Training assistance method and system based on two-way information feedback in driving training

Through graph neural network and knowledge graph, combined with collaborative filtering and deep reinforcement learning, personalized suggestions for two-way information feedback in driving training are realized, and the problems of insufficient data support and one-way information transmission in traditional training are solved, which improves training efficiency and Student experience.

CN118887061BActive Publication Date: 2024-12-03YIXIAN INTELLIGENCE
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411388826.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-10-08
Publication Date
2024-12-03
Estimated Expiration
2044-10-08

AI Technical Summary

Technical Problem

Traditional driving training lacks objective and scientific data support and timely and effective feedback mechanisms, resulting in low training efficiency and poor learning experience for students. The existing system mainly adopts a one-way information transmission model, ignoring the students' subjective feelings and personalized needs. .

Method used

Using a two-way information feedback system based on graph neural network and knowledge graph, by building driving scenario diagrams and driving knowledge networks, collecting and analyzing driver operation, vehicle status and environmental perception data, generating personalized training suggestions, and collaborative filtering and depth Reinforcement learning optimization suggestions are delivered to the driver in real time.

Benefits of technology

It improves the efficiency and effectiveness of driving training, dynamically optimizes the training plan, meets the personalized needs of students, and enhances the enthusiasm and learning experience of students.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118887061B_ABST
    Figure CN118887061B_ABST
Patent Text Reader

Abstract

The present invention provides a training assistance method and system based on two-way information feedback in driving training, which relates to the field of intelligent driving technology, including constructing a driving scene graph structure, inputting driver operation data, vehicle status data and environmental perception data into the driving scene graph structure, and obtaining a high-dimensional representation of the driving scene through message transmission and node update mechanism; performing subgraph matching on the high-dimensional representation of the driving scene in a driving knowledge network to find out the knowledge points and skill requirements most relevant to the current driving scene; generating personalized training suggestions based on the matching results and the driver's historical performance; using a collaborative filtering algorithm to optimize the generated training suggestions based on the learning experience of similar drivers; delivering the optimized training suggestions to the driver in real time through visual prompts, voice commands or tactile feedback; and recording the driver's response to the training suggestions, including whether the suggestions are executed, the execution effect and the change in driving performance after execution.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to technology, and in particular to a training assistance method and system based on two-way information feedback in driving training. Background Art

[0002] Traditional driver training mainly relies on the experience of instructors and the subjective feelings of trainees, lacking objective and scientific data support and timely and effective feedback mechanisms, resulting in low training efficiency and less than ideal learning experience for trainees.

[0003] In the existing technology, some driving training institutions have introduced computer-assisted teaching systems, using multimedia courseware, virtual simulation, online testing and other means to provide students with more intuitive and interactive learning content. However, these systems mainly focus on the presentation of training content and pay insufficient attention to the actual learning effect of students. The data generated by students during the learning process is not effectively utilized, and it is difficult for coaches to understand students' learning progress and weak links in a timely manner, and they are unable to provide targeted guidance.

[0004] In addition, most existing driving training assistance systems adopt a one-way information transmission mode, that is, the system provides training content and evaluation feedback to trainees, but lacks a channel for trainees to feedback their own status and needs to the system. This one-way interaction method ignores the trainees' subjective feelings and personalized needs, which is not conducive to the dynamic optimization of training programs and the mobilization of trainees' enthusiasm. Summary of the invention

[0005] The embodiments of the present invention provide a training assistance method and system based on two-way information feedback in driving training, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] Provides a training assistance method based on two-way information feedback in driving training, including:

[0008] Construct a driving scene graph structure, take the driver, vehicle, and road environment as nodes in the graph, and the interaction between them as edges, and use graph neural networks to model the driving scene; collect driver operation data, vehicle status data, and environmental perception data, where the driver operation data includes steering wheel angle, accelerator pedal position, and brake pedal force; the vehicle status data includes vehicle speed, acceleration, and yaw angular velocity; and the environmental perception data includes the position of surrounding vehicles and road sign recognition results; input the driver operation data, vehicle status data, and environmental perception data into the driving scene graph structure, and obtain a high-dimensional representation of the driving scene through message passing and node update mechanisms;

[0009] Construct a driving rules and skills system based on the knowledge graph, store the knowledge of traffic laws, safe driving skills, and vehicle control essentials in the form of entities and relationships, and form a structured driving knowledge network; perform subgraph matching on the high-dimensional representation of the driving scene in the driving knowledge network to find the most relevant knowledge points and skill requirements for the current driving scene; generate personalized training suggestions based on the matching results and the driver's historical performance, including operating skills, safety precautions and related theoretical knowledge;

[0010] A collaborative filtering algorithm is used to optimize the generated training suggestions based on the learning experience of similar drivers. The optimized training suggestions are delivered to the driver in real time through visual prompts, voice commands or tactile feedback. The driver's response to the training suggestions is recorded, including whether the suggestions are executed, the execution effect and the changes in driving performance after execution. The parameters of the graph neural network and the knowledge structure of the driving knowledge network are dynamically adjusted according to the response. Based on the updated graph neural network and driving knowledge network, the next driving scenario is analyzed and suggestions are generated, forming a closed-loop feedback mechanism for continuous optimization.

[0011] In an optional embodiment,

[0012] The driver operation data, vehicle status data and environment perception data are input into the driving scene graph structure. Through message passing and node update mechanism, a high-dimensional representation of the driving scene is obtained, including:

[0013] The collected driver operation data, vehicle status data and environment perception data are aligned according to the timestamps, and the timestamp of the vehicle status data is used as a reference to find the closest time between the timestamp of each data in the driver operation data and the environment perception data and the timestamp of the vehicle status data, and the data is matched to the closest time. For a time that does not have corresponding driver operation data or environment perception data, the data value at the time is estimated by using the interpolation method, so as to obtain multi-dimensional time series data with a unified time granularity;

[0014] Mapping the multidimensional time series data into a driving scene graph structure, representing the driver, vehicle, road and environment as four types of nodes of the graph, respectively, and the initial attributes of each type of node include the static attributes of the corresponding element;

[0015] At each time step, the attributes of the driver node are updated to the combination of the current driver operation data and the historical driver operation data statistics, the attributes of the vehicle node are updated to the combination of the current vehicle speed, acceleration, yaw rate, position and heading angle and the historical vehicle state statistics, the attributes of the road node and the environment node are updated to the combination of the road curvature, slope, road slipperiness and the relative position and speed of the surrounding vehicles and obstacles, the attributes of the edge between the driver node and the vehicle node are set to the value of the accelerator and brake operation signals and the change rate of the operation signals, and the attributes of the edge between the vehicle node and the road node are set to the influence coefficient of the road curvature and slope on the vehicle motion and the adhesion coefficient between the tire and the road;

[0016] A graph convolutional neural network is used to extract the feature representation of each node. Each layer of the graph convolutional neural network obtains the node feature matrix of the next layer of the network by multiplying the node feature matrix of the current layer of the network with the normalized adjacency matrix of the graph and the weight matrix of the current layer and processing the network through an activation function. The input layer of the graph convolutional neural network is the node attributes and edge attributes. By stacking multiple layers of graph convolutional neural networks, the neighborhood information of each node is integrated to obtain the feature vector of each node.

[0017] The attention mechanism is used to perform weighted summation on the feature vectors of the nodes output by the graph convolutional neural network to obtain a high-dimensional representation vector of the initial driving scene. The attention mechanism obtains the attention weight of each node by calculating the inner product of the feature vector of each node and the learnable query vector and performing exponential function and normalization processing. The feature vectors of all nodes are weighted summed based on the attention weight to obtain a scene representation vector. Finally, the scene representation vectors obtained at different time steps are merged into a time-series scene representation sequence, which is input into a recurrent neural network or a one-dimensional convolutional network to extract the evolution characteristics of the scene in the time dimension and output the final high-dimensional representation of the driving scene.

[0018] In an optional embodiment,

[0019] Based on the knowledge graph, a driving rules and skills system is constructed. The knowledge of traffic regulations, safe driving skills, and vehicle control essentials is stored in the form of entities and relationships to form a structured driving knowledge network. The high-dimensional representation of the driving scene is matched with subgraphs in the driving knowledge network to find the most relevant knowledge points and skill requirements for the current driving scene, including:

[0020] From the text materials of traffic regulations, driving teaching materials and safety manuals, named entity recognition and relationship extraction technology are used to extract driving-related entities and relationships. The entity types include driving behavior, driving scene elements, driving risks and driving skills, and the relationship types include attribute relationships, constraint relationships, causal relationships and operational relationships; the extracted entities and relationships are represented in the form of triples; knowledge from different sources is deduplicated and integrated, and a globally consistent driving knowledge graph is constructed using knowledge alignment technology; the driving knowledge graph is stored in a graph database;

[0021] Using a graph neural network to encode multimodal perception data of a driving scene into a scene semantic vector; extracting key entities and their neighbor nodes related to the driving scene according to the ontology definition of the driving knowledge graph to form a scene-related knowledge subgraph; calculating the semantic similarity between the scene semantic vector and each node in the knowledge subgraph through a graph neural network to find the Top-K knowledge nodes most relevant to the driving scene;

[0022] Taking the Top-K nodes as the starting point, a breadth-first search algorithm is used to traverse the knowledge subgraph to extract the shortest relationship path between nodes, where the shortest relationship path reflects the reasoning chain between the driving scenario and the knowledge nodes; based on the shortest relationship path obtained by reasoning, a natural language description is summarized to find out the knowledge points and skill requirements most relevant to the current driving scenario.

