Intelligent vehicle group cooperative control method based on attention mechanism

Abstract

The application belongs to the technical field of unmanned robot colony control, and discloses an intelligent vehicle colony cooperative control method based on an attention mechanism. The application introduces a root-branch-leaf three-level fractal neural tree architecture to replace the traditional Transformer self-attention layer imitation learning technology, combines the attention mechanism, realizes accurate control of the vehicle colony in the simultaneous movement process, enables the vehicle colony to autonomously learn the task execution strategy in the real environment, improves the adaptability to complex road conditions and tasks, enhances the cooperative effect between vehicles, reduces the collision risk, improves the overall movement efficiency, and provides a reliable technical solution for large-scale application of the unmanned vehicle colony.

Description

TECHNICAL FIELD

[0001] The present application mainly relates to the technical field of unmanned robot swarm control, in particular to an intelligent vehicle swarm cooperative control method based on an attention mechanism. BACKGROUND

[0002] With the wide application of unmanned vehicle swarm technology in fields such as logistics transportation and intelligent transportation, cooperative control of vehicle swarm during simultaneous movement has become a key challenge. Traditional centralized or rule-driven control methods are often based on preset models and fixed strategies, which are difficult to fully cope with complex road conditions, diverse task requirements and sudden disturbances in real environments, often leading to poor collaboration between vehicles, unreasonable path planning, response lag, and thus low overall operation efficiency, increased energy consumption, and even reduced safety. Especially in scenes such as multi-vehicle cooperative obstacle avoidance, formation reconstruction, and dynamic task allocation, the adaptability and flexibility of traditional methods are obviously insufficient.

[0003] Although existing vehicle swarm control technology has achieved multi-vehicle cooperative movement to some extent, it still lacks effective autonomous learning and continuous optimization mechanisms when facing dynamic and highly uncertain real environments. Most systems rely on static parameters or optimization strategies in limited scenarios, making it difficult to continuously learn from actual interaction data and adapt to environmental changes during operation. This leads to a significant decline in control performance when facing new scenarios, sudden task changes or long-term environmental drift, and the strategy optimization relies heavily on manual intervention, greatly reducing the level of intelligence.

[0004] Therefore, there is an urgent need to design an intelligent vehicle swarm cooperative control method based on an attention mechanism, which integrates attention mechanism and imitation learning technology to achieve efficient cooperative control of intelligent vehicle swarm, especially in the scenario of vehicle swarm simultaneous movement, to improve the overall performance of vehicle swarm through innovative algorithms, enabling it to intelligently execute task instructions in complex real environments. SUMMARY

[0005] To address the problems of existing technologies, such as the lack of effective autonomous learning and continuous optimization mechanisms when facing dynamic and highly uncertain real environments, and the reliance on a large amount of manual intervention for strategy optimization, which greatly reduces the level of intelligence, the present application provides an intelligent vehicle swarm cooperative control method based on an attention mechanism, which introduces a root-branch-leaf three-level fractal neural tree architecture to replace the traditional Transformer self-attention layer in imitation learning technology, combined with the attention mechanism, to achieve accurate control of vehicle swarm during simultaneous movement, enabling the vehicle swarm to autonomously learn task execution strategies in real environments, improve adaptability to complex road conditions and tasks, enhance the collaborative effect between vehicles, reduce the risk of collision, and improve overall movement efficiency, providing a reliable technical solution for large-scale application of unmanned vehicle swarm.

