Human driving experience-based interpretable deep learning autonomous lane change decision-making method and system

Abstract

The invention relates to the technical field of automatic driving, in particular to an interpretable deep learning autonomous lane change decision-making method and system based on human driving experience, and the method comprises the steps: collecting lane change data under different traffic conditions, screening and preprocessing the lane change data; generating a sample set containing driving style feature data, local interaction risk feature data and decision state vector feature data, and performing normalization and clustering processing on samples to obtain a training set and a verification set; and constructing a local interaction risk perception module, and dynamically calculating a traffic flow risk index based on a self-attention mechanism and a gating mechanism to reflect a traffic environment risk level. Data are clustered into an aggressive type, a normal type and a conservative type by integrating cognitive indexes related to lane changing behaviors and traditional kinematic characteristics, the driving style portrait can train a plurality of strategy networks with different fuzzy membership functions, and the problem of inconsistent decisions under different driving styles is effectively avoided.

Description

Technical Field

[0001] This invention relates to the field of UAV positioning error correction methods, and more specifically to an interpretable deep learning-based autonomous lane-changing decision-making method and system based on human driving experience. Background Technology

[0002] Autonomous lane changing is a typical task in autonomous driving decision-making and is of great significance in terms of traffic safety and efficiency. According to statistics from the National Highway Traffic Safety Administration (NHTSA) in the United States, lane-changing related accidents account for approximately 26.7% of all traffic accidents, significantly higher than the 20.9% of longitudinal following-related accidents, fully illustrating the complexity and high risk of lane-changing decisions. Unlike longitudinal driving, which only requires adjusting the longitudinal state of a single vehicle, lane changing involves game theory and interaction between multiple vehicles, reasoning about the intentions of surrounding vehicles, and high-level strategic decision-making. Therefore, it is widely regarded as a core indicator for measuring the interactive intelligence and decision-making capabilities of autonomous driving systems. How to generate lane-changing decisions that conform to both traffic regulations and human driving intuition while ensuring safety and efficiency has become a significant technical challenge in the field of intelligent driving.

[0003] While existing deep learning-based lane-change decision-making frameworks have achieved significant results, they still face three fundamental challenges in generating reliable decisions that align with human behavior patterns. First, current methods lack personalized modeling based on psychology. They simulate driving styles solely through physical quantities such as speed, acceleration, and relative distance, failing to capture key psychological factors like risk preference, yielding tendencies, time pressure, and subjective judgments of other vehicles' intentions, resulting in insufficient modeling of individual differences. Second, existing methods lack dynamic and interpretable situational risk assessment mechanisms. Human drivers naturally focus their attention on the most threatening target vehicles based on environmental changes and adjust their focus according to traffic conditions. However, most existing methods rely on fixed neighborhood topologies or static weighting schemes, failing to achieve adaptive risk perception and dynamic reasoning, thus making it difficult to reliably assess overall traffic risk in complex scenarios. Finally, existing deep learning models generally lack interpretability. Lane-change decisions often rely on deep neural networks for end-to-end policy learning, possessing powerful perception and decision-making capabilities, but their reasoning process is abstract and difficult to trace, making the key decision-making basis difficult to understand. In abnormal or high-risk scenarios, the model output may even conflict with human intuition. The lack of clear and timely explanations not only undermines the credibility of the system, but also creates obstacles for actual deployment and regulatory approval.

[0004] In summary, existing methods have shortcomings in modeling human driving psychology, dynamically perceiving traffic risks, and improving the interpretability of decisions, directly limiting the reliability and widespread adoption of autonomous driving lane-changing systems. Introducing a psychology-driven personalized modeling mechanism into a deep learning framework, combined with a dynamic interactive risk perception module, and enhancing decision interpretability through fuzzy logic and other methods, would not only generate lane-changing behaviors that better align with human driving intuition and social norms but also improve the system's transparency and credibility. This is of great significance for enhancing user trust, promoting the practical application of intelligent driving technology, and meeting the traceability requirements of future regulations. Summary of the Invention