[0023] In an optional embodiment,

[0024] Taking the Top-K nodes as the starting point, the knowledge subgraph is traversed using a breadth-first search algorithm to extract the shortest relationship path between nodes, including:

[0025] Taking the top-K knowledge nodes most relevant to the driving scenario as the center, extract the N-hop neighbor nodes of the top-K knowledge nodes in the driving knowledge graph to construct a knowledge subgraph containing more background knowledge, and the number of hops N of the N-hop neighbor nodes is adjusted according to the scale and density of the knowledge graph;

[0026] A breadth-first search algorithm is used to traverse the knowledge subgraph to explore all relationship paths between the Top-K knowledge nodes. The breadth-first search algorithm starts from the Top-K knowledge nodes, traverses the neighbor nodes of the Top-K knowledge nodes layer by layer, records the predecessor nodes and corresponding relationship edges of each knowledge node, and constructs a search tree of relationship paths;

[0027] In the search tree of the relationship path, the shortest path between the Top-K knowledge nodes is extracted as the reasoning chain, and the conditions of the shortest path include: the starting point and the end point of the path are both the Top-K knowledge nodes, the path does not contain repeated nodes, and the length of the path is the smallest; the shortest path is solved by a dynamic programming algorithm, the shortest distance between the knowledge nodes is updated through the state transfer equation, and the path is backtracked to obtain the node sequence on the shortest path;

[0028] Representing the shortest path as a sequence of relationship triples, wherein the relationship triples include two nodes and a relationship type connecting the two nodes, wherein the relationship type is a predefined relationship or a relationship customized according to a specific scenario;

[0029] In the case where there are multiple shortest paths between the Top-K knowledge nodes, a path fusion algorithm is used to merge the multiple shortest paths into a comprehensive path. The steps of the path fusion algorithm include: counting the frequency of nodes and relationships appearing in all shortest paths, sorting them from high to low by frequency, adding them to the comprehensive path in sequence, and deduplicating and compressing the comprehensive path.

[0030] In an optional embodiment,

[0031] Generate personalized training suggestions based on the matching results and the driver's historical performance, including operating skills, safety precautions and relevant theoretical knowledge; use collaborative filtering algorithms to optimize the generated training suggestions based on the learning experience of similar drivers, including:

[0032] Collect drivers' historical learning data, including completed knowledge points, test scores, and learning time, to form a driver-knowledge point rating matrix; use the Pearson correlation coefficient or cosine similarity measurement method to calculate the similarity between drivers based on their ratings of knowledge points; for the target driver, find the K neighbor drivers who are most similar to the target driver, use the historical ratings of the K neighbor drivers to predict the target driver's rating for each knowledge point, and generate a collaborative filtering Top-N knowledge point recommendation list from high to low according to the predicted ratings.

[0033] In an optional embodiment,

[0034] The method further comprises:

[0035] Based on the Top-N knowledge point recommendation list generated by collaborative filtering, the personalized recommendation problem is modeled as a Markov decision process, where the state is the driver's current learning state and historical interaction records, the action is each knowledge point in the Top-N knowledge point recommendation list, and the reward is the driver's immediate feedback on the recommended knowledge point, the long-term impact evaluation of the recommendation on the driver's overall learning effect, and the diversity of the recommended knowledge points;

[0036] The deep reinforcement learning algorithm is used to fit the action-state value through the Q-value function to learn the optimal knowledge point recommendation strategy; the driver characteristics, knowledge point characteristics, and historical interactions are embedded and vectorized and combined with the current time step information to form a state representation. A multi-layer perceptron or convolutional neural network is designed as the main structure of the Q network, which takes the state representation as input and outputs the Q value of each knowledge point in the Top-N recommendation list;

[0037] Adopting the experience replay mechanism, the state transition samples are stored in the replay pool, and small batch samples are randomly selected to train the Q network; at the same time, the target Q network and the evaluation Q network are constructed. The target Q network is used to generate the Q value target, and the parameters are regularly copied from the evaluation Q network;

[0038] When generating recommendations, the ε-greedy strategy is used to select actions, selecting the knowledge point with the highest Q value with a probability of 1-ε, and randomly exploring new knowledge points with a probability of ε, where ε decreases with the number of training rounds;

[0039] A fusion layer is designed, which takes the predicted scores of knowledge points generated by collaborative filtering and the Q values ​​of knowledge points output by the deep reinforcement learning algorithm as input, and generates a comprehensive score for each knowledge point through weighted summation or nonlinear transformation.

[0040] According to a second aspect of the embodiments of the present invention,

[0041] Provide a training assistance system based on two-way information feedback in driving training, including:

[0042] The first unit is used to construct a driving scene graph structure, taking the driver, vehicle, and road environment as nodes in the graph, and the interaction between them as edges, and using graph neural networks to model the driving scene; collecting driver operation data, vehicle status data, and environmental perception data, where the driver operation data includes steering wheel angle, accelerator pedal position, and brake pedal force; the vehicle status data includes vehicle speed, acceleration, and yaw angular velocity; and the environmental perception data includes the position of surrounding vehicles and road sign recognition results; inputting the driver operation data, vehicle status data, and environmental perception data into the driving scene graph structure, and obtaining a high-dimensional representation of the driving scene through message passing and node update mechanisms;

[0043] The second unit is used to build a driving rules and skills system based on the knowledge graph, store the knowledge of traffic regulations, safe driving skills, and vehicle control essentials in the form of entities and relationships, and form a structured driving knowledge network; perform sub-graph matching on the high-dimensional representation of the driving scene in the driving knowledge network to find the most relevant knowledge points and skill requirements for the current driving scene; generate personalized training suggestions based on the matching results and the driver's historical performance, including operating skills, safety precautions and related theoretical knowledge;

[0044] The third unit is used to optimize the generated training suggestions based on the learning experience of similar drivers using a collaborative filtering algorithm; deliver the optimized training suggestions to the driver in real time through visual prompts, voice commands or tactile feedback; record the driver's response to the training suggestions, including whether the suggestions are executed, the execution effect and the changes in driving performance after execution; dynamically adjust the parameters of the graph neural network and the knowledge structure of the driving knowledge network according to the response, and analyze and generate suggestions for the next driving scenario based on the updated graph neural network and driving knowledge network, forming a closed-loop feedback mechanism for continuous optimization.

[0045] According to a third aspect of the embodiments of the present invention,

[0046] An electronic device is provided, comprising:

[0047] processor;

[0048] a memory for storing processor-executable instructions;

[0049] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0050] A fourth aspect of the embodiments of the present invention is:

[0051] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0052] This technical solution achieves structured storage and efficient reasoning of driving knowledge by constructing a driving knowledge graph and combining multimodal perception data of driving scenarios. Through subgraph matching and the generation of reasoning chains, the system can provide highly relevant skill suggestions and knowledge points for the current driving scenario, thereby improving the intelligent decision-making ability of the autonomous driving system or providing effective driving guidance for human drivers.

[0053] This technical solution uses a breadth-first search algorithm to traverse the subgraphs in the driving knowledge graph, extract the shortest path between the Top-K knowledge nodes, and generate a comprehensive path through dynamic programming and path fusion algorithms. This solution can effectively match driving scenarios with knowledge points in the driving knowledge graph, form a reasoning chain, and provide intelligent knowledge reasoning and skill recommendations for driving scenarios. This method not only improves the efficiency of knowledge utilization, but also provides more comprehensive support for driving decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a flowchart of a training assistance method based on two-way information feedback in driving training according to an embodiment of the present invention;

[0055] Figure 2 The diagram is a structural diagram of a training assistance system based on two-way information feedback in driving training according to an embodiment of the present invention. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0058] Figure 1 FIG. 1 is a flow chart of a training assistance method based on two-way information feedback in driving training according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0059] S101. Construct a driving scene graph structure, take the driver, vehicle, and road environment as nodes in the graph, and the interaction relationship between them as edges, and use the graph neural network to model the driving scene; collect driver operation data, vehicle status data, and environmental perception data, where the driver operation data includes steering wheel angle, accelerator pedal position, and brake pedal force; the vehicle status data includes vehicle speed, acceleration, and yaw angular velocity; and the environmental perception data includes the position of surrounding vehicles and road sign recognition results; input the driver operation data, vehicle status data, and environmental perception data into the driving scene graph structure, and obtain a high-dimensional representation of the driving scene through message passing and node update mechanisms;

[0060] S102. Construct a driving rules and skills system based on the knowledge graph, store the knowledge of traffic laws, safe driving skills, and vehicle control essentials in the form of entities and relationships, and form a structured driving knowledge network; perform subgraph matching on the high-dimensional representation of the driving scene in the driving knowledge network to find the most relevant knowledge points and skill requirements for the current driving scene; generate personalized training suggestions based on the matching results and the driver's historical performance, including operating skills, safety precautions and related theoretical knowledge;

[0061] S103. Use collaborative filtering algorithms to optimize the generated training suggestions based on the learning experience of similar drivers; deliver the optimized training suggestions to the driver in real time through visual prompts, voice commands or tactile feedback; record the driver's response to the training suggestions, including whether the suggestions are executed, the execution effect and the changes in driving performance after execution; dynamically adjust the parameters of the graph neural network and the knowledge structure of the driving knowledge network according to the response, analyze the next driving scenario and generate suggestions based on the updated graph neural network and driving knowledge network, and form a closed-loop feedback mechanism for continuous optimization.