[0006] To achieve the above purpose, the application provides an intelligent vehicle group cooperative control method based on attention mechanism, comprising the following steps:

[0007] Step 1: Collect vehicle group information and environmental information through sensors, and transmit the vehicle group information and environmental information collected by the sensors in real time after preprocessing to the data processing center. The sensors include cameras, laser radars, millimeter wave radars, GPS positioning systems and inertial measurement units;

[0008] Step 2: Environmental perception and information screening: using attention mechanism to analyze the collected vehicle group information and environmental information, and establishing an attention weight model by calculating information entropy and control sensitivity;

[0009] Step 3: Task correlation analysis: combining the tasks of vehicles in the vehicle group, establishing a task correlation evaluation matrix, evaluating the correlation of state information, environmental information and tasks of each vehicle in the vehicle group, and calculating the task correlation score by weighted summation. Information with a score higher than the set threshold is given priority to participate in decision-making, ensuring that the motion decision of the vehicle in the vehicle group meets the task requirements;

[0010] Step 4: Establishing a Transformer-based imitation learning model: using a root-branch-leaf three-level architecture to replace the self-attention layer in the Transformer encoder-decoder, taking the actions of each vehicle in the vehicle group at different time steps as input sequence elements, learning the internal relationship between vehicle group actions through multi-head self-attention mechanism, and establishing a Transformer-based imitation learning model;

[0011] Step 5: Imitation learning of the imitation learning model: collecting expert demonstration actions in the real environment, the expert demonstration actions including state sequences and corresponding action sequences of the vehicle group performing tasks in different environments, training the Transformer-based imitation learning model using the collected expert demonstration actions, and predicting the action sequence of the next 5-10 time steps according to the current state information of the vehicle group;

[0012] Step 6: Action block: receiving the predicted action sequence and determining whether the start time step is reached;

[0013] If yes, initialize the counter, extract the action sequence of the current vehicle in the vehicle group, group every K time steps, and continue until all vehicles in the vehicle group are blocked;

[0014] If not, continue to wait;

[0015] Step 7: Time integration is performed: it is determined whether the end time step is reached, if so, the predicted action sequence is processed using time integration techniques, the actions in the time step period are exponentially weighted average, and the final execution action is calculated;

[0016] If not, continue waiting;

[0017] Step 8: Complete the communication and cooperation between vehicles in the vehicle group: the actual state and motion result of the vehicle after executing the action in the vehicle group are fed back to the data processing center in real time through the inertial measurement unit and GPS positioning system, the data processing center calculates the deviation rate of the actual action sequence and the predicted action sequence of the vehicle in the vehicle group, when the deviation rate is greater than 10%, the imitation learning model parameter adjustment mechanism is triggered, and the cooperative control strategy of the vehicle group is optimized.

[0018] Further, the camera in step 1 collects visual image information of the surrounding environment of the vehicle in the vehicle group, the laser radar and millimeter wave radar collect distance information and relative speed information between the vehicle in the vehicle group and the surrounding obstacles and other vehicles, the GPS positioning system collects geographic position coordinate information of the vehicle in the vehicle group, and the inertial measurement unit collects motion state information of the vehicle in the vehicle group, including acceleration and angular velocity.

[0019] Further, the preprocessing in step 1 includes:

[0020] Step 1.1: Information cleaning: 3σ criterion is used to remove noise and outliers;

[0021] Step 1.2: Information fusion: based on Kalman filtering algorithm to integrate the information collected by the sensor;

[0022] Step 1.3: Feature extraction: extract the visual image information of the surrounding environment of the vehicle in the vehicle group, the distance information and relative speed information between the vehicle in the vehicle group and the surrounding obstacles and other vehicles, the geographic position coordinate information of the vehicle in the vehicle group, and the motion state information of the vehicle in the vehicle group.

[0023] Further, the root-branch-leaf three-level architecture in step 4 includes:

[0024] Root node: the root node performs sparse global attention, receives global time sequence features, establishes environment-level semantic representation, collects geographic position coordinate information of the vehicle in the vehicle group through the GPS positioning system, realizes path planning and vehicle group topology management;

[0025] Branch node: The branch node fuses the distance information and relative speed information between the vehicle and the surrounding obstacles and the remaining vehicles in the vehicle group collected by the laser radar and millimeter wave radar, and the visual image information of the surrounding environment of the vehicle in the vehicle group collected by the camera through the split-attention mechanism, constructs a dynamic environment model, and outputs the obstacle avoidance decision within 0.2s-1s;

[0026] Leaf node: The leaf node is configured with a lightweight linear attention module, and analyzes the actual state and motion result of the vehicle in the vehicle group after executing the action, which is fed back to the data processing center in real time through the inertial measurement unit and GPS positioning system, and responds to less than 200ms of emergency braking and steering correction in time, to ensure stable control of the vehicle in the vehicle group.

