Multi-vehicle interactive driving decision-making and driving method and system based on knowledge driving

Through a knowledge-driven multi-vehicle interactive driving decision-making method, combined with neural relationship reasoning model and inverse reinforcement learning, the problem of flexibility and transparency of autonomous driving systems in complex traffic scenarios is solved, and safer and more efficient driving decisions are achieved.

CN118953403APending Publication Date: 2024-11-15CHINA MERCHANTS CHONGQING COMM RES & DESIGN INST
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411019990.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-29
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

In existing autonomous driving technology, rule-driven methods lack flexibility in complex traffic scenarios. Deep learning models have reduced prediction accuracy and opaque decision logic under sufficient training, which cannot meet the safety and interpretability requirements of autonomous driving systems.

Method used

Adopt knowledge-driven multi-vehicle interactive driving decision-making method, by obtaining vehicle key feature data, combining domain expert knowledge to train neural relationship reasoning models, using inverse reinforcement learning and Markov decision-making process to select the optimal driving strategy, and integrate data-driven and knowledge Driven decision logic.

Benefits of technology

Improves the adaptability and transparency of decision-making in autonomous driving systems, can respond in real time in complex traffic situations and provide explainable decision-making basis to achieve safer and more efficient driving.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118953403A_ABST
    Figure CN118953403A_ABST
Patent Text Reader

Abstract

The invention discloses a multi-vehicle interactive driving decision-making and driving method and system based on knowledge driving. Comprising the following steps: acquiring key feature data corresponding to multiple groups of different vehicles to construct a data set, and training a neural relationship reasoning model in combination with the data set and domain expert knowledge; deducing the inter-vehicle correlation and potential decision points corresponding to each sample in the data set through a trained neural relationship reasoning model; according to inference results corresponding to all the samples, inferring a reward function by adopting an inverse reinforcement learning method; obtaining the current key feature data of the target vehicle, and deducing the inter-vehicle correlation and potential decision points corresponding to the target vehicle through the trained neural relationship reasoning model; and selecting an optimal driving strategy through a Markov decision process based on the deduced reward function according to the deduced result corresponding to the target vehicle, thereby improving the adaptability and operation transparency of the automatic driving system.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of drive control systems particularly suitable for autonomous road vehicles, and in particular to a knowledge-driven multi-vehicle interactive driving decision, driving method and system. Background Art

[0002] Advances in autonomous driving technology have enabled the development of a variety of decision-making systems, which are mainly based on rule-driven and data-driven approaches. Rule-driven approaches directly apply traffic rules and behavioral guidelines through fixed logic and algorithms, and are suitable for handling standard traffic conditions. Data-driven approaches, especially deep learning models, learn to recognize complex traffic patterns by analyzing large amounts of data collected from on-board sensors such as radars and cameras.

[0003] However, rule-driven methods follow fixed procedures and instructions, which makes them generally unable to adapt effectively and lack flexibility when encountering complex traffic scenarios that are not preset or change rapidly. In addition, rule-driven systems may not be able to make accurate judgments when it is necessary to interpret ambiguous or conflicting traffic signs.

[0004] Although data-driven models, especially deep learning technology, perform well in pattern recognition, their prediction accuracy is heavily dependent on the coverage and quality of training data. When encountering situations that are not fully represented in the training set, the performance of the model may drop rapidly. In addition, the internal decision logic of the deep learning model is not transparent to users and developers, which is particularly critical in the field of autonomous driving, because safety requirements and potential legal liabilities require that the decision-making process must be explainable and verifiable. Summary of the invention

[0005] In view of the shortcomings of the existing technology, the present invention proposes a knowledge-driven multi-vehicle interactive driving decision, driving method and system, which can improve the flexibility and explainability of driving decisions. The specific technical solution is as follows:

[0006] In a first aspect, a knowledge-driven multi-vehicle interactive driving decision-making method is provided. In a first implementable manner of the first aspect, the method includes:

[0007] Obtain key feature data corresponding to multiple groups of different vehicles to build a data set, and combine the data set with domain expert knowledge to train the neural relational reasoning model;

[0008] Inferring the inter-vehicle relationships and potential decision points corresponding to each sample in the data set through a trained neural relational inference model;

[0009] According to the inference results corresponding to all samples, the reward function is inferred using the inverse reinforcement learning method;

[0010] Obtain the current key feature data of the target vehicle, and infer the inter-vehicle relationships and potential decision points corresponding to the target vehicle through the trained neural relational reasoning model;

[0011] Based on the inferred reward function, an optimal driving strategy is selected through a Markov decision process according to the inference result corresponding to the target vehicle.