[0062] In an optional embodiment,

[0063] The driver operation data, vehicle status data and environment perception data are input into the driving scene graph structure. Through message passing and node update mechanism, a high-dimensional representation of the driving scene is obtained, including:

[0064] The collected driver operation data, vehicle status data and environment perception data are aligned according to the timestamps, and the timestamp of the vehicle status data is used as a reference to find the closest time between the timestamp of each data in the driver operation data and the environment perception data and the timestamp of the vehicle status data, and the data is matched to the closest time. For a time that does not have corresponding driver operation data or environment perception data, the data value at the time is estimated by using the interpolation method, so as to obtain multi-dimensional time series data with a unified time granularity;

[0065] Mapping the multidimensional time series data into a driving scene graph structure, representing the driver, vehicle, road and environment as four types of nodes of the graph, respectively, and the initial attributes of each type of node include the static attributes of the corresponding element;

[0066] At each time step, the attributes of the driver node are updated to the combination of the current driver operation data and the historical driver operation data statistics, the attributes of the vehicle node are updated to the combination of the current vehicle speed, acceleration, yaw rate, position and heading angle and the historical vehicle state statistics, the attributes of the road node and the environment node are updated to the combination of the road curvature, slope, road slipperiness and the relative position and speed of the surrounding vehicles and obstacles, the attributes of the edge between the driver node and the vehicle node are set to the value of the accelerator and brake operation signals and the change rate of the operation signals, and the attributes of the edge between the vehicle node and the road node are set to the influence coefficient of the road curvature and slope on the vehicle motion and the adhesion coefficient between the tire and the road;

[0067] A graph convolutional neural network is used to extract the feature representation of each node. Each layer of the graph convolutional neural network obtains the node feature matrix of the next layer of the network by multiplying the node feature matrix of the current layer of the network with the normalized adjacency matrix of the graph and the weight matrix of the current layer and processing the network through an activation function. The input layer of the graph convolutional neural network is the node attributes and edge attributes. By stacking multiple layers of graph convolutional neural networks, the neighborhood information of each node is integrated to obtain the feature vector of each node.

[0068] The attention mechanism is used to perform weighted summation on the feature vectors of the nodes output by the graph convolutional neural network to obtain a high-dimensional representation vector of the initial driving scene. The attention mechanism obtains the attention weight of each node by calculating the inner product of the feature vector of each node and the learnable query vector and performing exponential function and normalization processing. The feature vectors of all nodes are weighted summed based on the attention weight to obtain a scene representation vector. Finally, the scene representation vectors obtained at different time steps are merged into a time-series scene representation sequence, which is input into a recurrent neural network or a one-dimensional convolutional network to extract the evolution characteristics of the scene in the time dimension and output the final high-dimensional representation of the driving scene.

[0069] For example, the core of this technical solution is to build a driving scene graph structure, combine driver operation data, vehicle status data and environmental perception data, and generate a high-dimensional representation of the driving scene through message passing and node update mechanism, thereby providing support for autonomous driving and driving behavior analysis. The following are the detailed steps to implement this technical solution, including specific technical details and data processing flow.

[0070] First, the system collects data from three aspects: Driver operation data: including operation signals such as throttle, brake, and steering wheel angle. Vehicle status data: including vehicle speed, acceleration, yaw rate, position, and heading angle. Environmental perception data: including road curvature, slope, road slipperiness, relative position and speed of surrounding vehicles and obstacles. The collection frequency of these data may be different, so their time granularity needs to be unified. The specific method is:

[0071] Based on the timestamp of the vehicle status data, find the moment in the driver operation data and environmental perception data that is closest to the timestamp of the vehicle status data, and align these data to the same time point. If the driver operation data or environmental perception data is missing at a certain moment, the data value at that moment is estimated through the interpolation algorithm. Interpolation can infer the data value at the missing moment by calculating the data change trend at adjacent moments. For example, assume that the vehicle status data is collected at a time interval of 100 milliseconds, while the driver operation data is collected at an interval of 200 milliseconds, and the environmental perception data is collected at an interval of 300 milliseconds. The time alignment process will find the closest time point to align the data, and interpolate the missing data, and finally generate multidimensional time series data with a unified time granularity.

[0072] Next, the aligned multi-dimensional time series data is mapped into a driving scene graph structure. The graph structure contains four types of nodes, representing:

[0073] Driver node: represents the statistical information of the driver's operation behavior and its historical operation behavior. Vehicle node: represents the statistical information of the vehicle's current state and its historical state. Road node: represents the geometric characteristics and physical conditions of the current road. Environment node: represents the surrounding obstacles, other vehicles and their motion characteristics. The initial attributes of each type of node contain static attributes related to the element. For example, the driver node can contain the driver's reaction time, driving experience, etc. The vehicle node can contain the vehicle's mass, power system parameters, etc. The road node can contain the speed limit of the road, lane width, etc. The environment node can contain weather conditions, traffic density, etc.

[0074] At each time step (i.e. each aligned time point), the attributes of each node are updated according to the data at the current moment:

[0075] Driver node: Its attributes are updated to the combination of the current driver operation data (such as throttle, brake, etc.) and the statistics of historical driver operation data. For example, at a certain moment, the driver's throttle pedal strength is 30%, and the historical operation statistics can include the average value and change rate of throttle usage in the past 5 seconds.

[0076] Vehicle node: Updated with the current vehicle speed, acceleration, yaw rate, position and heading angle, as well as historical statistics of these vehicle states. For example, if the current vehicle speed is 60 km / h, historical statistics may include the average speed and acceleration change trend in the past 10 seconds.

[0077] Road nodes and environment nodes: The attributes of these nodes are updated to include road curvature, slope, road slipperiness, and the relative positions and speeds of other vehicles and obstacles around. For example, at the current moment, the curvature of the road is 0.02, the slope is 3%, the relative speed of the surrounding vehicles is 15 km / h, and the obstacle distance is 50 meters.

[0078] At the same time, the attributes of the edges between the nodes will also be updated:

[0079] Edges between driver nodes and vehicle nodes: The attributes of the edges are set to the driver's throttle and brake operation signals, as well as the change rates of these operation signals. For example, at the current moment, the throttle is 30%, the brake is 0%, and the change rate is 2% / s.

[0080] Edges between vehicle nodes and road nodes: The attributes of the edges reflect the influence of road curvature and slope on vehicle motion, as well as the adhesion coefficient between tires and the road surface. For example, the slope of the road at the current moment has a 2% influence on the motion of the vehicle, and the adhesion coefficient between the tires and the road surface is 0.8.

[0081] In the driving scene graph, a graph convolutional network (GCN) is used to extract the feature representation of each node. Each layer of GCN extracts features through the following steps:

[0082] The node features of the current layer are weighted and summed with the adjacency matrix of the graph, and the information is propagated to the adjacent nodes by combining the structure of the graph and the correlation between nodes. The weighted features are processed by a nonlinear activation function to generate the node feature representation of the next layer. The initial data input to GCN includes the attributes of the node (such as driver operation signals, vehicle status, etc.) and the attributes of the edge (such as throttle, brake signals, the impact of the road on the vehicle, etc.). By stacking multiple layers of GCN, each node will gradually fuse the information of other nodes in its neighborhood, and finally obtain the feature vector of the node. This process effectively captures the dynamic relationship between the driver, vehicle, road and environment.

[0083] The feature vector of each node extracted by GCN represents the importance of the node in the entire driving scene. However, different nodes may contribute differently to the overall scene, so an attention mechanism is used to weight the sum of the feature vectors of the nodes to generate a high-dimensional representation of the driving scene.

[0084] Specifically, the attention mechanism calculates the attention weight of each node by performing an inner product operation on the feature vector of each node and a learnable query vector. These weights are then normalized and applied to the corresponding node features, and finally the weighted sum of the features of all nodes is taken to obtain the representation vector of the scene. For example, at a certain moment, the driver's operation may have a greater impact on the scene, so the weight of the driver node will be higher, while the weight of the environment node may be lower.

[0085] After obtaining the high-dimensional representation vector of the driving scene at each time step, these vectors are combined into a time series. Then, the evolution characteristics of the scene in the time dimension are further extracted through a recurrent neural network (RNN) or a one-dimensional convolutional network (1D-CNN). Recurrent neural network: extracts potential features in the time series by capturing the changing trend of the scene representation vector over time, which is suitable for processing continuous time series data. One-dimensional convolutional network: extracts local time features through a sliding convolution window, which is suitable for processing time series data with local features.