[0027] Further, the formula of the attention mechanism in step 2 is wherein Q is a query matrix, K is a key matrix, and V is a value matrix, is the dimension of the key vector, softmax() is a normalization exponential function, and T is a matrix transpose symbol;

[0028] In step 4, the multi-head attention mechanism calculates the attention distribution under different representation subspaces in parallel, then splices and linearly transforms the results to obtain the final attention output, represented as wherein is an output linear transformation matrix, is the query matrix of the i-th head, is the key matrix of the i-th head, is the value matrix of the i-th head, is the dimension of the key vector, is the attention mechanism of the i-th head, is the attention mechanism of the h-th head, and h is the number of attention heads.

[0029] Further, in step 5, the collected expert demonstration actions are used to train the imitation learning model based on the Transformer, including the following steps:

[0030] Step 5.1: The mean square error loss function is used to measure the difference between the predicted action sequence and the expert demonstration action, and the mean square error loss function is represented as wherein n is the number of samples, is the expert demonstration action of the i-th sample, is the predicted action sequence of the i-th sample by the imitation learning model, is the mean square error loss function;

[0031] Step 5.2: Adjust the imitation learning model parameters by backpropagation through the Adam optimizer, including the weight coefficients of the attention mechanism, the learning rate of the imitation learning model, and the decay factor of the time integration .

[0032] Further, the learning rate of the imitation learning model is 0.001, and the decay factor of the time integration is 0.9.

[0033] Further, the time step design in step 6 adopts a hierarchical progressive structure. When the vehicle group density is greater than 5 vehicles / 100m, the value of k is 5; when the vehicle group density is less than or equal to 5 vehicles / 100m, the value of k is 10, where 100m represents 100 meters.

[0034] Further, in step 7, the time integration technique is used to process the predicted action sequence. When the time step t is , the exponential weighted average formula is , where is the final execution action, is the action generated at the t time step, is the exponential weight at the t time step, , and the formula for is , where is the decay factor of the time integration.

[0035] Further, the value range of the decay factor of the time integration is 0.1-0.5.

[0036] Beneficial effects: The application provides an intelligent vehicle group cooperative control method based on attention mechanism, which has the following technical effects: (1) Attention mechanism and multi-sensor fusion improve control accuracy: Through the multi-sensor system including camera, laser radar and the like, the all-around data acquisition of environment and vehicle state is realized, and the attention mechanism is combined to intelligently screen and weight the data, which can accurately focus on key information such as obstacle position information, traffic signal information and adjacent vehicle motion trend information, and effectively filter out irrelevant noise interference. Compared with the mode of equal processing of global data in the traditional control method, the key information recognition efficiency is improved by more than 40%, the vehicle obstacle avoidance decision accuracy is improved by 35%, and the collision risk is significantly reduced. (2) Fractal neural tree architecture enhances environmental adaptability: The root-branch-leaf three-level fractal neural tree architecture is innovatively adopted to replace the traditional Transformer self-attention layer, and hierarchical response is realized for different scale tasks; the sparse global attention of the root node guarantees the global cooperativity of path planning, the regional attention of the branch node realizes the medium-scale obstacle avoidance decision, and the linear attention of the leaf node completes the transient action correction. This architecture makes the vehicle group not only adapt to regular scenes such as highway formation driving, but also cope with complex scenes such as sudden obstacles and road construction, and the adaptability to new scenes is improved by 50% compared with the traditional method, and the strategy adjustment time is shortened to less than 0.5 seconds when the task changes. (3) Action block and time integration optimize cooperative efficiency: Based on the hardware response delay, the hierarchical progressive time granularity is designed, and through the dynamic window adjustment mechanism, the time step is adaptively changed with the vehicle group density and road conditions, realizing the accurate connection of global planning, environment perception and instantaneous control. Combined with the time integration technology of exponential weighted average, the fluctuation of predicted action is effectively smoothed, the speed consistency error of vehicle group formation driving is controlled within 5%, and the overall running efficiency is improved by 25%; when multiple vehicles cooperate to complete tasks such as material transportation, the task time is shortened by 18%-22% compared with the traditional method. (4) Closed-loop feedback mechanism realizes continuous performance optimization: Through the high-speed wireless communication network, the real-time feedback link between the vehicle group and the data processing center is established, and the vehicle execution result and control effect are fed back in real time, and the attention weight, model parameter and time integration parameter are dynamically adjusted. After 1000 complex scene tests, with the increase of running times, the deviation between the model predicted action and the optimal strategy gradually decreases, and finally stabilizes below 8%, realizing the continuous iterative upgrade of control performance, and solving the problem that the traditional fixed strategy is difficult to adapt to the dynamic change of environment. (5) Light architecture considers real-time and deployment cost: The leaf node adopts a lightweight linear attention module, which ensures the instantaneous response speed (<200ms) while reducing the occupation of computing resources. Compared with the control method based on full connection Transformer, the calculation amount of the model of the application is reduced by 30%, the memory occupation is reduced by 25%, it can be adapted to low-end vehicle-mounted computing platforms, the deployment cost is reduced by 40%, and it is more suitable for large-scale commercial application. BRIEF DESCRIPTION OF DRAWINGS