[0012] In combination with the first implementable manner of the first aspect, in a second implementable manner of the first aspect, obtaining key feature data corresponding to the vehicle includes:

[0013] Acquiring original data corresponding to the vehicle, the original data including environmental data and vehicle status data;

[0014] Performing data cleaning on the original data, and performing data synchronization on the cleaned original data;

[0015] Feature extraction is performed on the synchronized raw data to obtain key feature data corresponding to the vehicle.

[0016] In combination with the first implementable manner of the first aspect, in a third implementable manner of the first aspect, training the neural relational reasoning model through the data set includes:

[0017] Maximizing the evidence lower bound optimization is adopted as the objective function for training the neural relational reasoning model.

[0018] In combination with the first implementable manner of the first aspect, in a fourth implementable manner of the first aspect, the reward function includes a safety reward component, an efficiency reward component, and a comfort reward component.

[0019] In the second aspect, a knowledge-driven multi-vehicle interactive driving method is provided, including:

[0020] Selecting an optimal driving strategy by using the knowledge-driven multi-vehicle interactive driving decision-making method as described in any one of the first to fourth possible implementations of the first aspect;

[0021] Generate corresponding vehicle control instructions based on the selected optimal driving strategy, and control the vehicle to perform corresponding operations according to the vehicle control instructions.

[0022] In a third aspect, a knowledge-driven multi-vehicle interactive driving decision system is provided. In a first implementable manner of the third aspect, the system includes:

[0023] The inference model training module is configured to obtain key feature data corresponding to multiple groups of different vehicles to build a data set, and train the neural relational inference model by combining the data set and domain expert knowledge;

[0024] A sample relationship inference module, configured to infer the inter-vehicle relationships and potential decision points corresponding to each sample in the data set through a trained neural relationship inference model;

[0025] A function inference module, configured to infer a reward function using an inverse reinforcement learning method based on the inference results corresponding to all samples;

[0026] A target relationship inference module, configured to obtain the current key feature data of the target vehicle and infer the inter-vehicle relationships and potential decision points corresponding to the target vehicle through the trained neural relationship inference model;

[0027] The driving strategy selection module is configured to select the optimal driving strategy through a Markov decision process based on the inferred reward function and the inference result corresponding to the target vehicle.

[0028] In combination with the first implementable manner of the third aspect, in a second implementable manner of the third aspect, the model training module includes:

[0029] A data acquisition unit, configured to acquire raw data corresponding to the vehicle, the raw data including environmental data and vehicle status data;

[0030] A data processing unit, configured to clean the raw data and synchronize the cleaned raw data;

[0031] The feature extraction unit is configured to perform feature extraction on the synchronized raw data to obtain key feature data corresponding to the vehicle.

[0032] In combination with the first implementable manner of the third aspect, in a third implementable manner of the third aspect, the model training module adopts maximizing the evidence lower bound optimization as the objective function of the neural relational reasoning model training.

[0033] In combination with the first implementable manner of the third aspect, in a fourth implementable manner of the third aspect, the reward function includes a safety reward component, an efficiency reward component, and a comfort reward component.

[0034] Fourthly, a knowledge-driven intelligent connected autonomous driving system is provided, including:

[0035] A decision module, configured to select an optimal driving strategy by using the knowledge-driven multi-vehicle interactive driving decision method as described in any one of the first to fourth implementation methods of the first aspect;

[0036] The execution module is configured to generate corresponding vehicle control instructions based on the selected optimal driving strategy, and control the vehicle to perform corresponding operations according to the vehicle control instructions.

