Heavy-duty car path planning and tracking method and system giving consideration to efficiency and energy-saving indexes

Abstract

The invention relates to the technical field of path planning and tracking, in particular to a heavy-duty car path planning and tracking method and system giving consideration to efficiency and energy-saving indexes, and the method comprises the steps: constructing a multi-dimensional state space fusing a vehicle state, environment information and real-time energy consumption; a reinforcement learning decision model is constructed and trained, the output of the reinforcement learning decision model is a layered action space, and a reward weight dynamic adjustment mechanism based on scene recognition is introduced to dynamically adjust the weight of safe, efficient and energy-saving rewards and drive model learning. And after training is completed, the model is deployed to a vehicle-mounted unit for real-time planning and control. During real vehicle operation, data are continuously collected, a model is updated through incremental learning, and an energy consumption-road condition association database is synchronously constructed. The database is inquired during real-time decision making, and model output is guided or finely adjusted by using a historical energy-saving strategy, so that rapid scene self-adaption of the energy-saving strategy is realized. Safe, efficient and energy-saving running of the heavy-duty car in a complex environment is achieved.

Description

Technical Field

[0001] This application relates to the field of path planning and tracking technology, specifically to a method and system for path planning and tracking of heavy-duty vehicles that takes into account both efficiency and energy-saving indicators. Background Technology

[0002] Traditional intelligent driving systems often adopt a modular design, separating path planning, vehicle control, and energy management into independent modules. The objective functions of each module are inconsistent: the planning module focuses on the shortest path, the control module focuses on tracking accuracy, and the energy-saving module adopts a fixed strategy. This results in low system coupling, making it difficult to achieve optimal global performance and unable to coordinate optimization in complex dynamic scenarios.

[0003] Existing related technologies, such as path planning methods based on Frenet coordinates and Cartesian coordinates, do not consider the dynamic interaction between the vehicle and obstacles and energy consumption indicators, and are not adaptable to complex dynamic environments. The state grid algorithm is used to generate candidate paths, but the sampling points are fixed, making it difficult to search for densely obstacle areas in a refined manner, and it also ignores energy consumption factors.

[0004] More importantly, existing methods lack scenario adaptability. Heavy vehicles need to prioritize energy conservation when frequently starting and stopping in congested areas, and focus on driving efficiency on highways. Fixed-weight optimization methods cannot adapt to these scenario differences, severely limiting the system's practicality and overall performance in real driving environments. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a method and system for heavy-duty vehicle route planning and tracking that balances efficiency and energy-saving indicators.

[0006] In a first aspect, the present invention provides a method for route planning and tracking of heavy-duty vehicles that takes into account both efficiency and energy-saving indicators, comprising the following steps: S1. Constructing a multi-dimensional state space required for a decision-making model based on reinforcement learning, wherein the multi-dimensional state space integrates the vehicle's own state, environmental dynamic information, traffic environment semantic information, and real-time energy consumption indicators. S2. Based on the multidimensional state space, construct and train the decision model. The training process includes: The output of the decision model is defined as a hierarchical action space, which includes a high-level action space for path planning and a low-level action space for vehicle control. Define a comprehensive reward function consisting of a weighted average of security rewards, efficiency rewards, and energy-saving rewards; A scene recognition-based dynamic adjustment mechanism for reward weights is introduced, namely: during model training, the weights of each reward item in the comprehensive reward function are dynamically adjusted according to the real-time identified driving scene type; The decision model is trained and learned with the goal of maximizing the comprehensive reward function. S3. Deploy the trained decision model to the on-board computing unit for real-time path planning and vehicle control; S4. During actual vehicle operation, driving data is continuously collected, and the parameters of the decision-making model are updated based on the incremental learning mechanism; at the same time, an energy consumption-road condition correlation database is built and continuously updated based on the collected data. S5. During the real-time decision-making process, the energy consumption-road condition correlation database is queried based on the current road condition characteristics, and the output of the decision-making model is guided or fine-tuned using the retrieved historical energy-saving strategies, so as to achieve rapid scenario adaptation of energy-saving strategies.