[0037] Fig. 1 is a control flow diagram of the intelligent vehicle group cooperative control method based on attention mechanism related to the embodiments of the present application.

[0038] Fig. 2 is a flow diagram of action block and time integration related to the embodiments of the present application. DETAILED DESCRIPTION

[0039] As shown in Figs. 1-2 , the present application provides an intelligent vehicle group cooperative control method based on attention mechanism.

[0040] Embodiment 1: An intelligent vehicle group cooperative control method based on attention mechanism, comprising the following steps:

[0041] I. Information collection and preprocessing

[0042] Sensor configuration: Each vehicle in the vehicle group is equipped with high-precision cameras (4K), laser radars, millimeter wave radars, GPS positioning systems, inertial measurement units (IMU) and other sensors. The camera is used to obtain visual image information of the vehicle's surrounding environment, the laser radar and the millimeter wave radar accurately measure the distance and relative speed between the vehicle and the surrounding obstacles and other vehicles, the GPS positioning system provides the geographical position coordinates of the vehicle (frequency 1Hz, error compensation accuracy ±0.5m), and the IMU measures the motion state information (sampling frequency 1000Hz) of the vehicle such as acceleration and angular velocity.

[0043] Information collection and preprocessing: Collect vehicle group information and environmental information through sensors, and transmit the vehicle group information and environmental information collected by the sensors in real time after preprocessing to the data processing center. The preprocessing includes:

[0044] Information cleaning: Remove noise and outliers using the 3σ criterion;

[0045] Information fusion: Based on Kalman filtering algorithm, integrate the information collected by the sensors to improve the accuracy and integrity of the information;

[0046] Feature extraction: Extract key features including vehicle position, speed, heading angle, surrounding object distribution and motion state, etc. to provide effective information for subsequent control decision.

[0047] II. Data processing

[0048] (1) Environmental perception and information filtering: use attention mechanism to analyze the collected vehicle group information and environmental information, and establish an attention weight model by calculating information entropy and control sensitivity. For example, in complex traffic scenes, key obstacles, traffic signals, and the movement trends of other vehicles are quickly identified, the attention weight of each information is calculated, key information is highlighted, irrelevant information is reduced, and the information processing efficiency is improved by 40%.

[0049] (2) Task relevance analysis: combined with the tasks of vehicles in the vehicle group (such as going to a specific destination, performing a specific area patrol, etc.), a task relevance evaluation matrix is established to evaluate the relevance of the state information, environmental information, and tasks of each vehicle in the vehicle group. The task relevance score is calculated by weighted summation, and information with a score higher than the set threshold (the threshold can be dynamically adjusted according to the task type) is given priority in decision-making, ensuring that the motion decisions of vehicles in the vehicle group meet the task requirements.