[0037] Beneficial effects: The knowledge-driven multi-vehicle interactive driving decision, driving method and system of the present invention can automatically learn the dynamic interaction mode between vehicles from data through the neural relationship reasoning model, and integrate the knowledge of field experts into the learning process, providing a clear logical basis for the decision-making process, thereby significantly improving the adaptability of the autonomous driving system and the transparency of operation. Through the inverse reinforcement learning method, the decision rules can be refined and optimized, so that the autonomous driving system can not only respond based on real-time data when facing complex traffic conditions, but also perform logical reasoning according to the current environment, and flexibly provide explainable decision-making basis, thereby achieving safer and more efficient driving decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the specific implementation of the present invention, the following will briefly introduce the drawings required for use in the specific implementation. In all the drawings, each element or part is not necessarily drawn according to the actual scale.

[0039] Figure 1 A flow chart of a knowledge-driven multi-vehicle interactive driving decision-making method provided by an embodiment of the present invention;

[0040] Figure 2 A flow chart of a multi-vehicle interactive driving method based on knowledge-driven according to an embodiment of the present invention;

[0041] Figure 3 A system block diagram of a knowledge-driven multi-vehicle interactive driving decision system provided by an embodiment of the present invention;

[0042] Figure 4 A system block diagram of a knowledge-driven intelligent connected autonomous driving system provided in accordance with an embodiment of the present invention. DETAILED DESCRIPTION

[0043] The following embodiments of the technical solution of the present invention are described in detail in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and are therefore only used as examples, and cannot be used to limit the protection scope of the present invention.

[0044] like Figure 1 The flowchart of the knowledge-driven multi-vehicle interactive driving decision-making method shown in FIG. 1 includes:

[0045] Step 1: Obtain key feature data corresponding to multiple groups of different vehicles to build a data set, and train the neural relational reasoning model based on the data set and domain expert knowledge;

[0046] Step 2: Inferring the inter-vehicle relationships and potential decision points corresponding to each sample in the data set through the trained neural relational inference model;

[0047] Step 3: Based on the inference results corresponding to all samples, the reward function is inferred using the inverse reinforcement learning method;

[0048] Step 4: Obtain the current key feature data of the target vehicle, and infer the inter-vehicle relationships and potential decision points corresponding to the target vehicle through the trained neural relational reasoning model;

[0049] Step 5: Based on the inferred reward function and the inference result corresponding to the target vehicle, an optimal driving strategy is selected through a Markov decision process.

[0050] Specifically, first, multiple sets of key feature data can be obtained to build a data set, and each set of key feature data corresponds to a different vehicle. The data set can be used to train the neural relational reasoning model to obtain the interactive relationship diagram between vehicles. Then, combined with the knowledge of domain experts, the interactive relationship diagram can be further understood, explained, and made transparent, and a neural relational reasoning model that can be used to infer the dynamic workshop relationships and potential decision points between vehicles can be obtained. In this way, human intuitive understanding and domain rules can be integrated into the data-driven model, improving the transparency and explainability of decision-making.

[0051] Then, the inter-vehicle relationships and potential decision points of each sample in the data set can be inferred through the trained neural relational reasoning model. After that, based on the inter-vehicle relationships and potential decision points corresponding to each sample, the inverse reinforcement learning method is used to infer the optimal reward function to refine and optimize the decision rules. Then, the current key feature data of the target vehicle can be obtained, and the key feature data can be input into the trained neural relational reasoning model based on domain expert knowledge to infer the inter-vehicle relationships and potential decision points between the current target vehicle and the surrounding vehicles, thereby accurately analyzing the complex dynamic relationship between vehicles and improving the accuracy of the overall decision.

[0052] Finally, based on the inferred reward function, the optimal driving strategy can be selected through the Markov decision process according to the vehicle-to-vehicle relationships and potential decision points corresponding to the target vehicle, so as to consider future states in real-time decision-making, optimize long-term behavior rather than just short-term benefits, and make decisions more comprehensive and forward-looking.