[0007] As a further limitation of the technical solution of the present invention, S1 specifically includes: Raw data of the vehicle's surrounding environment is collected by onboard LiDAR, millimeter-wave radar, cameras, and ultrasonic sensors; the raw data is then filtered, spatiotemporally aligned, and correlated to obtain dynamic environmental information including obstacle positions, speeds, and movement trends. The static information of road network, lane lines and traffic signs from high-precision maps is semantically fused with the dynamic environmental information to generate traffic environment semantic information. Based on the universal characteristic curve of the engine and the vehicle dynamics model of heavy-duty vehicles, the instantaneous fuel consumption rate and cumulative energy consumption are calculated according to the current vehicle status and driving instructions, which serve as real-time energy consumption indicators. The vehicle's own state, environmental dynamics, traffic environment semantics, and real-time energy consumption indicators are vectorized, encoded, and fused to form a multi-dimensional state space.

[0008] As a further limitation of the technical solution of this invention, the specific calculation formula of the comprehensive reward function is as follows:

[0009] in, These are the weighting coefficients for security rewards, efficiency rewards, and energy-saving rewards, respectively, and they satisfy the following conditions: ; Security rewards The calculation formula is:

[0010] in, The distance between the vehicle and the nearest obstacle; This is the safe distance threshold; This is the collision distance threshold; The reward value for safe driving; This is the penalty value when a collision occurs; This is the distance penalty coefficient; High-efficiency rewards The calculation formula is:

[0011] in, For the desired speed; This represents the travel time already elapsed; The preset maximum travel time; and These are the reward weighting coefficients for speed and time, respectively. Energy efficiency bonus The calculation formula is:

[0012] in, Energy consumption at the current moment; This is the preset maximum energy consumption; This is the weighting coefficient for energy-saving rewards.

[0013] The specific calculation formula of the comprehensive reward function is clarified. By weighted integration of three types of rewards—safety, efficiency, and energy saving—the multi-objective optimization requirements are quantified into a computable mathematical model, enabling the decision-making model to clearly perceive the priority of different performance indicators and avoid performance imbalance caused by single-objective optimization.

[0014] The safety reward, designed using a piecewise function, dynamically adjusts the reward value based on the distance between the vehicle and obstacles. This encourages vehicles to maintain a safe distance while effectively penalizing close-range risks and collision behaviors, thus reinforcing the model's safe driving tendency. The efficiency reward considers both driving speed and driving time, guiding vehicles to improve driving efficiency within a reasonable range. The energy-saving reward is directly linked to actual energy consumption, incentivizing the model to choose low-energy-consumption strategies. These three elements work together to achieve multi-objective balanced optimization.

[0015] As a further limitation of the technical solution of this invention, in S2, the constructed decision model adopts an Actor-Critic reinforcement learning architecture, wherein: Actor networks, as policy networks, are used to generate specific actions in a hierarchical action space based on a multidimensional state space. The Critic network, as a value network, is used to evaluate the expected long-term cumulative reward that can be obtained by performing the action in the current state; The decision model is trained through collaborative iterative optimization using an Actor network and a Critic network; the steps of the collaborative iterative optimization include: (a) In a simulated or real environment, the decision model performs actions based on the Actor network, obtains new states and immediate rewards after interacting with the environment, and stores the experience data of the interaction in the experience replay pool. (b) Sample a batch of empirical data from the empirical replay pool, calculate the value estimation error of the current Critic network using the advantage function, minimize the error using the gradient descent algorithm, and update the parameters of the Critic network. (c) Based on the value assessment provided by the updated Critic network, calculate the performance gradient of the Actor network and update the parameters of the Actor network along the gradient direction so that the policy tends to select actions that can obtain higher long-term cumulative rewards. (d) Repeat steps (a) to (c) until the policy performance of the decision model converges or reaches the preset training objective. An experience replay pool mechanism is introduced, which breaks the temporal correlation of data by storing and sampling interactive experience data, reducing sample redundancy and correlation interference during training, and improving the efficiency and robustness of model training.