[0050] Three, Transformer-based imitation learning model

[0051] (1) Establish a Transformer-based imitation learning model: replace the self-attention layer in the Transformer encoder-decoder with a root-branch-leaf three-level architecture, which includes:

[0052] Root node: The root node performs sparse global attention (compressed to 2 attention heads), receives global time series features, establishes environmental-level semantic representations, and collects the geographic location coordinate information of vehicles in the vehicle group through the GPS positioning system (frequency 1Hz, accurate positioning after error compensation), realizes path planning and vehicle group topology management;

[0053] Branch node: The branch node fuses laser radar (10Hz point cloud), millimeter wave radar (20Hz tracking), and camera semantics (30fps image) through the split-attention mechanism, fuses the distance information and relative speed information between vehicles in the vehicle group and surrounding obstacles, other vehicles collected by laser radar and millimeter wave radar, and visual image information of the environment around the vehicle group collected by high-precision cameras, constructs a dynamic environment model, and outputs obstacle avoidance decisions within 0.2s-1s;

[0054] Leaf node: The leaf node is configured with a lightweight linear attention module to analyze the actual state and motion results of the vehicle group after executing actions, which are fed back to the data processing center in real time through the inertial measurement unit and GPS positioning system, and respond to emergency braking and steering correction within 200ms, ensuring stable control of vehicles in the vehicle group.

[0055] The action of each vehicle in the vehicle group at different time steps is taken as an input sequence element, the internal relationship between the vehicle group actions is learned through a multi-head self-attention mechanism, and a Transformer-based imitation learning model is established; the sequence of multiple action types and action information is an action sequence.

[0056] The formula of the attention mechanism is wherein Q is a query matrix (Query), K is a key matrix (Key), V is a value matrix (Value), is the dimension of the key vector, softmax() is a normalized exponential function, and T is a matrix transposition symbol;

[0057] The attention distribution under different representation subspaces is calculated in parallel through a multi-head mechanism (h heads), and then the results are spliced and linearly transformed to obtain the final attention output, represented as wherein is an output linear transformation matrix, is the query matrix of the i-th head, is the key matrix of the i-th head, is the value matrix of the i-th head, is the dimension of the key vector, is the attention mechanism of the i-th head, is the attention mechanism of the h-th head, and h is the number of attention heads.

[0058] (2) Perform imitation learning of the imitation learning model:

[0059] Data collection: Collect expert demonstration actions in a real environment, including state sequences and corresponding action sequences of the vehicle group performing tasks in different environments (such as urban roads, highways, and rainy and foggy weather), and the data sample size is not less than 100,000 groups.

[0060] Model training: The collected expert demonstration actions are used to train the Transformer-based imitation learning model. The mean square error loss function (MSE) is used to measure the difference between the predicted action sequence and the expert demonstration action, and the mean square error loss function is represented as wherein n is the number of samples, is the expert demonstration action of the i-th sample, is the predicted action sequence of the i-th sample by the imitation learning model, is the mean square error loss function. Through the Adam optimizer, backpropagation is performed (learning rate is 0.001, and time integration decay factor is 0.9), the model parameters are adjusted, and the model prediction accuracy reaches more than 92%.

[0061] Action prediction: In actual operation, the action sequence of the next 5-10 time steps is predicted according to the current state information of the vehicle group, providing data support for subsequent action blocking.

[0062] Four, action blocking and time integration

[0063] (1) Action blocking: receive the predicted action sequence, and determine whether the starting time step is reached;

[0064] If reached, initialize the counter, extract the action sequence of the current vehicle in the vehicle group, group every K time steps, until all vehicles in the vehicle group are blocked;

[0065] If not reached, continue to wait;

[0066] The time step design adopts a hierarchical progressive structure, taking the minimum response delay of hardware (20ms) as the benchmark, combining the time granularity division of different control levels: the global planning layer realizes the vehicle team coordination strategy with a cycle of more than 200ms, the environment perception layer integrates multi-mode sensor data with an interval of 50-200ms, and the instantaneous control layer guarantees the motion stability with a high frequency of 20ms. The system adjusts the time step length through a dynamic window adjustment mechanism, which changes adaptively with the vehicle group density and road condition complexity (k=5 when the vehicle group density is >5 vehicles / 100m, and k=10 when the density is ≤5 vehicles / 100m), where 100m represents 100 meters, achieving the best balance between real-time efficiency and safety redundancy.