[0053] In this embodiment, optionally, in step 1, obtaining key feature data corresponding to the vehicle includes:

[0054] Acquiring original data corresponding to the vehicle, the original data including environmental data and vehicle status data;

[0055] Performing data cleaning on the original data, and performing data synchronization on the cleaned original data;

[0056] Feature extraction is performed on the synchronized raw data to obtain key feature data corresponding to the vehicle.

[0057] Specifically, first, the original data corresponding to the vehicle can be obtained, which includes the environmental data around the vehicle and the vehicle status data. The environmental data includes but is not limited to the vehicle position, the distance and speed between nearby objects such as other vehicles, pedestrians, and obstacles, etc. The vehicle status data includes but is not limited to speed, acceleration, steering wheel angle, brake status, etc.

[0058] Then, the acquired raw data can be cleaned to remove invalid, erroneous or incomplete data records, and correct outliers and deviations in the data. After the raw data is cleaned, the data from different data sources can be aligned according to time, so as to accurately associate and analyze the relative dynamics between vehicles later. It should be understood that this embodiment adopts the existing data cleaning method, and its specific method steps belong to the conventional technical means in this field, which will not be repeated here.

[0059] Finally, the key feature data of the vehicle can be extracted from the synchronized raw data. The key feature data may include but are not limited to: road type, traffic density, road type, traffic density, vehicle position, speed, acceleration, vehicle spacing, vehicle behavior patterns, etc.

[0060] In this embodiment, optionally, in step 1, training the neural relational reasoning model using the data set includes:

[0061] Maximizing the evidence lower bound optimization is adopted as the objective function for training the neural relational reasoning model.

[0062] Specifically, the neural relational reasoning model includes an encoder and a decoder, where the encoder can be designed as a graph neural network to infer the potential probability distribution of vehicle-to-vehicle interactions as shown below:

[0063]

[0064] Among them, q φ (z ij x) is the conditional probability distribution, indicating that given the vehicle state x, there is a certain interaction z between vehicle i and vehicle j. ij The probability of φ is the encoder parameter.

[0065] The decoder can also be designed as a graph neural network to simulate the dynamic system of the vehicle, specifically:

[0066]

[0067] in, Represents the state x at a given previous time step t-1 and interaction graph z, the state of vehicle i at time t The conditional probability distribution of .

[0068] During the training process, the objective function of the neural relational reasoning model training is to maximize the evidence lower bound optimization, which mainly includes two items, one of which is the reconstruction loss, which aims to maximize the log-likelihood of the state x under a given interaction graph. The other is the Kullback-Leibler divergence, which is used to measure the difference between the inferred potential probability distribution and the prior distribution p(z), specifically:

[0069]

[0070] Among them, θ is the decoder parameter.

[0071] By inputting the key feature data corresponding to the vehicle into the trained neural relational reasoning model, we can obtain the vehicle's corresponding workshop relationships, that is, the vehicle's status, and potential decision points, that is, the actions that the vehicle may take.

[0072] The encoder can be used to interpret vehicle behavior trajectories and infer a potential interaction graph, where nodes represent vehicles and edges represent vehicle-to-vehicle relationships inferred based on dynamic interactions. The decoder uses the structured information in the interaction graph to simulate or predict future vehicle trajectories. Through this process, the model's understanding of the interaction dynamics is checked to be sufficient to accurately predict actual vehicle motion, and expert knowledge is used to supplement and correct it.

[0073] In this embodiment, optionally, in step 3, the reward function includes a safety reward component, an efficiency reward component and a comfort reward component.

[0074] Specifically, the reward function can be structurally optimized based on safety rewards, efficiency rewards, and comfort rewards. Among them, safety rewards include but are not limited to complying with traffic rules such as traffic signs and traffic lights, maintaining a safe following distance, reasonably controlling speed, driving behavior standards, and having emergency response capabilities. Efficiency rewards include but are not limited to path optimization, speed optimization, traffic coordination, etc. Comfort rewards include but are not limited to driving stability, noise and vibration control, etc.