[0016] Utilizing the advantage function Calculate the value estimation error, the advantage function Through formula Calculation, where Let the action value function be... This is the state value function.

[0017] As a further limitation of the technical solution of the present invention, in S2, the dynamic adjustment mechanism for reward weight based on scene recognition specifically includes: The types of driving scenarios that can be identified in real time include at least traffic congestion scenarios, highway scenarios, and ordinary road scenarios; When a congestion scenario is identified, the weight of energy-saving rewards in the comprehensive reward function is increased, while the weight of efficiency rewards is decreased. When a high-speed scenario is identified, the weight of the efficiency reward in the comprehensive reward function is increased, while the weight of the energy-saving reward is decreased.

[0018] As a further limitation of the technical solution of the present invention, in S4, the step of updating the decision model parameters based on the incremental learning mechanism includes: (a) During actual vehicle operation, real-time data on vehicle status, environmental perception, control commands, energy consumption and reward feedback are collected synchronously at a preset sampling frequency. The data is then processed by removing outliers, filling in missing values ​​and aligning time sequences to generate standardized data samples. (b) The standardized data samples are temporarily stored in the first experience pool according to the driving scenario classification; the data in the first experience pool are periodically evaluated and ranked based on the potential value of each data point for improving the current decision-making model, and high-value data is selected and transferred to the second experience pool using a priority storage strategy. (c) When the preset triggering conditions are met, the incremental learning process is started, and data is extracted from the first experience pool and the second experience pool in a preset ratio to form a training set. The parameters of the currently deployed decision model are used as the initial values ​​for small-batch training, and the convergence of the model's comprehensive reward loss function is used as the training termination condition to obtain the updated model parameters. The triggering conditions include at least the new data volume reaching a threshold, the model performance continuously declining, or the detection of new unrecorded scenarios. (d) Perform offline playback verification and online restricted trial operation on the updated model parameters; only when the verification and trial operation results meet the preset performance and safety standards will the updated model parameters be deployed to the vehicle computing unit to replace the original model parameters; if they do not meet the standards, the original model parameters will be rolled back and the corresponding data will be added to re-trigger incremental learning.

[0019] A dual-experience pool structure is adopted to manage real-vehicle driving data in a hierarchical manner. By selecting high-value data and storing it in the historical experience pool, the quality of training data is ensured, and the impact of redundant data on training efficiency is reduced. At the same time, the data preprocessing process effectively removes outliers and fills in missing values, ensuring the reliability of training data. Multi-dimensional incremental learning trigger conditions are set to avoid invalid model updates due to insufficient data volume, and can respond promptly to situations such as model performance degradation and the emergence of new scenarios, ensuring the timeliness and necessity of model updates. The method of small-batch training and initial parameter reuse reduces the computational resource consumption of real-vehicle updates and avoids performance fluctuations during the model update process.

[0020] A dual verification mechanism of offline verification and online trial operation is introduced to strictly control the performance and security standards of the updated model, thus avoiding the security risks caused by the deployment of unqualified models.

[0021] As a further limitation of the technical solution of this invention, the establishment of the energy consumption-road condition correlation database includes: The database storage structure is defined, using a combination of road slope intervals, traffic flow state levels, and vehicle speed ranges as the scenario index key, and the corresponding optimal energy consumption strategy parameters and execution effect evaluation values ​​as the associated data values, and an index is established. Specifically, the road slope intervals are divided according to a preset gradient, the traffic flow state levels are divided into three levels—smooth, slow, and congested—based on real-time vehicle density, and the vehicle speed ranges are divided according to preset speed intervals. The optimal energy consumption strategy parameters include the target acceleration range and engine speed range; the execution effect evaluation values ​​include the actual energy consumption value and the tracking accuracy error. During vehicle operation, real-time road gradient, traffic flow status, vehicle speed, and corresponding vehicle energy consumption data and execution strategy parameters are collected under different driving scenarios, and the data is cleaned and classified according to the scenarios. For each type of scenario, the optimal energy consumption strategy is selected from multiple sets of historical data, with energy consumption as the core optimization target and driving safety and tracking accuracy as constraints. The optimal energy consumption strategy is associated with the corresponding scenario features and execution effect evaluation values ​​and stored in the energy consumption-road condition association database.