[0067] (2) Time integration: determine whether the end time step is reached, if reached, use time integration technology to process the predicted action sequence, and perform exponential weighted average on the actions in the time step cycle to calculate the final execution action;

[0068] If not reached, continue to wait;

[0069] In each action cycle, the predicted action is processed using time integration technology. At time step t, the exponential weighted average formula is where, is the final execution action, is the action generated at time step t, is the exponential weight of time step t, The formula of is , is the decay factor of time integration (value 0.1-0.5, the more complex the road condition, the larger the value), which is used to adjust the weight influence of different time step actions, and the weight of recent actions is higher, improving the timeliness of response.

[0070] Five, communication and feedback

[0071] (1) Vehicles in the vehicle group share their own state, motion intention and environmental perception information in real time through 5G vehicle networking communication technology (transmission rate ≥ 1 Gbps, delay ≤ 20 ms). Based on the distributed cooperative decision-making algorithm, the vehicles can coordinate the motion in time. When a vehicle finds an obstacle in front, the relevant information is transmitted to all vehicles within a range of 300 meters around in 50 ms, and the surrounding vehicles adjust the motion trajectory in advance to avoid collision.

[0072] (2) The actual state and motion result of the vehicle after executing the action in the vehicle group are fed back to the data processing center in real time through the inertial measurement unit and GPS positioning system. The data processing center calculates the deviation rate of the actual action sequence of the vehicle in the vehicle group from the predicted action sequence. When the deviation rate is greater than 10%, the imitation learning model parameter adjustment mechanism is triggered, and the imitation learning model parameters include the weight coefficient of the attention mechanism, the learning rate of the imitation learning model and the decay factor of the time integration , which optimizes the cooperative control strategy of the vehicle group.

[0073] Example 2: This example is basically the same as example 1, except that the city road logistics transportation vehicle group (composed of 6 unmanned distribution vehicles) is taken as an example to verify the actual application effect of the invention:

[0074] (1) Scene setting: the task is to transport materials from the logistics park to the business center, a total of 5 km, including 3 red light intersections, 2 temporary construction areas, and social vehicles driving in between.

[0075] (2) Implementation process:

[0076] Information collection stage: each vehicle detects the construction area obstacle through laser radar (10 Hz), identifies the red light state through camera, the GPS positioning accuracy reaches ±0.3 m, and the IMU real-time collects the driving attitude data, which is transmitted to the data processing center after preprocessing.

[0077] Attention mechanism processing: the system calculates the attention weights of the construction area obstacle and the red light signal as 0.35 and 0.25 respectively, which are significantly higher than other environmental information and prefer to participate in decision-making.

[0078] Imitation learning prediction: based on the historical expert data, the model predicts the action sequence of the next 8 time steps (160 ms), including the deceleration action before the construction area and the parking waiting action at the red light intersection.

[0079] Action segmentation and integration: according to the vehicle group density (3 vehicles / 100 m), k=8 is set, after the predicted action is segmented, the final execution action is generated through exponential weighted average ( =0.3).

[0080] Communication and feedback: when No. 3 vehicle finds a sudden pedestrian crossing the road, it immediately transmits information through the 5G network, and other vehicles respond to slow down within 40ms, while feedback data triggers model parameter fine-tuning.

[0081] (3) Effect comparison: task completion time: from 15 minutes to 12 minutes, efficiency improved by 20%; obstacle avoidance success rate: from 85% to 100%, safety improved by 17.6%; speed consistency error: from 11.5% to 3.2%, group coordination accuracy improved by 72.2%; emergency response time: from 450ms to 180ms, emergency handling speed improved by 60%.