[0005] In view of this, the present invention provides an interpretable deep learning-based autonomous lane-changing decision-making method and system based on human driving experience, aiming to solve the above-mentioned technical problems.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: An interpretable deep learning-based autonomous lane-changing decision-making method based on human driving experience includes the following steps: S1. Collect lane-changing data under different traffic conditions, filter and preprocess the lane-changing data to generate a sample set containing driving style feature data, local interaction risk feature data and decision state vector feature data, and normalize and cluster the samples to obtain the training set and validation set. S2. Construct a local interactive risk perception module, dynamically calculate the traffic flow risk index based on the self-attention mechanism and gating mechanism. The risk index is obtained by the normalized weighted sum of the gating risk scores and is used to reflect the risk level of the traffic environment. S3. Construct an interpretable deep learning decision module based on a fuzzy neural network, which receives the decision state vector and outputs the lane-changing behavior decision result. The decision module includes a fuzzification layer, a rule layer, a normalization layer and an output layer. The final decision is made by triggering intensity calculation through fuzzy rules. S4. Based on the backpropagation algorithm, train a fuzzy neural network to construct the weight parameters of an interpretable deep learning decision module. Verify the decision effect through a validation set and visualize the activation degree of fuzzy rules to extract interpretable lane-changing rules.

[0007] Furthermore, the lane-changing data filtering and preprocessing in S1 includes: Load the NGSIM and HighD datasets, analyze vehicle position, speed and lane-changing behavior at a sampling frequency of 10Hz, and store them as raw data scene fragments; filter non-passenger vehicle samples and retain only left lane-changing operations, remove multiple lane-changing records, and generate valid lane-changing behavior scene fragments; model lane-changing behavior into preparation phase, execution phase and adjustment phase, and segment each phase time with the vehicle lateral speed threshold.

[0008] Furthermore, the clustering processing of driving style feature data in S1 uses the K-means algorithm to cluster the samples into three categories: conservative, normal, and aggressive, corresponding to driving style labels. The driving style characteristic data includes waiting time before lane change, speed difference between the vehicle and the vehicle in front at the start of lane change, longitudinal relative distance between the vehicle and the vehicle to the left front at the start of lane change, and longitudinal relative distance between the vehicle and the vehicle to the left rear at the start of lane change.

[0009] Furthermore, the execution of the local interaction risk perception module in S2 includes: The local interaction risk feature vectors of each vehicle are activated using the LeakyReLU activation function. Projecting to higher-dimensional space: (1); in, For trainable weight matrix, It is the bias vector; Calculate query vector and key vector : (2); in, , The weight matrix is ​​trainable. Calculate attention weights based on the self-attention mechanism; (3); in, This represents the attention weight of vehicle j relative to vehicle i, and N is the total number of interactive vehicles; Generate scalar gated values ​​based on gating mechanisms: (4); in, For trainable weight matrix, It is the bias vector; Computational heuristic local risk score: (5); in, , , For learnable non-negative weights for velocity, acceleration, and distance, To ensure numerical stability and avoid a very small value when divided by zero For direction indication binary variables, This indicates whether the vehicle is in front or behind. Let be the relative speed between this vehicle and vehicle i. Let be the relative acceleration between the vehicle and vehicle i. This is the absolute distance between vehicle i and the current vehicle; The global attention score for each vehicle is obtained by calculating the average of each column of the attention matrix: (7); in, Let be the average value of each column of the attention matrix for vehicle k; Calculate the traffic flow risk index It is represented as the normalized weighted sum of the gating risk scores: (8); in, Let J be the scalar gating value for vehicle j. For vehicle j, a heuristic local risk score is given.

[0010] Furthermore, in S3, the fuzzy neural network constructs an interpretable deep learning decision module. The fuzzification layer maps the input features through K Gaussian membership functions, constructs a rule layer, calculates the fuzzy rule triggering intensity through the rule layer, constructs a normalization layer to normalize the rule activation values, constructs an output layer and obtains the final aggregated output through linear mapping, and finally calculates the probability of executing lane change.

[0011] Furthermore, the Gaussian membership function mapping formula for the fuzzification layer in step 3 is: (9); in, As input features, The center of the k-th fuzzy set of the j-th input is... Let be the variance of the k-th fuzzy set of the j-th input.