[0022] As a further limitation of the technical solution of the present invention, in S5, the step of implementing scenario adaptation of energy-saving strategy based on energy consumption-road condition association database includes: Based on the road gradient, traffic flow status and vehicle speed collected in real time by the vehicle, the current scene features are generated, and similar historical scene records are retrieved in the energy consumption-road condition association database. Based on the search results and the preset similarity threshold, the system may choose to invoke the best historical strategy, use it as a reference for the basic strategy, or trigger the model to regenerate the initial strategy. For an initial or basic strategy determined based on historical records, the strategy parameters are fine-tuned in real time by combining dynamic information about the current environment and using an online optimization algorithm with the comprehensive reward function as the objective, to generate the final execution strategy. The actual execution effect of the final execution strategy will be added as a new record to the energy consumption-road condition association database.

[0023] Secondly, the present invention also provides a heavy-duty vehicle route planning and tracking system that takes into account both efficiency and energy-saving indicators, comprising: The state space construction module is used to construct the multi-dimensional state space required for the decision model based on reinforcement learning. The multi-dimensional state space integrates the vehicle's own state, environmental dynamic information, traffic environment semantic information, and real-time energy consumption indicators. The model training module is used to construct and train a decision model based on the multidimensional state space. The training process includes: The output of the decision model is defined as a hierarchical action space, which includes a high-level action space for path planning and a low-level action space for vehicle control. Define a comprehensive reward function consisting of a weighted average of security rewards, efficiency rewards, and energy-saving rewards; A scene recognition-based dynamic adjustment mechanism for reward weights is introduced, namely: during model training, the weights of each reward item in the comprehensive reward function are dynamically adjusted according to the real-time identified driving scene type; The decision model is trained and learned with the goal of maximizing the comprehensive reward function. The model deployment and execution module is used to deploy the trained decision model to the on-board computing unit for real-time path planning and vehicle control. The online learning and database construction module is used to continuously collect driving data during real vehicle operation and update the parameters of the decision model based on the incremental learning mechanism; at the same time, it builds and continuously updates the energy consumption-road condition correlation database based on the collected data. The scenario-adaptive decision module is used to query the energy consumption-road condition association database based on the current road condition characteristics during the real-time decision-making process, and use the retrieved historical energy-saving strategies to guide or fine-tune the output of the decision model, so as to achieve rapid scenario adaptation of energy-saving strategies.

[0024] As can be seen from the above technical solutions, this application has the following advantages: It breaks through the modular separation limitations of path planning, vehicle control, and energy management in traditional intelligent driving systems, achieving global collaborative optimization of planning, control, and energy conservation. By adopting a hierarchical action space design, path planning and vehicle control tasks are processed in layers, ensuring the rationality of path planning and improving the accuracy and reliability of decision-making in complex traffic environments. Furthermore, by introducing a dynamic weight adjustment mechanism based on scene recognition and an energy consumption-road condition correlation database, the system can adapt to the different needs of various driving scenarios such as congestion, highways, and ordinary roads, avoiding the shortcomings of fixed-weight methods that cannot adapt to scene changes.

[0025] By using an incremental learning mechanism to achieve continuous iterative optimization of the decision-making model, combined with the dynamic updates of the energy consumption-road condition correlation database, the system can continuously adapt to new driving scenarios and road condition characteristics, thereby improving its practicality and adaptability in real, complex, and dynamic environments. Attached Figure Description

[0026] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a flowchart illustrating the method provided in an embodiment of the present invention.