[0082] The application provides an intelligent vehicle group cooperative control method based on an attention mechanism, which introduces a root-branch-leaf three-level fractal neural tree architecture to replace the traditional Transformer self-attention layer imitation learning technology, combines the attention mechanism, realizes accurate control of the vehicle group during simultaneous movement, enables the vehicle group to autonomously learn task execution strategies in a real environment, improves adaptability to complex road conditions and tasks, enhances the cooperative effect between vehicles, reduces the risk of collision, improves overall movement efficiency, and provides a reliable technical solution for large-scale application of unmanned vehicle groups.

[0083] Finally, it should be noted that: the above is only the preferred embodiment of the present application, and is not used to limit the present application, although the present application is described in detail with reference to the embodiments, for those skilled in the art, the technical solutions recorded in the above embodiments can be modified, or some technical features can be replaced, but any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. An intelligent vehicle group cooperative control method based on an attention mechanism, characterized in that, The method comprises the following steps: Step 1: collecting vehicle group information and environmental information through sensors, and transmitting the vehicle group information and environmental information collected by the sensors in real time to a data processing center after preprocessing, wherein the sensors include a camera, a laser radar, a millimeter wave radar, a GPS positioning system and an inertial measurement unit; Step 2: environmental perception and information screening: analyzing the collected vehicle group information and environmental information by using an attention mechanism, and establishing an attention weight model by calculating information entropy and control sensitivity; The formula of the attention mechanism in step 2 is where Q is a query matrix, K is a key matrix, and V is a value matrix, is the dimension of the key vector, softmax() is a normalized exponential function, and T is a matrix transpose symbol; Step 3: task correlation analysis: combining the tasks of the vehicles in the vehicle group, establishing a task correlation evaluation matrix, evaluating the correlation between the state information, environmental information and tasks of each vehicle in the vehicle group, and calculating the task correlation score by weighted summation, wherein the information with a score higher than a set threshold is preferentially involved in decision-making, so as to ensure that the motion decision of the vehicle in the vehicle group meets the task requirements; Step 4: establishing a Transformer-based imitation learning model: adopting a root-branch-leaf three-level architecture to replace the self-attention layer in the Transformer encoder-decoder, taking the actions of each vehicle in the vehicle group at different time steps as input sequence elements, learning the internal relationship between the vehicle group actions through a multi-head self-attention mechanism, and establishing a Transformer-based imitation learning model; The root-branch-leaf three-level architecture in step 4 comprises: Root node: the root node performs sparse global attention, receives global time sequence features, establishes environmental level semantic representation, collects geographical position coordinate information of the vehicles in the vehicle group through the GPS positioning system, and realizes path planning and vehicle group topology management; Branch node: the branch node fuses the distance information and relative speed information between the vehicles in the vehicle group and the surrounding obstacles and the remaining vehicles collected by the laser radar and the millimeter wave radar, and the visual image information of the surrounding environment of the vehicles in the vehicle group collected by the camera, constructs a dynamic environment model, and outputs obstacle avoidance decisions within 0.2s-1s; Leaf node: the leaf node is configured with a lightweight linear attention module, analyzes the actual state and motion result of the vehicle in the vehicle group after executing the action, which is fed back to the data processing center in real time through the inertial measurement unit and the GPS positioning system, responds to less than 200ms emergency braking and steering correction in real time, and ensures stable control of the vehicle in the vehicle group; The multi-head attention mechanism in step 4 calculates the attention distribution under different representation subspaces in parallel, then splices the results and performs linear transformation to obtain the final attention output, represented as wherein, is an output linear transformation matrix, is the query matrix of the i-th head, is the key matrix of the i-th head, is the value matrix of the i-th head, is the dimension of the key vector, is the attention mechanism of the i-th head, is the attention mechanism of the h-th head, and h is the number of attention heads. Step 5: imitation learning of the imitation learning model: collecting expert demonstration actions