[0012] Furthermore, each rule layer r is composed of fuzzy sets, and the corresponding trigger strength is: (10); in, Let r be the emission intensity of the r-th rule. The input is the membership degree in the fuzzy set. The index of the fuzzy set assigned to the input in the rule; Normalize the rule activation values: (11); in, R is the normalized weight of the rule, representing the relative contribution of rule r to the final output, and R is the total number of fuzzy rules in the network. The final aggregated output is obtained through linear function mapping. : (12) in, This is the output corresponding to rule r. For the input feature vector, A learnable weight matrix for rules. These are learnable biases used for rules; The probability of performing a lane change is: (13); in, This is the predicted probability of lane changing.

[0013] Furthermore, in step S4, the decision-making effect is verified using a validation set, and the activation level of fuzzy rules is visualized to extract interpretable lane-changing rules. The model parameters are optimized using the backpropagation algorithm, and the training loss function is set to minimize the binary cross-entropy loss: (14); in, For the true labels of the samples, Predict the probability of lane change for the sample; The fuzzy membership parameters and rule-level linear coefficients obtained from training are substituted into the decision model of the local computing unit, and the decision effect is verified on the validation set divided in S1. For the playback of the scene segments in the dataset, the activation degree of different fuzzy rules is visualized. Based on the degree of activation of each membership function in different lane-changing stages, lane-changing solution rules with high interpretability are extracted.

[0014] An interpretable deep learning autonomous lane-changing decision-making system based on human driving experience includes: a lane-changing data filtering and feature preprocessing module, a local interaction risk perception module, an interpretable deep learning decision-making construction module, and a decision model training and fuzzy rule reasoning module. The lane change data filtering and feature preprocessing module is used to collect diverse lane change data under different traffic conditions and perform effective data filtering and preprocessing. The local interactive risk perception module is used to construct local interactive risk perception, dynamically focus on key vehicles that affect lane changing behavior, and output a traffic flow risk index. The interpretable deep learning decision building module is used to construct interpretable deep learning decisions based on traffic context information using fuzzy neural networks, and output lane-changing behavior decision results. The decision model training and fuzzy rule inference module is used to optimize model parameters according to the backpropagation algorithm, set the training loss function to minimize the binary cross-entropy loss, substitute the fuzzy membership parameters and rule-level linear coefficients obtained from the training into the decision model of the local computing unit, and verify the decision effect on the partitioned validation set.

[0015] This invention proposes an unsupervised driving style recognition method based on psychological principles. This method integrates cognitive indicators related to lane-changing behavior with traditional kinematic features to cluster data into three categories: aggressive, normal, and conservative. This driving style profile can train multiple policy networks with different fuzzy membership functions, thereby improving personalization and effectively avoiding inconsistencies in decision-making under different driving styles.

[0016] This invention proposes a local interaction perception module based on human thinking patterns. This module integrates attention mechanisms, adaptive gating technology, and heuristic risk scoring to dynamically identify the most influential interacting vehicles during lane changes. By combining relative motion characteristics with weighted interaction saliency and contextual risk, this module can calculate a traffic flow risk index, thereby improving human-like risk perception capabilities and decision consistency.

[0017] This invention combines fuzzy rules with a policy neural network. The traffic flow risk index guides risk perception decisions, while rule extraction and visualization enable a transparent, human-like, and traceable reasoning process, thereby enhancing user trust and reliability.