[0075] Combined with the above domain expert knowledge, the reward function can be specifically set as:

[0076] R(x,a)=ω1·R safe (x, a)+ω2·R eff (x, a)+ω3·R comfort (x, a);

[0077] Among them, a is the action taken, R safe is the security bonus component, R eff is the efficiency reward component, R comfort is the comfort reward component, ω1, ω2, and ω3 are the weights corresponding to the safety reward, efficiency reward, and comfort reward, respectively.

[0078] The state of the sample vehicle and the possible actions can be determined based on the inference results corresponding to the sample, and then the inverse reinforcement learning method is used to optimize the above reward function to obtain the optimal reward function for subsequent decision-making selection, which is:

[0079]

[0080] Among them, π ω is the optimal strategy based on ω1, ω2, and ω3, π is the probability of taking each possible action a when a state x is observed, H is the entropy of the strategy, which is used to encourage exploration, and λ is the weight of the regularization term, which is used to adjust the importance of the entropy term in the optimization process.

[0081] After the reward function is inferred, the reward function can be applied to the Markov decision process. The state of the target vehicle is determined according to the relationship between the vehicles corresponding to the target vehicle. According to the state and potential decision points corresponding to the target vehicle, the value of each state of the target vehicle is iteratively updated to select the optimal driving strategy for the target vehicle. The specific calculation formula of the value function is as follows:

[0082]

[0083] Among them, V new (x) is the value of state x after the current iteration, max a represents the action a that maximizes the expression, γ is the discount factor, and represents the present value of future rewards. P(x′x, a) is the probability that the target vehicle moves from state x to state x′ through action a, V old (x′) is the value of state x′ in the previous iteration.

[0084] This iteration is continued until the value function meets the convergence condition. In this embodiment, the convergence condition can be set to V new (x) and V old The difference of (x) is less than the preset threshold. When the convergence condition is met, the optimal driving strategy can be extracted from the value function.

[0085] like Figure 2 The flowchart of the knowledge-driven multi-vehicle interactive driving method shown in FIG. 1 includes:

[0086] Step S1, selecting the optimal driving strategy using the above-mentioned knowledge-driven multi-vehicle interactive driving decision-making method;

[0087] Step S2: Generate corresponding vehicle control instructions based on the selected optimal driving strategy, and control the vehicle to perform corresponding operations according to the vehicle control instructions.

[0088] Specifically, first, the knowledge-driven multi-vehicle interactive driving decision-making method provided in this embodiment can be used to select the optimal driving strategy. And an optimal action a is specified for the current state of the target vehicle through the optimal driving strategy. * , specifically:

[0089]

[0090] In this way, when faced with complex traffic conditions, the autonomous driving system can not only respond based on real-time data, but also perform logical reasoning based on domain knowledge and the current environment, and flexibly provide explainable decision-making basis, thereby achieving safer and more efficient driving decisions.

[0091] Then, corresponding vehicle control instructions may be generated based on the selected optimal action, such as adjusting the throttle instruction, the brake instruction or the steering wheel instruction, etc. The driving system may send the generated vehicle control instructions to the corresponding execution device, and the execution device may control the vehicle to perform the corresponding operation according to the vehicle control instructions.

[0092] like Figure 3 The multi-vehicle interactive driving decision system shown in the figure is based on knowledge-driven, and the decision system includes:

[0093] The inference model training module is configured to obtain key feature data corresponding to multiple groups of different vehicles to build a data set, and train the neural relational inference model by combining the data set and domain expert knowledge;

[0094] A sample relationship inference module, configured to infer the inter-vehicle relationships and potential decision points corresponding to each sample in the data set through a trained neural relationship inference model;

[0095] A function inference module, configured to infer a reward function using an inverse reinforcement learning method based on the inference results corresponding to all samples;

[0096] A target relationship inference module, configured to obtain the current key feature data of the target vehicle and infer the inter-vehicle relationships and potential decision points corresponding to the target vehicle through the trained neural relationship inference model;

[0097] The driving strategy selection module is configured to select the optimal driving strategy through a Markov decision process based on the inferred reward function and the inference result corresponding to the target vehicle.