[0028] Figure 2 A block diagram of a system provided in an embodiment of the present invention. Detailed Implementation

[0029] To make the purpose, features, and advantages of this application more apparent and understandable, specific embodiments and accompanying drawings will be used to clearly and completely describe the technical solution protected by this application. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0030] Unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this application and in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0031] like Figure 1 As shown, this embodiment of the invention provides a method for heavy-duty vehicle route planning and tracking that balances efficiency and energy saving indicators, including the following steps: S1. Constructing the multi-dimensional state space required for a reinforcement learning-based decision-making model, wherein the multi-dimensional state space integrates the vehicle's own state, environmental dynamics, traffic environment semantics, and real-time energy consumption indicators; specifically including: (1) Raw data collection of the vehicle's surrounding environment, as detailed below: LiDAR is installed at the front of the vehicle roof to collect 3D point cloud data of obstacles within a set range, used to accurately obtain obstacle outlines and distance information; millimeter-wave radar is installed on both sides of the front and rear bumpers to collect target speed and distance data, adapted for high-speed moving target detection; cameras are installed at the rearview mirror position on the windshield and around the vehicle body to collect image data, used to identify traffic lights, lane lines and obstacle types; ultrasonic sensors are installed on the front and rear bumpers to collect data on nearby obstacles within a range of 0.1-5 meters, used for obstacle avoidance assistance in low-speed scenarios; By using the time synchronization protocol of vehicle Ethernet and CAN bus, the timestamp accuracy of each sensor data is calibrated to the microsecond level, ensuring the spatiotemporal consistency of multi-source data at the same time and avoiding data misalignment caused by acquisition delay.

[0032] (2) Raw data processing and acquisition of dynamic environmental information, as detailed below: The lidar point cloud data is filtered to remove ground noise, then downsampled by voxel filtering, and finally smoothed by Kalman filtering to remove obstacle position and velocity data; the millimeter-wave radar data is filtered by median filtering to remove random noise; and the camera image data is filtered by Gaussian filtering to remove salt-and-pepper noise and improve subsequent recognition accuracy. Based on the timestamp of the lidar, and taking into account the position offset, angle deviation, focal length, and pixel size of each sensor, the data from the millimeter-wave radar, camera, and ultrasonic sensor are mapped to a unified vehicle coordinate system through coordinate transformation to achieve spatial alignment; for data with temporal delays, linear interpolation is used to complete the time and ensure the temporal consistency of the data. A joint probabilistic data association algorithm is adopted to match targets detected by multiple sources of sensors based on the characteristics of obstacles such as position, speed and size. Duplicate targets are eliminated and missing targets are added. Finally, the output includes environmental dynamic information containing the unique identifier of the obstacle, real-time position, speed, acceleration and motion trend.

[0033] (3) Generation of semantic information about the traffic environment, as detailed below: Static information such as road network topology, lane line type, traffic signs, and road curvature is extracted from high-precision maps, and a static raster map is generated according to a preset raster size. Obstacles in the dynamic environmental information are semantically labeled by category and overlaid with a static grid map. Static road attributes, dynamic obstacle semantics, and obstacle movement status are recorded in the grid to generate a grid map or topology map containing multi-layer semantic information, which clarifies the passable areas, constraint rules, and dynamic risk sources for vehicle travel.

[0034] (4) Real-time energy consumption index calculation, as follows: Based on the universal characteristic curve of heavy-duty vehicle engines, a mapping model of engine speed, torque and instantaneous fuel consumption rate is established; combined with the whole vehicle dynamics model, parameters such as vehicle mass, drag coefficient, rolling resistance coefficient and transmission efficiency are input to construct a calculation model of vehicle driving resistance and power demand. The vehicle's CAN bus collects real-time vehicle status parameters such as current engine speed, torque, transmission gear, vehicle speed, and acceleration, as well as driving command parameters such as steering wheel angle, throttle opening, and brake pedal travel. The collected real-time parameters are input into the engine fuel consumption rate mapping model to calculate the instantaneous fuel consumption rate. Based on the integral of the instantaneous fuel consumption rate and driving time, combined with the vehicle dynamics model correction, the cumulative energy consumption is calculated to form a real-time energy consumption index.