[0086] Finally, after feature extraction in the time dimension, the system outputs the final high-dimensional representation of the driving scene, which can be used as input for subsequent tasks such as autonomous driving decision-making, driving behavior analysis, or event prediction.

[0087] Assume that the collected driver operation data is: 30% throttle, 0% brake, 5° steering angle, the vehicle status data is: speed 60 km / h, acceleration 0.5 m / s², yaw rate 0.02 rad / s, and the environmental perception data is: road curvature 0.02, slope 3%, obstacle distance 50 meters, obstacle relative speed 15 km / h. Through the above steps, the system maps these data to a graph structure, and after processing by GCN and attention mechanism, a high-dimensional representation of the driving scene is obtained.

[0088] This technical solution constructs a driving scene graph structure and combines graph convolutional neural networks and attention mechanisms to extract complex relationships and dynamic information in driving scenes, and ultimately generates a high-dimensional representation of the driving scene. This high-dimensional representation provides an effective feature basis for autonomous driving systems and driving behavior analysis.

[0089] In an optional embodiment,

[0090] Based on the knowledge graph, a driving rules and skills system is constructed. The knowledge of traffic regulations, safe driving skills, and vehicle control essentials is stored in the form of entities and relationships to form a structured driving knowledge network. The high-dimensional representation of the driving scene is matched with subgraphs in the driving knowledge network to find the most relevant knowledge points and skill requirements for the current driving scene, including:

[0091] From the text materials of traffic regulations, driving teaching materials and safety manuals, named entity recognition and relationship extraction technology are used to extract driving-related entities and relationships. The entity types include driving behavior, driving scene elements, driving risks and driving skills, and the relationship types include attribute relationships, constraint relationships, causal relationships and operational relationships; the extracted entities and relationships are represented in the form of triples; knowledge from different sources is deduplicated and integrated, and a globally consistent driving knowledge graph is constructed using knowledge alignment technology; the driving knowledge graph is stored in a graph database;

[0092] Using a graph neural network to encode multimodal perception data of a driving scene into a scene semantic vector; extracting key entities and their neighbor nodes related to the driving scene according to the ontology definition of the driving knowledge graph to form a scene-related knowledge subgraph; calculating the semantic similarity between the scene semantic vector and each node in the knowledge subgraph through a graph neural network to find the Top-K knowledge nodes most relevant to the driving scene;

[0093] Taking the Top-K nodes as the starting point, a breadth-first search algorithm is used to traverse the knowledge subgraph to extract the shortest relationship path between nodes, where the shortest relationship path reflects the reasoning chain between the driving scenario and the knowledge nodes; based on the shortest relationship path obtained by reasoning, a natural language description is summarized to find out the knowledge points and skill requirements most relevant to the current driving scenario.

[0094] Exemplarily, the core of this technical solution is to build a driving rules and skill system based on the knowledge graph, store information such as driving behavior, traffic regulations, and safe driving skills in a structured form, and perform sub-graph matching based on the high-dimensional representation of the driving scene to find the knowledge points and skill requirements that are most relevant to the current driving scene.

[0095] First, we use natural language processing (NLP) technology to extract driving-related entities and relationships from text materials such as traffic regulations, driving teaching materials, and safe driving manuals. The specific steps are as follows:

[0096] Use named entity recognition (NER) technology to extract key entities in the driving field. For example: "driving behavior", "driving scene elements", "driving risks", "driving skills", etc. Entity types include: Driving behavior: such as "steering", "braking", "lane change", etc. Driving scene elements: such as "road curvature", "traffic lights", "obstacles ahead", etc. Driving risks: such as "skidding", "collision risk", "blind spot", etc. Driving skills: such as "emergency braking operation", "turning skills in rainy days", "lane keeping skills", etc.

[0097] Use relation extraction technology to extract the relationship between entities. Relationship types include:

[0098] Attribute relationship: such as "the steering angle is 30 degrees". Constraint relationship: such as "you must stop at a red light". Causal relationship: such as "sudden braking causes the vehicle to skid". Operational relationship: such as "slow down when the road is slippery in rainy days".

[0099] The extracted entities and relations are represented in the form of triples, and the form of triples is: (entity 1, relation, entity 2); for example:

[0100] ("driver", "execution", "emergency braking operation"); ("rainy day", "cause", "slippery road surface"); ("lane keeping skills", "applicable to", "urban roads"); In this way, driving-related knowledge is structured into triplets to form actionable knowledge units.

[0101] Due to the diversity of knowledge sources, there may be repeated expressions or conflicts of the same knowledge in different texts. Therefore, knowledge alignment technology is used to remove duplicates and merge knowledge from different sources to build a globally consistent driving knowledge graph. The specific steps include:

[0102] Entity alignment: Identify different expressions of the same entity in different sources and unify them. For example, "emergency brake" and "emergency stop operation" can be represented as the same entity. Relation alignment: Similarly, unify different expressions of the same relationship. Through these steps, a unified and globally consistent driving knowledge graph is finally constructed.

[0103] The constructed driving knowledge graph is stored in a graph database. The graph database can efficiently store and query graph structure data and support complex graph operations and path queries, which provides a basis for subsequent subgraph matching and reasoning.

[0104] In the autonomous driving scenario, the system obtains multimodal perception data from multiple sensors (such as cameras, radars, lidars, etc.), including vision, radar signals, vehicle status, etc. These data are first processed into a high-dimensional representation of the driving scene.

[0105] In order to connect these multimodal data with the structured knowledge in the knowledge graph, a graph neural network (GNN) is used to encode the multimodal data and convert it into a scene semantic vector, which represents the core semantic information of the current driving scene.

[0106] For example, in a certain driving scenario, the sensor detects a decelerating vehicle ahead on the road, a slight bend in the road, light rain, etc. After these multimodal perception data are processed by the graph neural network, a semantic vector is generated, representing the core semantic features of the scene.

[0107] According to the ontology definition of the driving knowledge graph, the key entities and their neighbor nodes related to the current driving scenario are extracted from the global knowledge graph to form a scenario-related knowledge subgraph. The specific steps are as follows:

[0108] Find the entities most relevant to the scene semantic vector from the driving knowledge graph. For example, if the scene contains information such as "rainy day", "slight curve", "vehicle ahead slows down", then the entities related to these scene elements may include "driving skills in rainy days", "slowing down skills on curves", etc. Extract the neighbor nodes directly related to these entities and build a subgraph containing the core elements of the scene.

[0109] Use graph neural network to calculate the semantic similarity between the scene semantic vector and each node in the knowledge subgraph, and find the Top-K knowledge nodes most relevant to the driving scene. The specific process is as follows:

[0110] Match the scene semantic vector with the feature vector of each node in the subgraph and calculate their similarity. Select the top-K most relevant knowledge nodes based on the similarity. For example, in a rainy day scene, the system may select knowledge nodes related to "driving skills in rainy days", "driving skills on curves", "vehicle deceleration skills", etc.

[0111] After obtaining the top-K most relevant knowledge nodes, we use the breadth-first search (BFS) algorithm to traverse the knowledge subgraph starting from these nodes and extract the shortest relationship paths between nodes. These paths reflect the reasoning chain between driving scenarios and knowledge nodes.

[0112] For example, assuming that the current scene contains information such as "rainy day", "curve", and "vehicle ahead slows down", the system finds the following shortest reasoning path through breadth-first search: "rainy day" → "slippery road surface" → "vehicle speed should be reduced"; "curve" → "curve driving skills" → "should slow down and maintain a certain lane position"; "vehicle ahead slows down" → "should maintain a safe distance and slow down".

[0113] Based on the shortest relationship path obtained through reasoning, the system summarizes the reasoning chain into a natural language description and provides driving suggestions or skill requirements to the driver or the autonomous driving system. These suggestions can help the autonomous driving system make decisions or provide guidance to the driver.

[0114] For example, based on the above reasoning path, the system can generate the following natural language description: "According to the current road conditions in rainy days, you should reduce your speed and maintain a safe distance from the vehicle in front"; "When driving on a curve, it is recommended to slow down and maintain lane position to avoid the risk of skidding."

[0115] To better illustrate the whole process, here is a specific application of an example driving scenario:

[0116] Driving scenario: In light rain, the vehicle is driving on a slight curve, and there is a decelerating vehicle in front. Multimodal perception data: The camera detects that the vehicle in front is decelerating, the radar detects that the distance between the vehicle and the vehicle in front is decreasing, and the sensor detects that the road surface is slippery. Scene semantic vector: The system encodes these perception data into scene semantic vectors, reflecting core information such as rainy days, curves, and deceleration of the vehicle in front. Subgraph matching: Extract knowledge subgraphs related to rainy days, curves, and deceleration of the vehicle in front from the driving knowledge graph, and find the most relevant Top-K knowledge nodes, such as "driving skills in rainy days", "deceleration skills on curves", etc. Reasoning chain: Through breadth-first search, the system finds reasoning chains such as "rainy days → slippery road surface → speed should be reduced", "curve → speed should be reduced". Driving advice: Based on the reasoning chain, the system generates advice: "It is currently rainy and the road surface is slippery. It is recommended to slow down and keep a safe distance from the vehicle in front."