[0018] This invention addresses the shortcomings of existing methods in terms of personalized modeling, contextual understanding, and decision interpretability, significantly improving the human-likeness of autonomous driving lane-changing decisions and user trust, and is applicable to various scenarios such as highways and urban roads. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0020] Figure 1 This is a flowchart of the automatic lane-changing decision-making method of the present invention; Figure 2 This is a flowchart of the lane change data filtering and feature preprocessing of the present invention; Figure 3 This is a schematic diagram of the phased modeling of lane-changing behavior according to the present invention; Figure 4This is a schematic diagram of the partial interactive risk perception module structure of the present invention; Figure 5 This is a schematic diagram of the structure of the interpretable deep learning decision module of the present invention. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] like Figure 1-5 As shown, this patent abstracts the proposed automatic lane-changing decision-making method and system into a lane-changing data filtering and feature preprocessing module, a local interaction risk perception module, an interpretable deep learning decision-making construction module, and a decision model training and fuzzy rule reasoning module. The modeling flowcharts of each module are as follows: Step 1: First, collect diverse lane-changing data under different traffic conditions, and then perform effective data filtering and preprocessing. The flowchart for Step 1 is as follows: Figure 1 As shown: (1.1) Load the NGSIM and HighD datasets, analyze the vehicle position, speed and lane-changing behavior in the datasets at a sampling frequency of 10Hz, and store them as original data scene fragments.

[0023] (1.2) Filter non-passenger vehicle samples and retain only left lane change operations, remove multiple lane change records to ensure sample independence, and generate effective lane change behavior scene fragments.

[0024] (1.3) such as Figure 3 As shown, the lane-changing behavior is modeled into three stages: preparation, execution, and adjustment, corresponding to four time points, namely the time when the lane-changing intention is generated. The moment the lane-changing behavior begins Crossing the middle of the lane and the end of the lane change Among them, the start time of the lane-changing behavior. The lane change ends when the lateral speed exceeds 0.34 m / s, and the lane change is completed when the lateral speed decreases to approximately 0.2 m / s. The moment of crossing the center line is also considered. This refers to the moment when the center of the vehicle crosses the lane.

[0025] (1.4) Generate corresponding data samples for each time step. Each sample contains feature data and labels. The feature data includes driving style feature data, local interaction risk feature data, and decision state vector feature data, as shown in Tables 1, 2, and 3, respectively. The labels are binary classification labels, with 1 for the execution or adjustment stage and 0 for the preparation stage.

[0026] Table 1 Driving Style Characteristics Data

[0027] Table 2 Local Interaction Risk Characteristics Data

[0028] Table 3. Feature Data of Decision State Vector

[0029] (1.5) The table in Table 1 The four features were normalized, and the K-means algorithm was used to cluster the samples in the data into three categories: conservative, normal, and aggressive, corresponding to driving style labels. Among them, the conservative type is characterized by a long waiting time, a preference for larger gaps, and stable control behavior; the normal type is characterized by a moderate reaction time and a balanced operating strategy; and the aggressive type is characterized by a short waiting time, acceptance of small gaps, and active control of input. 20% of each subset of the dataset is randomly selected as the validation set, and 80% is used as the training set.

[0030] Step 2: Construct a local interactive risk perception module, dynamically focusing on key vehicles affecting lane-changing behavior, and outputting a traffic flow risk index. The flowchart for Step 2 is as follows. Figure 2 As shown: (2.1) The local interaction risk feature vectors of each vehicle are activated by the LeakyReLU activation function. Projecting to higher-dimensional space: (1); in, For trainable weight matrix, This is the bias vector.

[0031] (2.2) Calculate the query vector and key vector : (2); in, , This is a trainable weight matrix.

[0032] (2.3) Calculate attention weights based on self-attention mechanism: (3); in, This represents the attention weight of vehicle j relative to vehicle i, and N is the total number of interactive vehicles.

[0033] (2.4) Based on the gating mechanism, a scalar gate is generated for each vehicle to adaptively adjust the degree of influence of each vehicle: (4); in, For trainable weight matrix, This is the bias vector.

[0034] (2.5) To evaluate and perceive high-risk scenarios such as rapid approach, acceleration, or close proximity, a heuristic local risk score is calculated: (5); in, , , For learnable non-negative weights for velocity, acceleration, and distance, To ensure numerical stability and avoid a very small value when divided by zero For direction indication binary variables, This indicates whether the vehicle is in front or behind. Let be the relative speed between this vehicle and vehicle i. Let be the relative acceleration between the vehicle and vehicle i. This is the absolute distance between vehicle i and the current vehicle.