[0035] (5) Vectorized encoding and fusion, specifically implemented as follows: The vehicle's own state (position (x,y), heading angle θ, speed V, etc.), environmental dynamic information (obstacle position, speed, etc.), traffic environment semantic information (road curvature, traffic sign coding, etc.), and real-time energy consumption indicators (instantaneous fuel consumption rate, cumulative energy consumption) are standardized and then converted into fixed-dimensional vectors; the semantic information is converted into numerical vectors using one-hot encoding or word embedding. By combining splicing fusion with attention-weighted fusion, vectors of various dimensions are spliced ​​together to form an initial fusion vector. Through the attention mechanism, key obstacles in the dynamic environmental information and core constraint rules in the semantic information of the traffic environment are given higher weights, ultimately forming a multi-dimensional state space vector with fixed dimensions and prominent features, which is used as the input of the decision model.

[0036] S2. Based on the multidimensional state space, construct and train the decision model. The training process includes: The output of the decision model is defined as a hierarchical action space, which includes a high-level action space for path planning and a low-level action space for vehicle control. Define a comprehensive reward function consisting of a weighted average of security rewards, efficiency rewards, and energy-saving rewards; the specific formula for calculating the comprehensive reward function is as follows:

[0037] in, These are the weighting coefficients for security rewards, efficiency rewards, and energy-saving rewards, respectively, and they satisfy the following conditions: ; Security rewards The calculation formula is:

[0038] in, The distance between the vehicle and the nearest obstacle; This is the safe distance threshold; This is the collision distance threshold; The reward value for safe driving; This is the penalty value when a collision occurs; This is the distance penalty coefficient; High-efficiency rewards The calculation formula is:

[0039] in, For the desired speed; This represents the travel time already elapsed; The preset maximum travel time; and These are the reward weighting coefficients for speed and time, respectively. Energy efficiency bonus The calculation formula is:

[0040] in, Energy consumption at the current moment; This is the preset maximum energy consumption; This is the weighting coefficient for energy-saving rewards.