[0117] This technical solution achieves structured storage and efficient reasoning of driving knowledge by constructing a driving knowledge graph and combining multimodal perception data of driving scenarios. Through subgraph matching and the generation of reasoning chains, the system can provide highly relevant skill suggestions and knowledge points for the current driving scenario, thereby improving the intelligent decision-making ability of the autonomous driving system or providing effective driving guidance for human drivers.

[0118] In an optional embodiment,

[0119] Taking the Top-K nodes as the starting point, the knowledge subgraph is traversed using a breadth-first search algorithm to extract the shortest relationship path between nodes, including:

[0120] Taking the top-K knowledge nodes most relevant to the driving scenario as the center, extract the N-hop neighbor nodes of the top-K knowledge nodes in the driving knowledge graph to construct a knowledge subgraph containing more background knowledge, and the number of hops N of the N-hop neighbor nodes is adjusted according to the scale and density of the knowledge graph;

[0121] A breadth-first search algorithm is used to traverse the knowledge subgraph to explore all relationship paths between the Top-K knowledge nodes. The breadth-first search algorithm starts from the Top-K knowledge nodes, traverses the neighbor nodes of the Top-K knowledge nodes layer by layer, records the predecessor nodes and corresponding relationship edges of each knowledge node, and constructs a search tree of relationship paths;

[0122] In the search tree of the relationship path, the shortest path between the Top-K knowledge nodes is extracted as the reasoning chain, and the conditions of the shortest path include: the starting point and the end point of the path are both the Top-K knowledge nodes, the path does not contain repeated nodes, and the length of the path is the smallest; the shortest path is solved by a dynamic programming algorithm, the shortest distance between the knowledge nodes is updated through the state transfer equation, and the path is backtracked to obtain the node sequence on the shortest path;

[0123] Representing the shortest path as a sequence of relationship triples, wherein the relationship triples include two nodes and a relationship type connecting the two nodes, wherein the relationship type is a predefined relationship or a relationship customized according to a specific scenario;

[0124] In the case where there are multiple shortest paths between the Top-K knowledge nodes, a path fusion algorithm is used to merge the multiple shortest paths into a comprehensive path. The steps of the path fusion algorithm include: counting the frequency of nodes and relationships appearing in all shortest paths, sorting them from high to low by frequency, adding them to the comprehensive path in sequence, and deduplicating and compressing the comprehensive path.

[0125] Exemplarily, the core of this technical solution is based on the driving knowledge graph, and uses a breadth-first search algorithm to extract the shortest relationship path between the knowledge nodes most relevant to the driving scenario in the subgraph, and generates a comprehensive path through a path fusion algorithm to form a reasoning chain.

[0126] First, extract the top-K knowledge nodes that are most relevant to the current scene from the high-dimensional representation of the driving scene. These nodes represent knowledge points in the driving knowledge graph that are highly relevant to the current driving situation. For example, in a rainy driving scene, the top-K knowledge nodes that may be extracted include "driving skills in rainy days", "slippery roads", "speed control", etc.

[0127] In order to enrich the knowledge background, starting from the Top-K knowledge nodes, their N-hop neighbor nodes are extracted to construct the knowledge subgraph. The extraction of neighbor nodes is based on the scale and density of the driving knowledge graph, and the N value can be dynamically adjusted according to the actual application scenario.

[0128] Definition of hop count N: Hop count N represents the number of steps when expanding outward along the relationship edge starting from a node in the knowledge graph. For example, N=1 represents a direct neighbor node, and N=2 represents a neighbor node of a neighbor node.

[0129] Through this expansion, a subgraph containing more background knowledge can be constructed. For example, if one of the Top-K nodes is "driving skills in rainy days", its 1-hop neighbor nodes may include "slippery road surface", "should slow down", etc., and its 2-hop neighbor nodes may include "tire adhesion coefficient", etc.

[0130] In the constructed knowledge subgraph, the breadth-first search (BFS) algorithm is used to traverse the nodes in the graph to explore all relationship paths between the Top-K knowledge nodes. The specific steps are as follows:

[0131] Search starting point: Starting from any of the Top-K knowledge nodes, start traversing its neighbor nodes. Traversal layer by layer: During the search process, traverse all neighbor nodes connected to the current node layer by layer until all nodes related to the Top-K nodes are traversed. Record predecessor nodes and relationship edges: During the traversal process, record the predecessor node of each node (that is, which node it is accessed from) and the relationship edge connecting the two nodes. This process generates a search tree, in which each path represents a path from one Top-K node to another Top-K node through several intermediate nodes.

[0132] In the search tree, the system needs to extract the shortest relationship path between the Top-K knowledge nodes, that is, the shortest path from one Top-K node to another Top-K node. The extraction of the shortest path follows the following conditions:

[0133] The starting point and the end point are both Top-K nodes: The path must be from one Top-K node to another Top-K node and cannot contain other irrelevant nodes. The path does not contain duplicate nodes: The nodes in each path must not be repeated to ensure the validity of the path. The path length is minimum: The path length is determined by the number of nodes passed through, and the system will give priority to the path with the least number of nodes.

[0134] In order to efficiently solve the shortest path, the system uses a dynamic programming algorithm. Dynamic programming gradually finds the shortest path by continuously updating the shortest distance between nodes. Each time it updates, the system records the predecessor node of the current node so that after finding the shortest path, it can trace back the path and obtain the node sequence from the starting point to the end point.

[0135] For example, if there are three different paths from the node "driving skills in rainy weather" to the node "lane keeping skills", the system will dynamically update the length of each path and select the shortest one. Finally, the system traces back the path and obtains a node sequence, such as "driving skills in rainy weather" → "slippery road surface" → "speed control" → "lane keeping skills".

[0136] Each shortest path can be represented as a series of relationship triplets, that is, the relationship between two nodes. For example: (node ​​1, relationship, node 2); in the shortest path, each triple represents the relationship between two nodes. For example:

[0137] ("Rainy day driving tips", "Cause", "Slippery road surface"); ("Slippery road surface", "Leads to", "Speed ​​should be reduced"); These triples can be predefined relations, such as "Cause", "Leads to", "Applicable to", etc., or they can be customized according to specific driving scenarios.

[0138] Represent the shortest path as a sequence of triplets. For example:

[0139] (node ​​A, relationship 1, node B) → (node ​​B, relationship 2, node C) → (node ​​C, relationship 3, node D);

[0140] This sequence reflects the chain of reasoning between knowledge nodes. For example:

[0141] ("Rainy driving tips", "Cause", "Slippery road") → ("Slippery road", "Leads to", "Speed ​​should be reduced") → ("Speed ​​should be reduced", "Applicable", "Lane keeping tips");

[0142] In some cases, there may be multiple shortest paths between the top-K nodes. For example, there may be multiple shortest paths of the same length between "driving skills in rainy weather" and "lane keeping skills" through different intermediate nodes.

[0143] In order to solve the redundancy problem of multiple shortest paths, the system uses a path fusion algorithm to merge multiple shortest paths into a comprehensive path. The specific steps are as follows:

[0144] Count the frequency of nodes and relationships: Calculate the frequency of each node and relationship in all shortest paths. The higher the frequency, the more common the node or relationship is in multiple paths. Sort by frequency: Sort the nodes and relationships from high to low according to the frequency of occurrence. Add to the comprehensive path in sequence: Add nodes and relationships to the comprehensive path in sequence according to the frequency. De-duplicate and compress: When generating a comprehensive path, deduplicate the repeated nodes to ensure the simplicity of the path.

[0145] Assume that there are two shortest paths from "Rainy Weather Driving Skills" to "Lane Keeping Skills":

[0146] "Driving tips in rainy weather" → "Slippery road" → "Speed ​​should be reduced" → "Lane keeping tips";

[0147] "Driving tips in rainy weather" → "Obstructed vision" → "Speed ​​should be reduced" → "Lane keeping tips";

[0148] Through the path fusion algorithm, the system will count the frequency of each node. Since "vehicle speed should be reduced" appears in both paths with the highest frequency, it will appear in the comprehensive path first. In the end, the system may generate the following comprehensive path:

[0149] "Driving tips in rainy weather" → "Speed ​​should be reduced" → "Lane keeping tips".

[0150] This technical solution uses a breadth-first search algorithm to traverse the subgraphs in the driving knowledge graph, extract the shortest path between the Top-K knowledge nodes, and generate a comprehensive path through dynamic programming and path fusion algorithms. This solution can effectively match driving scenarios with knowledge points in the driving knowledge graph, form a reasoning chain, and provide intelligent knowledge reasoning and skill recommendations for driving scenarios. This method not only improves the efficiency of knowledge utilization, but also provides more comprehensive support for driving decisions.

[0151] In an optional embodiment,

[0152] Generate personalized training suggestions based on the matching results and the driver's historical performance, including operating skills, safety precautions and relevant theoretical knowledge; use collaborative filtering algorithms to optimize the generated training suggestions based on the learning experience of similar drivers, including:

[0153] Collect drivers' historical learning data, including completed knowledge points, test scores, and learning time, to form a driver-knowledge point rating matrix; use the Pearson correlation coefficient or cosine similarity measurement method to calculate the similarity between drivers based on their ratings of knowledge points; for the target driver, find the K neighbor drivers who are most similar to the target driver, use the historical ratings of the K neighbor drivers to predict the target driver's rating for each knowledge point, and generate a collaborative filtering Top-N knowledge point recommendation list from high to low according to the predicted ratings.