[0098] Specifically, the decision-making system consists of an inference model training module, a sample relationship inference module, a function inference module, a target relationship inference module, and a driving strategy selection module. Among them, the inference model training module can obtain multiple sets of key feature data to build a data set, and each set of key feature data corresponds to a different vehicle. The neural relationship inference model can be trained through the data set, and the interaction relationship diagram obtained through training can be further understood, explained, and made transparent in combination with the knowledge of domain experts, so as to obtain a neural relationship inference model that can be used to infer the dynamic workshop relationships and potential decision points between vehicles.

[0099] The sample relationship inference module can infer the workshop relationships and potential decision points corresponding to each sample vehicle through the trained neural relationship inference model based on the key feature data corresponding to each sample in the data set. The function inference module can infer the optimal reward function based on the workshop relationships and potential decision points corresponding to each sample and combined with the domain expert knowledge using the inverse reinforcement learning method, thereby integrating human intuitive understanding and domain rules into the data-driven model, thereby improving the transparency and explainability of decision-making.

[0100] The target relationship inference module can obtain the current key feature data of the target vehicle and input the key feature data into the trained neural relationship inference model to infer the vehicle-to-vehicle relationship and potential decision points between the current target vehicle and surrounding vehicles. This provides a clear logical basis for the subsequent decision-making process and improves the adaptability of the autonomous driving system and the transparency of operation.

[0101] The driving strategy selection module can select the optimal driving strategy through the Markov decision process based on the inferred reward function, according to the vehicle-to-vehicle relationships and potential decision points corresponding to the target vehicle, so as to consider future states in real-time decision-making, optimize long-term behavior rather than just short-term benefits, and make decisions more comprehensive and forward-looking.

[0102] In this embodiment, optionally, the model training module includes:

[0103] A data acquisition unit, configured to acquire raw data corresponding to the vehicle, the raw data including environmental data and vehicle status data;

[0104] A data processing unit, configured to clean the raw data and synchronize the cleaned raw data;

[0105] The feature extraction unit is configured to perform feature extraction on the synchronized raw data to obtain key feature data corresponding to the vehicle.

[0106] Specifically, the model training module includes a data acquisition unit, a data processing unit and a feature extraction unit. The data acquisition unit can acquire original data corresponding to the vehicle, which includes environmental data around the vehicle and vehicle status data.

[0107] The acquired raw data can then be cleaned to remove invalid, erroneous or incomplete data records and correct outliers and deviations in the data. After the raw data is cleaned, the data from different data sources can be aligned in time so that the relative dynamics between vehicles can be accurately correlated and analyzed later.

[0108] In this embodiment, optionally, the model training module adopts maximizing the evidence lower bound optimization as the objective function of the neural relational reasoning model training.

[0109] In this embodiment, optionally, the reward function includes a safety reward component, an efficiency reward component and a comfort reward component.

[0110] like Figure 4 The system block diagram of the knowledge-driven intelligent connected autonomous driving system shown in FIG. 1 includes:

[0111] A decision module, configured to select an optimal driving strategy using the above-mentioned knowledge-driven multi-vehicle interactive driving decision method;

[0112] The execution module is configured to generate corresponding vehicle control instructions based on the selected optimal driving strategy, and control the vehicle to perform corresponding operations according to the vehicle control instructions.

[0113] Specifically, the driving system is composed of a decision module and an execution module. Among them, the decision module can firstly select the optimal driving strategy by using the knowledge-driven multi-vehicle interactive driving decision method provided in this embodiment. And specify an optimal action for the current state of the target vehicle through the optimal driving strategy. The execution module can generate corresponding vehicle control instructions based on the selected optimal action, such as adjusting the throttle instruction, the brake instruction or the steering wheel instruction. The driving system can send the generated vehicle control instruction to the corresponding execution device, and the execution device can control the vehicle to perform the corresponding operation according to the vehicle control instruction.

