each dimension constraint are determined. A scenario constraint pattern is generated based on the multidimensional constraint expression and the action boundaries, and a hierarchical correspondence between the scenario constraint pattern and the inference constraint network is established.

7. A vehicle-mounted device interaction anomaly detection system based on TBOX and scene awareness, used to implement the method of any one of claims 1-6, characterized in that, include: The first unit is used to acquire the driving scenario features of the in-vehicle device at the current moment, and to construct a scenario constraint model based on the driving scenario features; Based on semantic ontology reasoning, the vehicle's driving state, environmental conditions and hierarchical safety rules are mapped into a semantic constraint space, and a legality interval for interactive behavior based on fuzzy boundaries is defined in the semantic constraint space. The operation behavior data is semantically parsed and mapped to obtain preliminary semantic labels. The second unit is used to dynamically associate the preliminary semantic tags with the historical interaction sequence, calculate the multidimensional similarity measure between the current operation behavior and the historical interaction sequence within the legality range of the interaction behavior, and obtain the association analysis result by fusing contextual information. The third unit is used to design a knowledge-driven hierarchical reasoning strategy, which jointly makes decisions with the association analysis results and the semantic constraint space, evaluates the intention rationality and scenario adaptability of the operation behavior through the hierarchical reasoning strategy, and generates a semantic reasoning result with confidence based on the legality interval of the interaction behavior. The fourth unit is used to control the response execution mode of the vehicle-mounted device based on the semantic reasoning results and their confidence levels, and to collect user feedback information and environmental change information after execution. The feedback data is used to adaptively adjust the semantic constraint space to achieve dynamic optimization of the model.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images