[0154] For example, first, the system needs to collect historical learning data of each driver. This data is used to evaluate the driver's learning status and knowledge mastery, mainly including the following:

[0155] Completed knowledge points: A list of knowledge points that the driver has learned and completed, recording the driver's learning progress on specific operating skills, safety precautions, theoretical knowledge, etc. For example, a driver may have completed knowledge points such as "driving skills in rainy days" and "overtaking skills".

[0156] Test score: The test result after learning each knowledge point can usually be a quantitative score, reflecting the driver's mastery of the knowledge point. For example, the driver scored 85 points in the "Rainy Day Driving Skills" test and 90 points in the "Overtaking Skills" test.

[0157] Learning time: The learning time of the driver on each knowledge point is used to evaluate the driver's dedication to different knowledge points. For example, a driver spent 2 hours on "Rainy Day Driving Skills" and 1 hour on "Overtaking Skills".

[0158] These data constitute a driver-knowledge point scoring matrix, in which each driver has a score for each knowledge point, and the score can be calculated comprehensively based on factors such as test scores and learning time.

[0159] Using the above data, the system constructs a driver-knowledge point scoring matrix, where rows represent drivers, columns represent knowledge points, and each cell of the matrix represents the driver's score for the knowledge point.

[0160] In order to generate personalized training recommendations, the system needs to find other drivers that are most similar to the target driver. Specifically, the similarity measurement method can adopt the following two common methods:

[0161] Pearson correlation coefficient: used to measure the linear correlation between two drivers' scores on knowledge points. It can eliminate the absolute value difference of the scores and focus on the relative change trend of the scores. Cosine similarity: used to measure the angle between the score vectors of two drivers, reflecting the similarity between them. Cosine similarity is more suitable for situations where the scoring scale is consistent. Through these methods, the system can calculate the similarity between each driver. Specifically, the system will compare the scores of two drivers on the same knowledge point. The closer the scores are, the higher the similarity.

[0162] The system calculates the similarity between drivers A and B. Assuming cosine similarity is used, the system compares the scores of drivers A and B on the common learning knowledge points (i.e., "driving skills in rainy weather" and "overtaking skills") to derive the similarity between them. For example, the calculation result may show that the similarity between drivers A and B is 0.9, indicating that they are very similar in learning behavior.

[0163] After calculating the similarity between drivers, the system finds the K most similar neighbor drivers for each target driver. The learning history and scores of these neighbor drivers will be used to predict the target driver's score for knowledge points that have not yet been learned.

[0164] Next, the system uses the historical scores of K neighboring drivers to predict the target driver's score for each unlearned knowledge point. For example, driver A has not yet learned the "emergency braking technique", and the system can predict that A's score for this knowledge point may be 85 points based on the scores of B and C on this knowledge point (B scored 88 points and C scored 82 points).

[0165] Based on the predicted score, the system generates a Top-N knowledge point recommendation list for the target driver, which contains the knowledge points that the target driver needs to learn most. The knowledge points in the recommendation list are sorted from high to low according to the predicted score. For example, the recommendation list generated by the system for driver A may be as follows: emergency braking skills (predicted score: 85); highway driving (predicted score: 80); in this way, the system can recommend the top priority knowledge points that driver A has not yet learned to help him conduct personalized learning.

[0166] The collaborative filtering algorithm can dynamically adjust the recommended content as the driver's learning progresses. When the driver has completed learning new knowledge points and updated the scoring matrix, the system will recalculate the similarity and generate an updated Top-N recommendation list based on the new scoring matrix. In this way, the system can continuously optimize training suggestions as the driver grows.

[0167] The system should design a feedback mechanism to collect feedback from drivers on their learning of recommended knowledge points. This can include data such as the driver's test scores and learning time after learning the recommended knowledge points. Based on this feedback, the system can further adjust the collaborative filtering model to optimize the recommendation effect. For example, if the driver has a significant learning effect on certain recommended knowledge points (such as a high test score), the system can prioritize the recommendation of advanced knowledge related to these knowledge points.

[0168] This technical solution collects the driver's historical learning data, constructs a driver-knowledge point scoring matrix, and uses a collaborative filtering algorithm to generate personalized training suggestions. By calculating the similarity between drivers, the system can find the neighbors most similar to the target driver, and predict the target driver's score for unlearned knowledge points based on the learning history of these neighbors, thereby generating optimized training suggestions. This solution not only takes into account the individual needs of drivers, but also further improves the training effect through dynamic optimization and feedback mechanisms, providing strong support for personalized learning of drivers.

[0169] In an optional embodiment,

[0170] The method further comprises:

[0171] Based on the Top-N knowledge point recommendation list generated by collaborative filtering, the personalized recommendation problem is modeled as a Markov decision process, where the state is the driver's current learning state and historical interaction records, the action is each knowledge point in the Top-N knowledge point recommendation list, and the reward is the driver's immediate feedback on the recommended knowledge point, the long-term impact evaluation of the recommendation on the driver's overall learning effect, and the diversity of the recommended knowledge points;

[0172] The deep reinforcement learning algorithm is used to fit the action-state value through the Q-value function to learn the optimal knowledge point recommendation strategy; the driver characteristics, knowledge point characteristics, and historical interactions are embedded and vectorized and combined with the current time step information to form a state representation. A multi-layer perceptron or convolutional neural network is designed as the main structure of the Q network, which takes the state representation as input and outputs the Q value of each knowledge point in the Top-N recommendation list;

[0173] Adopting the experience replay mechanism, the state transition samples are stored in the replay pool, and small batch samples are randomly selected to train the Q network; at the same time, the target Q network and the evaluation Q network are constructed. The target Q network is used to generate the Q value target, and the parameters are regularly copied from the evaluation Q network;

[0174] When generating recommendations, the ε-greedy strategy is used to select actions, selecting the knowledge point with the highest Q value with a probability of 1-ε, and randomly exploring new knowledge points with a probability of ε, where ε decreases with the number of training rounds;

[0175] A fusion layer is designed, which takes the predicted scores of knowledge points generated by collaborative filtering and the Q values ​​of knowledge points output by the deep reinforcement learning algorithm as input, and generates a comprehensive score for each knowledge point through weighted summation or nonlinear transformation.

[0176] For example, in order to optimize personalized recommendations, the recommendation system is modeled as a Markov decision process (MDP). In the MDP, the system finds the optimal knowledge point recommendation strategy through a series of interactions among states, actions, and rewards.

[0177] State: The state indicates the driver's current learning state and its historical interaction records. Specifically, the state includes information such as the driver's learning progress, completed knowledge points, test scores, and learning time. For example, a driver's state may be: completed "Rainy Day Driving Skills" and "Overtaking Skills", with scores of 85 and 90 points respectively, and a total learning time of 3 hours.

[0178] Action: An action is to select a knowledge point from the Top-N knowledge point recommendation list for recommendation. Each knowledge point can be considered as an action. For example, the current system recommends Top-N knowledge points for drivers, including "emergency braking skills", "highway driving skills", and "night driving skills", and each knowledge point corresponds to an optional action.

[0179] Reward: Rewards are evaluated comprehensively based on the driver's immediate feedback on the recommended knowledge points, the long-term impact of the recommendation on the driver's overall learning effect, and the diversity of the recommended knowledge points. Immediate feedback can be the test score of the driver after learning the knowledge point, the long-term impact can be the contribution of the knowledge point to the entire learning path, and diversity means avoiding recommending too much repetitive content.

[0180] Assuming that the driver has completed "Rainy Day Driving Skills" and "Overtaking Skills" in his current learning state, the system generates a Top-N recommendation list, including "Emergency Braking Skills", "Highway Driving Skills" and "Night Driving Skills". The system's task is to select the best one among these knowledge points for recommendation, and continuously optimize the recommendation strategy through deep reinforcement learning algorithms.

[0181] In deep reinforcement learning, the system uses the Q-value function to fit the "action-state value", that is, to determine the long-term rewards that can be obtained by choosing a certain action (i.e. recommending a certain knowledge point) in the current state. By continuously optimizing the Q-value function, the system can learn how to select the best recommended knowledge point under different learning states.

[0182] State Representation: In order to be input into the Q network, the driver's learning state, knowledge point features and historical interaction records need to be embedded and vectorized. The embedding vector can map discrete states and actions into low-dimensional continuous vectors, which is convenient for model processing. The features of embedding vectorization include: Driver characteristics: such as learning progress, scores of completed knowledge points, learning time, etc. Knowledge point features: such as the difficulty of knowledge points, related theoretical knowledge, safety precautions, etc. Historical interaction records: such as the driver's feedback on previously recommended knowledge points and learning performance on different knowledge points.

[0183] For example, the embedding vector representation of driver A may include the scores of the knowledge points he has learned (such as 85 points, 90 points), the learning time (such as 3 hours), etc., while the embedding vector of the knowledge point "emergency braking skills" may include its difficulty coefficient, related safe driving theories, etc.