[0035] (2.6) The global attention score for each vehicle is obtained by calculating the average of each column of the attention matrix: (6); (2.7) Standardize the vehicle attention score: (7); in, Let be the average value of each column of the attention matrix for vehicle k; (2.8) Calculate the traffic flow risk index It is represented as the normalized weighted sum of the gating risk scores: (8); in, Let J be the scalar gating value for vehicle j. For the heuristic local risk score of vehicle j, A lower value indicates a more relaxed traffic environment and lower pressure, while a higher value indicates significant surrounding risks and higher task complexity.

[0036] Step 3: Construct an interpretable deep learning decision-making module based on traffic context information using a fuzzy neural network, and output the lane-changing behavior decision results. The flowchart for Step 3 is as follows: Figure 3 As shown: (3.1) Construct an input layer to receive the decision state vector of the current traffic scenario as shown in Table 3. ,in This is the output of step 2.

[0037] (3.2) Construct a fuzzing layer to map each input vector through K Gaussian membership functions: (9); in, As input features, The center of the k-th fuzzy set of the j-th input is... Let be the variance of the k-th fuzzy set of the j-th input.

[0038] (3.3) Construct a rule layer, where each rule r is composed of fuzzy sets, and the corresponding trigger strength is: (10); in, Let r be the emission intensity (activation level) of the r-th rule. The input is the membership degree in the fuzzy set. This is the index of the fuzzy set assigned to the input in the rule.

[0039] (3.4) Construct a normalization layer to normalize the rule activation values: (11); in, R is the normalized weight of the rule, representing the relative contribution of rule r to the final output, and R is the total number of fuzzy rules in the network.

[0040] (3.5) Construct the output layer and obtain the final aggregated output through linear function mapping. : (12); in, Let be the output corresponding to rule r, and let x be the input feature vector. A learnable weight matrix for rules. These are learnable biases used for rules.

[0041] Finally, the probability of performing a lane change can be obtained as follows: (13); in, This is the predicted probability of lane changing.

[0042] Step 4: Decision model weight training and fuzzy rule inference.

[0043] (4.1) Optimize the model parameters according to the backpropagation algorithm, and set the training loss function to minimize the binary cross-entropy loss: (14); in, For the true labels of the samples, Predict the probability of lane change for the sample.

[0044] (4.2) Substitute the fuzzy membership parameters and regular linear coefficients obtained from the training into the decision model of the local computing unit, and verify the decision effect on the validation set divided in step 1.

[0045] (4.3) For the playback of scene segments in the dataset, the activation degree of different fuzzy rules is visualized. Based on the degree of activation of each membership function in different lane-changing stages, the lane-changing solution rules with high interpretability are extracted as shown in the table below.

[0046]

[0047] 1. Existing lane-changing decision-making methods lack personalized modeling based on psychology. They typically simulate driving style solely through physical parameters such as speed, acceleration, and relative distance, making it difficult to capture key psychological factors. For example, some drivers, despite driving at low speeds, can change lanes quickly and react nimbly, indicating a high risk tolerance and decisiveness; while others, despite driving at higher speeds, delay lane changes due to a cautious tendency. This invention proposes an unsupervised driving style recognition method based on psychological principles. This method integrates cognitive indicators related to lane-changing behavior with traditional kinematic features, clustering data into three categories: aggressive, normal, and conservative (step 1). This driving style profile can train multiple policy networks with different fuzzy membership functions, thereby improving personalization and effectively avoiding inconsistencies in decision-making under different driving styles.

[0048] 2. Existing methods lack the ability to perceive and assess overall traffic risk in a context-aware, dynamic, and interpretable manner. Human drivers naturally focus on the most threatening vehicles based on changing intentions and the surrounding environment, but most current methods fail to reproduce this adaptive reasoning ability. This invention proposes a local interaction perception module based on human thinking patterns. This module integrates attention mechanisms, adaptive gating technology, and heuristic risk scoring (step 2) to dynamically identify the most influential interacting vehicles during lane changes. By combining relative motion characteristics with weighted interaction saliency and contextual risk, this module can calculate a traffic flow risk index, thereby improving human-like risk perception and decision consistency.