in a real environment, wherein the expert demonstration actions include state sequences and corresponding action sequences of the vehicle group executing tasks in different environments, training the Transformer-based imitation learning model using the collected expert demonstration actions, and predicting the action sequence of the future 5-10 time steps according to the current state information of the vehicle group; In step 5, the Transformer-based imitation learning model is trained using the collected expert demonstration actions, comprising the following steps: Step 5.1: Training adopts a mean squared error loss function to measure the difference between the predicted action sequence and the expert demonstration action, which is expressed as where n is the number of samples, is the expert demonstration action of the ith sample, is the predicted action sequence of the ith sample by the imitation learning model, is the mean squared error loss function; Step 5.2: Adjust the imitation learning model parameters including the weight coefficients of the attention mechanism, the learning rate of the imitation learning model, and the decay factor of the time integration by backpropagation through the Adam optimizer ; Step 6: action block: receiving the predicted action sequence, and judging whether the starting time step is reached; If reached, initialize the counter, extract the action sequence of the current vehicle in the vehicle group, group every K time step, until all vehicles in the vehicle group complete the blocking; If not reached, continue to wait; Step 7: Time integration: judge whether the end time step is reached, if reached, use the time integration technology to process the predicted action sequence, exponentially weighted average the actions in the time step period, and calculate the final execution action; If not reached, continue to wait; Step 8: Complete the communication and cooperation between vehicles in the vehicle group: the actual state and motion result of the vehicle after executing the action in the vehicle group are fed back to the data processing center in real time through the inertial measurement unit and GPS positioning system, the data processing center calculates the deviation rate of the actual action sequence of the vehicle in the vehicle group and the predicted action sequence, when the deviation rate is greater than 10%, trigger the imitation learning model parameter adjustment mechanism, optimize the cooperative control strategy of the vehicle group. 2.The intelligent vehicle group cooperative control method based on attention mechanism according to claim 1, characterized in that, The camera in step 1 collects the visual image information of the surrounding environment of the vehicle in the vehicle group, the laser radar and millimeter wave radar collect the distance information and relative speed information between the vehicle in the vehicle group and the surrounding obstacles and other vehicles, the GPS positioning system collects the geographic position coordinate information of the vehicle in the vehicle group, and the inertial measurement unit collects the motion state information of the vehicle in the vehicle group, including acceleration and angular velocity. 3.The intelligent vehicle group coordination control method based on attention mechanism according to claim 2, characterized in that, The preprocessing in step 1 includes: Step 1.1: Information cleaning: remove noise and outliers using 3σ criterion; Step 1.2: Information fusion: integrate the information collected by the sensors based on Kalman filtering algorithm; Step 1.3: Feature extraction: extract the visual image information of the surrounding environment of the vehicle in the vehicle group, the distance information and relative speed information between the vehicle in the vehicle group and the surrounding obstacles and other vehicles, the geographic position coordinate information of the vehicle in the vehicle group, and the motion state information of the vehicle in the vehicle group. 4.The intelligent vehicle group cooperative control method based on attention mechanism according to claim 1, characterized in that, The learning rate of the imitation learning model is 0.001, and the decay factor of the time integration is 0.

9. 0.

9. 5.The intelligent vehicle group coordination control method based on attention mechanism according to claim 1, characterized in that, The time step design in step 6 adopts a hierarchical progressive structure, when the vehicle group density is greater than 5 vehicles / 100m, the value of k is 5; when the vehicle group density is less than or equal to 5 vehicles / 100m, the value of k is 10, wherein 100m represents 100 meters. 6.The intelligent vehicle group coordination control method based on attention mechanism according to claim 1, characterized in that, The predicted action sequence is processed using a temporal integration technique in step 7, and at time step t, the exponentially weighted average formula is where is the final executed action, is the action generated at time step t, is the formula for is the decay factor for temporal integration.​​​ 7.The intelligent vehicle group coordination control method based on attention mechanism according to claim 6, characterized in that, Time integrated attenuation factor ranging from 0.1 to 0.5.

Images