[0184] Design of Q network: Q network is a multi-layer perceptron (MLP) or convolutional neural network (CNN), which takes state representation as input and outputs the Q value of each knowledge point. In the Top-N recommendation list, each knowledge point has a corresponding Q value, which indicates the long-term return of recommending the knowledge point. The main structure of Q network includes input layer, several hidden layers and output layer. The hidden layer can improve the expression ability of the model through nonlinear activation function.

[0185] In order to improve training efficiency and stability, the system adopts the experience replay mechanism. The specific steps are as follows:

[0186] Experience replay pool: During system operation, state transition samples (i.e., the process of the driver transferring from one learning state to another) will be stored in the experience replay pool. Each sample contains the current state, the selected knowledge point (action), the reward, and the new state (i.e., the state after the driver has completed the knowledge point learning). Randomly extract small batch samples: When training the Q network, the system will randomly extract small batch samples from the experience replay pool for training. This can break the temporal correlation between samples and improve the training effect.

[0187] Target Q network and evaluation Q network: The system adopts a dual Q network architecture, including a target Q network and an evaluation Q network. The target Q network is used to generate target Q values, and the evaluation Q network is used to update model parameters. In order to maintain the stability of the Q network, the parameters of the target Q network are regularly copied from the evaluation Q network.

[0188] For example, driver A transfers from state S1 (having completed "rainy day driving skills" and "overtaking skills") to state S2 (having completed the learning of "emergency braking skills" and scored 85 points). This transfer process will be stored in the experience replay pool for use in Q network training.

[0189] In the process of knowledge point recommendation, the system adopts the ε-greedy strategy to balance the relationship between exploration and utilization:

[0190] Utilization: With a probability of 1-ε, the system selects the knowledge point with the highest Q value for recommendation, that is, it uses the current optimal strategy. Exploration: With a probability of ε, the system randomly selects a knowledge point for recommendation, that is, it conducts exploration to find potential better knowledge points.

[0191] As training progresses, the ε value will gradually decrease, and the system will gradually shift from initial exploration to more utilization. For example, in the early stages of training, ε may be set to 0.5, and the system has a 50% chance of randomly recommending knowledge points; in the later stages of training, ε may drop to 0.1, and the system is more inclined to recommend knowledge points with the highest Q value.

[0192] In order to comprehensively utilize the knowledge point prediction scores generated by collaborative filtering and the Q value output by the deep reinforcement learning algorithm, the system designs a fusion layer. The task of the fusion layer is to combine these two scoring criteria to generate the final comprehensive score for each knowledge point. The specific implementation methods include:

[0193] Weighted summation: The score generated by collaborative filtering and the Q value output by deep reinforcement learning are weighted and summed according to certain weights. The weights can be adjusted according to different scenarios. For example, the weight of the collaborative filtering score can be set to 0.4, and the weight of the Q value can be set to 0.6. Nonlinear transformation: The collaborative filtering prediction score and Q value are transformed by nonlinear functions (such as sigmoid, tanh, etc.) and then fused. Nonlinear transformation can enhance the complexity and expressiveness of the model.

[0194] After the fusion layer generates a comprehensive score for each knowledge point, the system generates a final recommendation list from high to low based on the comprehensive score. For example, at a certain moment, the recommendation list for driver A may be as follows:

[0195] Emergency braking skills (comprehensive score: 0.92); night driving skills (comprehensive score: 0.88); highway driving skills (comprehensive score: 0.85); the system will give priority to recommending knowledge points with the highest comprehensive scores to help drivers conduct personalized learning.

[0196] This technical solution optimizes the personalized training suggestions for drivers by modeling the personalized recommendation problem as a Markov decision process, introducing a deep reinforcement learning algorithm, and combining it with the Top-N recommendation list generated by collaborative filtering. Through reinforcement learning mechanisms such as Q-value function fitting, experience replay, and ε-greedy strategy, the system can learn the optimal knowledge point recommendation strategy. At the same time, by integrating the scores and Q-values ​​of collaborative filtering, the system can generate a comprehensive score to further improve the accuracy and diversity of recommendations.

[0197] Figure 2 FIG. 1 is a structural diagram of a training assistance system based on two-way information feedback in driving training according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0198] The first unit is used to construct a driving scene graph structure, taking the driver, vehicle, and road environment as nodes in the graph, and the interaction between them as edges, and using graph neural networks to model the driving scene; collecting driver operation data, vehicle status data, and environmental perception data, where the driver operation data includes steering wheel angle, accelerator pedal position, and brake pedal force; the vehicle status data includes vehicle speed, acceleration, and yaw angular velocity; and the environmental perception data includes the position of surrounding vehicles and road sign recognition results; inputting the driver operation data, vehicle status data, and environmental perception data into the driving scene graph structure, and obtaining a high-dimensional representation of the driving scene through message passing and node update mechanisms;

[0199] The second unit is used to build a driving rules and skills system based on the knowledge graph, store the knowledge of traffic regulations, safe driving skills, and vehicle control essentials in the form of entities and relationships, and form a structured driving knowledge network; perform sub-graph matching on the high-dimensional representation of the driving scene in the driving knowledge network to find the most relevant knowledge points and skill requirements for the current driving scene; generate personalized training suggestions based on the matching results and the driver's historical performance, including operating skills, safety precautions and related theoretical knowledge;

[0200] The third unit is used to optimize the generated training suggestions based on the learning experience of similar drivers using a collaborative filtering algorithm; deliver the optimized training suggestions to the driver in real time through visual prompts, voice commands or tactile feedback; record the driver's response to the training suggestions, including whether the suggestions are executed, the execution effect and the changes in driving performance after execution; dynamically adjust the parameters of the graph neural network and the knowledge structure of the driving knowledge network according to the response, and analyze and generate suggestions for the next driving scenario based on the updated graph neural network and driving knowledge network, forming a closed-loop feedback mechanism for continuous optimization.

[0201] According to a third aspect of the embodiments of the present invention,

[0202] An electronic device is provided, comprising:

[0203] processor;

[0204] a memory for storing processor-executable instructions;

[0205] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0206] A fourth aspect of the embodiments of the present invention is:

[0207] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0208] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0209] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. Training assistance method based on two-way information feedback in driving training, It is characterized in that include: Construct a driving scene graph structure, taking the driver, vehicle, and road environment as nodes in the graph, and the interactions between them as edges, and use graph neural networks to model the driving scene; Collect driver operation data, vehicle status data and environmental perception data, where driver operation data includes steering wheel angle, accelerator pedal position, brake pedal force, vehicle status data includes vehicle speed, acceleration, yaw angular velocity, and environmental perception data includes surrounding vehicle positions and road sign recognition results; input driver operation data, vehicle status data and environmental perception data into the driving scene graph structure, and obtain a high-dimensional representation of the driving scene through message passing and node update mechanism; Build a driving rules and skills system based on the knowledge graph, store the knowledge of traffic laws, safe driving skills, and vehicle control essentials in the form of entities and relationships, and form a structured driving knowledge network; The high-dimensional representation of the driving scene is matched with subgraphs in the driving knowledge network to find the knowledge points and skill requirements most relevant to the current driving scene. Based on the matching results and the driver's historical performance, personalized training suggestions are generated, including operating skills, safety precautions and related theoretical knowledge. A collaborative filtering algorithm is used to optimize the generated training suggestions based on the learning experience of similar drivers; Deliver optimized training suggestions to drivers in real time through visual prompts, voice commands or tactile feedback; Record the driver's response to the training suggestions, including whether the suggestions were implemented, the implementation effect, and the changes in driving performance after implementation; dynamically adjust the parameters of the graph neural network and the knowledge structure of the driving knowledge network according to the response situation, and analyze and generate suggestions for the next driving scenario based on the updated graph neural network and driving knowledge network, forming a closed-loop feedback mechanism for continuous optimization; Build a driving rules and skills system based on the knowledge graph, store the knowledge of traffic laws, safe driving skills, and vehicle control essentials in the form of entities and relationships, and form a structured driving knowledge network; The high-dimensional representation of the driving scene is matched with subgraphs in the driving knowledge network to find out the most relevant knowledge points and skill requirements for the current driving scene, including: From the text materials of traffic regulations, driving teaching materials and safety manuals, named entity recognition and relationship extraction technology are used to extract driving-related entities and relationships. The entity types include driving behavior, driving scene elements, driving risks and driving skills, and the relationship types include attribute relationships, constraint relationships, causal relationships and operational relationships; the extracted entities and relationships are represented in the form of triples; knowledge from different sources is deduplicated and integrated, and a globally consistent driving knowledge graph is constructed using knowledge alignment technology; the driving knowledge graph is stored in a graph database; Using a graph neural network to encode multimodal perception data of a driving scene into a scene semantic vector; extracting key entities and their neighbor nodes related to the driving scene according to the ontology definition of the driving knowledge graph to form a scene-related knowledge subgraph; calculating the semantic similarity between the scene semantic vector and each node in the knowledge subgraph through a graph neural network to find the Top-K knowledge nodes most relevant to the driving scene; Taking the Top-K nodes as the starting point, a breadth-first search algorithm is used to traverse the knowledge subgraph to extract the shortest relationship path between nodes, where the shortest relationship path reflects the reasoning chain between the driving scenario and the knowledge nodes; based on the shortest relationship path obtained by reasoning, a natural language description is summarized to find the knowledge points and skill requirements most relevant to the current driving scenario; Taking the Top-K nodes as the starting point, the knowledge subgraph is traversed using a breadth-first search algorithm to extract the shortest relationship path between nodes, including: Taking the top-K knowledge nodes most relevant to the driving scenario as the center, extract the N-hop neighbor nodes of the top-K knowledge nodes in the driving knowledge graph to construct a knowledge subgraph containing more background knowledge, and the number of hops N of the N-hop neighbor nodes is adjusted according to the scale and density of the knowledge graph; A breadth-first search algorithm is used to traverse the knowledge subgraph to explore all relationship paths between the Top-K knowledge nodes. The breadth-first search algorithm starts from the Top-K knowledge nodes, traverses the neighbor nodes of the Top-K knowledge nodes layer by layer, records the predecessor nodes and corresponding relationship edges of each knowledge node, and constructs a search tree of relationship paths; In the search tree of the relationship path, the shortest path between the Top-K knowledge nodes is extracted as the reasoning chain, and the conditions of the shortest path include: the starting point and the end point of the path are both the Top-K knowledge nodes, the path does not contain repeated nodes, and the length of the path is the smallest; the shortest path is solved by a dynamic programming algorithm, the shortest distance between the knowledge nodes is updated through the state transfer equation, and the path is backtracked to obtain the node sequence on the shortest path; Representing the shortest path as a sequence of relationship triples, wherein the relationship triples include two nodes and a relationship type connecting the two nodes, wherein the relationship type is a predefined relationship or a relationship customized according to a specific scenario; In the case where there are multiple shortest paths between the Top-K knowledge nodes, a path fusion algorithm is used to merge the multiple shortest paths into a comprehensive path. The steps of the path fusion algorithm include: counting the frequency of nodes and relationships appearing in all shortest paths, sorting them from high to low by frequency, adding them to the comprehensive path in sequence, and deduplicating and compressing the comprehensive path.