[0049] 3. Existing deep learning-based lane-changing decision-making models generally lack interpretability, which limits their ability to support decision traceability and build user trust. This invention combines fuzzy rules with a policy neural network (step 3). The traffic flow risk index guides risk perception decisions (the input of step 3 uses the traffic flow risk index from step 2), while rule extraction and visualization enable a transparent, human-like, and traceable reasoning process (step 4, rule extraction), thereby enhancing user trust and reliability.

[0050] In summary, this invention addresses the shortcomings of existing methods in terms of personalized modeling, contextual understanding, and decision interpretability, significantly improving the human-likeness of autonomous driving lane-changing decisions and user trust, and is applicable to various scenarios such as highways and urban roads.

[0051] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0052] The various embodiments described in this specification are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An interpretable deep learning-based autonomous lane-changing decision-making method based on human driving experience, characterized in that, Includes the following steps: S1. Collect lane-changing data under different traffic conditions, filter and preprocess the lane-changing data to generate a sample set containing driving style feature data, local interaction risk feature data and decision state vector feature data, and normalize and cluster the samples to obtain the training set and validation set. S2. Construct a local interactive risk perception module, dynamically calculate the traffic flow risk index based on the self-attention mechanism and gating mechanism. The risk index is obtained by the normalized weighted sum of the gating risk scores and is used to reflect the risk level of the traffic environment. S3. Construct an interpretable deep learning decision module based on a fuzzy neural network, which receives the decision state vector and outputs the lane-changing behavior decision result. The decision module includes a fuzzification layer, a rule layer, a normalization layer and an output layer. The final decision is made by triggering intensity calculation through fuzzy rules. S4. Based on the backpropagation algorithm, train a fuzzy neural network to construct the weight parameters of an interpretable deep learning decision module. Verify the decision effect through a validation set and visualize the activation degree of fuzzy rules to extract interpretable lane-changing rules.

2. The interpretable deep learning-based autonomous lane-changing decision-making method based on human driving experience according to claim 1, characterized in that, The lane change data filtering and preprocessing in S1 includes: Load the NGSIM and HighD datasets, analyze vehicle position, speed and lane-changing behavior at a sampling frequency of 10Hz, and store them as raw data scene fragments; filter non-passenger vehicle samples and retain only left lane-changing operations, remove multiple lane-changing records, and generate valid lane-changing behavior scene fragments; model lane-changing behavior into preparation phase, execution phase and adjustment phase, and segment each phase time with the vehicle lateral speed threshold.

3. The interpretable deep learning-based autonomous lane-changing decision-making method based on human driving experience according to claim 2, characterized in that, The clustering processing of driving style feature data in S1 uses the K-means algorithm to cluster the samples into three categories: conservative, normal, and aggressive, corresponding to driving style labels. The driving style characteristic data includes waiting time before lane change, speed difference between the vehicle and the vehicle in front at the start of lane change, longitudinal relative distance between the vehicle and the vehicle to the left front at the start of lane change, and longitudinal relative distance between the vehicle and the vehicle to the left rear at the start of lane change.

4. The interpretable deep learning-based autonomous lane-changing decision-making method based on human driving experience according to claim 2, characterized in that, The execution of the local interaction risk perception module in S2 includes: The local interaction risk feature vectors of each vehicle are activated using the LeakyReLU activation function. Projection to higher-dimensional space: （1）； in, For trainable weight matrix, It is the bias vector; Calculate query vector and key vector : （2）； in, , The weight matrix is ​​trainable. Calculate attention weights based on the self-attention mechanism; （3）； in, This represents the attention weight of vehicle j relative to vehicle i, and N is the total number of interactive vehicles; Generate scalar gated values ​​based on gating mechanisms: （4）； in, For trainable weight matrix, It is the bias vector; Computational heuristic local risk score: （5）； in, , , For learnable non-negative weights for velocity, acceleration, and distance, To ensure numerical stability and avoid a very small value when divided by zero For direction indication binary variables, This indicates whether the vehicle is in front or behind. Let be the relative speed between this vehicle and vehicle i. Let be the relative acceleration between the vehicle and vehicle i. This is the absolute distance between vehicle i and this vehicle; The global attention score for each vehicle is obtained by calculating the average of each column of the attention matrix: （6）； N is the total number of interactive vehicles; Standardize the vehicle's attention score: （7）； in, Let be the average value of each column of the attention matrix for vehicle k; Calculate the traffic flow risk index It is represented as the normalized weighted sum of the gating risk scores: （8）； in, Let J be the scalar gating value for vehicle j. For vehicle j, a heuristic local risk score is given.