[0114] 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 by equivalents. These modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and specification of the present invention.

Claims

1. A knowledge-driven multi-vehicle interactive driving decision-making method, characterized in that: include: Obtain key feature data corresponding to multiple groups of different vehicles to build a data set, and combine the data set with domain expert knowledge to train the neural relational reasoning model; Inferring the inter-vehicle relationships and potential decision points corresponding to each sample in the data set through a trained neural relational inference model; According to the inference results corresponding to all samples, the reward function is inferred using the inverse reinforcement learning method; Obtain the current key feature data of the target vehicle, and infer the inter-vehicle relationships and potential decision points corresponding to the target vehicle through the trained neural relational reasoning model; Based on the inferred reward function, an optimal driving strategy is selected through a Markov decision process according to the inference result corresponding to the target vehicle.

2. The knowledge-driven multi-vehicle interactive driving decision-making method according to claim 1 is characterized in that: Obtain key feature data corresponding to the vehicle, including: Acquiring original data corresponding to the vehicle, the original data including environmental data and vehicle status data; Performing data cleaning on the original data, and performing data synchronization on the cleaned original data; Feature extraction is performed on the synchronized raw data to obtain key feature data corresponding to the vehicle.

3. The knowledge-driven multi-vehicle interactive driving decision-making method according to claim 1 is characterized in that: The neural relational reasoning model is trained with datasets including: Maximizing the evidence lower bound optimization is adopted as the objective function for training the neural relational reasoning model.

4. The knowledge-driven multi-vehicle interactive driving decision-making method according to claim 1 is characterized by: The reward function includes a safety reward component, an efficiency reward component and a comfort reward component.

5. A knowledge-driven multi-vehicle interactive driving method, characterized in that: include: Selecting the optimal driving strategy by using the knowledge-driven multi-vehicle interactive driving decision-making method as described in any one of claims 1 to 4; Generate corresponding vehicle control instructions based on the selected optimal driving strategy, and control the vehicle to perform corresponding operations according to the vehicle control instructions.

6. A knowledge-driven multi-vehicle interactive driving decision system, characterized in that: include: The inference model training module is configured to obtain key feature data corresponding to multiple groups of different vehicles to build a data set, and train the neural relational inference model by combining the data set and domain expert knowledge; A sample relationship inference module, configured to infer the inter-vehicle relationships and potential decision points corresponding to each sample in the data set through a trained neural relationship inference model; A function inference module, configured to infer a reward function using an inverse reinforcement learning method based on the inference results corresponding to all samples; A target relationship inference module, configured to obtain the current key feature data of the target vehicle and infer the inter-vehicle relationships and potential decision points corresponding to the target vehicle through the trained neural relationship inference model; The driving strategy selection module is configured to select the optimal driving strategy through a Markov decision process based on the inferred reward function and the inference result corresponding to the target vehicle.

7. The knowledge-driven multi-vehicle interactive driving decision system according to claim 6 is characterized in that: The model training module includes: A data acquisition unit, configured to acquire raw data corresponding to the vehicle, the raw data including environmental data and key feature data; A data processing unit, configured to clean the raw data and synchronize the cleaned raw data; The feature extraction unit is configured to perform feature extraction on the synchronized raw data to obtain key feature data corresponding to the vehicle.

8. The knowledge-driven multi-vehicle interactive driving decision system according to claim 6 is characterized in that: The model training module adopts maximizing the evidence lower bound optimization as the objective function of the neural relational reasoning model training.

9. The knowledge-driven multi-vehicle interactive driving decision system according to claim 6 is characterized by: The reward function includes a safety reward component, an efficiency reward component and a comfort reward component.

10. A knowledge-driven intelligent connected autonomous driving system, characterized in that: include: A decision module, configured to select an optimal driving strategy using the knowledge-driven multi-vehicle interactive driving decision method as described in any one of claims 1 to 4; The execution module is configured to generate corresponding vehicle control instructions based on the selected optimal driving strategy, and control the vehicle to perform corresponding operations according to the vehicle control instructions.