2. The method according to claim 1, It is characterized in that The driver operation data, vehicle status data and environment perception data are input into the driving scene graph structure. Through message passing and node update mechanism, a high-dimensional representation of the driving scene is obtained, including: The collected driver operation data, vehicle status data and environment perception data are aligned according to the timestamps, and the timestamp of the vehicle status data is used as a reference to find the closest time between the timestamp of each data in the driver operation data and the environment perception data and the timestamp of the vehicle status data, and the data is matched to the closest time. For a time that does not have corresponding driver operation data or environment perception data, the data value at the time is estimated by using the interpolation method, so as to obtain multi-dimensional time series data with a unified time granularity; Mapping the multidimensional time series data into a driving scene graph structure, representing the driver, vehicle, road and environment as four types of nodes of the graph, respectively, and the initial attributes of each type of node include the static attributes of the corresponding element; At each time step, the attributes of the driver node are updated to the combination of the current driver operation data and the historical driver operation data statistics, the attributes of the vehicle node are updated to the combination of the current vehicle speed, acceleration, yaw rate, position and heading angle and the historical vehicle state statistics, the attributes of the road node and the environment node are updated to the combination of the road curvature, slope, road slipperiness and the relative position and speed of the surrounding vehicles and obstacles, the attributes of the edge between the driver node and the vehicle node are set to the value of the accelerator and brake operation signals and the change rate of the operation signals, and the attributes of the edge between the vehicle node and the road node are set to the influence coefficient of the road curvature and slope on the vehicle motion and the adhesion coefficient between the tire and the road; A graph convolutional neural network is used to extract the feature representation of each node. Each layer of the graph convolutional neural network obtains the node feature matrix of the next layer of the network by multiplying the node feature matrix of the current layer of the network with the normalized adjacency matrix of the graph and the weight matrix of the current layer and processing the network through an activation function. The input layer of the graph convolutional neural network is the node attributes and edge attributes. By stacking multiple layers of graph convolutional neural networks, the neighborhood information of each node is integrated to obtain the feature vector of each node. The attention mechanism is used to perform weighted summation on the feature vectors of the nodes output by the graph convolutional neural network to obtain a high-dimensional representation vector of the initial driving scene. The attention mechanism obtains the attention weight of each node by calculating the inner product of the feature vector of each node and the learnable query vector and performing exponential function and normalization processing. The feature vectors of all nodes are weighted summed based on the attention weight to obtain a scene representation vector. Finally, the scene representation vectors obtained at different time steps are merged into a time-series scene representation sequence, which is input into a recurrent neural network or a one-dimensional convolutional network to extract the evolution characteristics of the scene in the time dimension and output the final high-dimensional representation of the driving scene.

3. The method according to claim 1, It is characterized in that Generate personalized training suggestions based on the matching results and the driver's historical performance, including operating skills, safety precautions and relevant theoretical knowledge; Using collaborative filtering algorithms, the generated training suggestions are optimized based on the learning experience of similar drivers, including: Collect drivers' historical learning data, including completed knowledge points, test scores, and learning time, to form a driver-knowledge point rating matrix; use the Pearson correlation coefficient or cosine similarity measurement method to calculate the similarity between drivers based on their ratings of knowledge points; for the target driver, find the K neighbor drivers who are most similar to the target driver, use the historical ratings of the K neighbor drivers to predict the target driver's rating for each knowledge point, and generate a collaborative filtering Top-N knowledge point recommendation list from high to low according to the predicted ratings.

4. The method according to claim 3, It is characterized in that The method further comprises: Based on the Top-N knowledge point recommendation list generated by collaborative filtering, the personalized recommendation problem is modeled as a Markov decision process, where the state is the driver's current learning state and historical interaction records, the action is each knowledge point in the Top-N knowledge point recommendation list, and the reward is the driver's immediate feedback on the recommended knowledge point, the long-term impact evaluation of the recommendation on the driver's overall learning effect, and the diversity of the recommended knowledge points; The deep reinforcement learning algorithm is used to fit the action-state value through the Q-value function to learn the optimal knowledge point recommendation strategy; the driver characteristics, knowledge point characteristics, and historical interactions are embedded and vectorized and combined with the current time step information to form a state representation. A multi-layer perceptron or convolutional neural network is designed as the main structure of the Q network, which takes the state representation as input and outputs the Q value of each knowledge point in the Top-N recommendation list; Adopting the experience replay mechanism, the state transition samples are stored in the replay pool, and small batch samples are randomly selected to train the Q network; at the same time, the target Q network and the evaluation Q network are constructed. The target Q network is used to generate the Q value target, and the parameters are regularly copied from the evaluation Q network; When generating recommendations, the ε-greedy strategy is used to select actions, selecting the knowledge point with the highest Q value with a probability of 1-ε, and randomly exploring new knowledge points with a probability of ε, where ε decreases with the number of training rounds; A fusion layer is designed, which takes the predicted scores of knowledge points generated by collaborative filtering and the Q values ​​of knowledge points output by the deep reinforcement learning algorithm as input, and generates a comprehensive score for each knowledge point through weighted summation or nonlinear transformation.

5. A training assistance system based on two-way information feedback in driving training, used to implement the method according to any one of claims 1 to 4, It is characterized in that include: The first unit is used to construct a driving scene graph structure, taking the driver, vehicle, and road environment as nodes in the graph, and the interaction between them as edges, and using graph neural networks to model the driving scene; Collect driver operation data, vehicle status data and environmental perception data, where driver operation data includes steering wheel angle, accelerator pedal position, brake pedal force, vehicle status data includes vehicle speed, acceleration, yaw angular velocity, and environmental perception data includes surrounding vehicle positions and road sign recognition results; input driver operation data, vehicle status data and environmental perception data into the driving scene graph structure, and obtain a high-dimensional representation of the driving scene through message passing and node update mechanism; The second unit is used to build a driving rules and skills system based on the knowledge graph, storing the knowledge of traffic laws, safe driving skills, and vehicle control essentials in the form of entities and relationships to form a structured driving knowledge network; The high-dimensional representation of the driving scene is matched with subgraphs in the driving knowledge network to find the knowledge points and skill requirements most relevant to the current driving scene. Based on the matching results and the driver's historical performance, personalized training suggestions are generated, including operating skills, safety precautions and related theoretical knowledge. The third unit is used to optimize the generated training suggestions based on the learning experience of similar drivers using a collaborative filtering algorithm; Deliver optimized training suggestions to drivers in real time through visual prompts, voice commands or tactile feedback; Record the driver's response to training suggestions, including whether the suggestions are implemented, the implementation effect, and the changes in driving performance after implementation; dynamically adjust the parameters of the graph neural network and the knowledge structure of the driving knowledge network according to the response, and analyze and generate suggestions for the next driving scenario based on the updated graph neural network and driving knowledge network, forming a closed-loop feedback mechanism for continuous optimization.

6. An electronic device, It is characterized in that include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 4.

7. A computer-readable storage medium having computer program instructions stored thereon, It is characterized in that When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 4 is implemented.