5. The interpretable deep learning-based autonomous lane-changing decision-making method based on human driving experience according to claim 1, characterized in that, In S3, the fuzzy neural network constructs an interpretable deep learning decision module. The fuzzification layer maps the input features through K Gaussian membership functions, constructs a rule layer, calculates the fuzzy rule triggering intensity through the rule layer, constructs a normalization layer to normalize the rule activation values, constructs an output layer and obtains the final aggregated output through linear mapping, and finally calculates the probability of executing lane change.

6. The interpretable deep learning-based autonomous lane-changing decision-making method based on human driving experience according to claim 5, characterized in that, The Gaussian membership function mapping formula for the fuzzy layer in step 3 is: （9）； in, For input features, The center of the k-th fuzzy set of the j-th input is... Let be the variance of the k-th fuzzy set of the j-th input.

7. The interpretable deep learning-based autonomous lane-changing decision-making method based on human driving experience according to claim 5, characterized in that, Each rule layer r is composed of fuzzy sets, and the corresponding trigger strength is: （10）； in, Let r be the emission intensity of the r-th rule. The input is the membership degree in the fuzzy set. The index of the fuzzy set assigned to the input in the rule; Normalize the rule activation values: （11）； in, R is the normalized weight of the rule, representing the relative contribution of rule r to the final output, and R is the total number of fuzzy rules in the network. The final aggregated output is obtained through linear function mapping. : （12）； in, This is the output corresponding to rule r. For the input feature vector, A learnable weight matrix for rules. These are learnable biases used for rules; The probability of performing a lane change is: （13）； in, This is the predicted probability of lane changing.

8. The interpretable deep learning-based autonomous lane-changing decision-making method based on human driving experience according to claim 1, characterized in that, In step S4, the decision-making effect is verified using a validation set, and the activation level of fuzzy rules is visualized to extract interpretable lane-changing rules. Specifically: The model parameters are optimized using the backpropagation algorithm, and the training loss function is set to minimize the binary cross-entropy loss: （14）； in, For the true labels of the samples, Predict the probability of lane change for the sample; The fuzzy membership parameters and rule-level linear coefficients obtained from training are substituted into the decision model of the local computing unit, and the decision effect is verified on the validation set divided in S1. For the playback of the scene segments in the dataset, the activation degree of different fuzzy rules is visualized. Based on the degree of activation of each membership function in different lane-changing stages, lane-changing solution rules with high interpretability are extracted.

9. An interpretable deep learning-based autonomous lane-changing decision-making system based on human driving experience, applicable to any of the interpretable deep learning-based autonomous lane-changing decision-making methods based on human driving experience as described in claims 1-8, characterized in that, include: The module includes a lane-change data filtering and feature preprocessing module, a local interaction risk perception module, an interpretable deep learning decision building module, and a decision model training and fuzzy rule reasoning module. The lane change data filtering and feature preprocessing module is used to collect diverse lane change data under different traffic conditions and perform effective data filtering and preprocessing. The local interactive risk perception module is used to construct local interactive risk perception, dynamically focus on key vehicles that affect lane changing behavior, and output a traffic flow risk index. The interpretable deep learning decision building module is used to construct interpretable deep learning decisions based on traffic context information using fuzzy neural networks, and output lane-changing behavior decision results. The decision model training and fuzzy rule inference module is used to optimize model parameters according to the backpropagation algorithm, set the training loss function to minimize the binary cross-entropy loss, substitute the fuzzy membership parameters and rule-level linear coefficients obtained from the training into the decision model of the local computing unit, and verify the decision effect on the partitioned validation set.

Images