[0041] It should be noted that the distance penalty coefficient Speed ​​reward weighting coefficient Time reward weighting coefficient and energy-saving reward weighting coefficient The determination method adopts a two-stage mechanism of offline calibration and online adaptive adjustment. The specific steps are as follows: 1. Steps for offline calibration of basic coefficients: Real-world driving data of heavy-duty vehicles was collected under typical scenarios (congestion, highways, and ordinary roads), covering safe driving samples, efficient driving samples, and energy-saving driving samples under different vehicle speeds, obstacle distances, and road gradients. Each sample included key parameters such as obstacle distance, actual vehicle speed, driving time, and energy consumption. Simultaneously, safe distance thresholds were determined based on heavy-duty vehicle design parameters (such as vehicle dimensions, braking performance, and engine power). Collision distance threshold Preset maximum driving time Preset maximum energy consumption Basic parameters; A multi-objective optimization model is constructed, with constraints including a safety constraint satisfaction rate of ≥99%, a high efficiency indicator achievement rate of ≥95%, and an energy saving rate improvement of ≥10% compared to the baseline strategy. The objective is to maximize the fit between the comprehensive reward function and the actual driving performance. A genetic algorithm is used to solve for the initial coefficient combination. Distance penalty coefficient The initial value range is 5-15, and its calibration must ensure that the distance between the vehicle and the obstacle is close to the collision threshold. Security rewards Reduced to a preset penalty threshold (not higher than) 50 Furthermore, the gradient of the reward change matches the response characteristics of the braking system; Speed ​​reward weighting coefficient The initial value range is 0.3-0.6, and the time reward weighting coefficient is... The initial value range is 0.4-0.7, and satisfies... + =1, calibration must ensure the vehicle is at the desired speed. High efficiency reward during driving The maximum value is reached, and here the maximum value is no less than (0.8). +0.8 ); Energy saving reward weighting coefficient The initial value range is 0.2-0.5, and calibration must ensure that the vehicle's energy consumption reaches the optimal range (not exceeding 0.3). Energy-saving incentives Not less than 0.7 And when energy consumption exceeds the standard (above 0.8), The reward is negative; Substitute the initial coefficients into the simulation environment and verify the model performance in a test scenario that includes complex obstacle interaction and scene switching. If problems such as insufficient safety distance, unsatisfactory efficiency, or poor energy-saving effect occur, fine-tune the coefficients according to the principle of safety first and efficiency and energy saving coordination until the preset performance indicators are met. 2. Steps for online adaptive adjustment of coefficients: In actual vehicle operation, the coefficient adjustment will be initiated when any of the following conditions are met: ① The percentage of times the safety reward is negative exceeds 5% within 500 consecutive decision cycles; ② The average efficiency reward is lower than the preset efficiency threshold (not higher than 60% of the initial calibration maximum value); ③ The average energy-saving reward is lower than the preset energy-saving threshold (not higher than 60% of the initial calibration maximum value); ④ The vehicle driving scenario changes (such as from a highway scenario to a congested scenario) and continues for 30 decision cycles. Dynamic adjustment rules: When close-range avoidance occurs repeatedly ( < <0.5 When increasing the value by 0.5-1.0 each time,... This continues until such scenarios account for less than 1%; when there is no close-range risk for a long period and the safety reward is excessively high (average above 0.9). When ), decrease by an increment of 0.3-0.8 each time. Balancing safety and efficiency; In high-speed scenarios, if the actual vehicle speed is lower than the expected speed 80% and continuing for 20 decision cycles, increasing (0.05-0.1 each time), decrease (0.05-0.1 each time); In ordinary roads or congested scenarios, if the driving time exceeds 120%, increase (0.05-0.1 each time), decrease (0.05-0.1 each time), always maintain + =1; In congested scenarios, if actual energy consumption exceeds 150% of the optimal energy consumption range, increase... (0.03-0.08 each time); In high-speed scenarios, if the energy-saving reward is too high, causing the vehicle speed to be lower than the expected speed, reduce... (0.03-0.08 each time); After each adjustment, the coefficients must be maintained within ±30% of the initial value range of the offline calibration to avoid extreme values ​​causing model instability. If the comprehensive performance indicators (safety, efficiency, energy saving) meet the preset requirements for 1000 consecutive decision cycles after adjustment, the current coefficients will be fixed as the benchmark values ​​for subsequent adjustments.

[0042] A scene recognition-based dynamic reward weight adjustment mechanism is introduced, namely: during model training, the weight of each reward item in the comprehensive reward function is dynamically adjusted according to the real-time identified driving scene type; with the goal of maximizing the comprehensive reward function, the decision model is driven to train and learn. In this step, the constructed decision model adopts the Actor-Critic reinforcement learning architecture, where: Actor networks, as policy networks, are used to generate specific actions in a hierarchical action space based on a multidimensional state space. The Critic network, as a value network, is used to evaluate the expected long-term cumulative reward that can be obtained by performing the action in the current state; The decision model is trained through collaborative iterative optimization using an Actor network and a Critic network; the steps of the collaborative iterative optimization include: (a) In a simulated or real environment, the decision model performs actions based on the Actor network, obtains new states and immediate rewards after interacting with the environment, and stores the experience data of the interaction in the experience replay pool. (b) Sample a batch of empirical data from the empirical replay pool, calculate the value estimation error of the current Critic network using the advantage function, minimize the error using the gradient descent algorithm, and update the parameters of the Critic network. (c) Based on the value assessment provided by the updated Critic network, calculate the performance gradient of the Actor network and update the parameters of the Actor network along the gradient direction so that the policy tends to select actions that can obtain higher long-term cumulative rewards. (d) Repeat steps (a) to (c) until the policy performance of the decision model converges or reaches the preset training objective.

[0043] It should be noted that the specific steps for dynamically adjusting the reward weights based on scene recognition and training the decision model with the goal of maximizing the comprehensive reward function during model training are as follows: 1. Training initialization steps: The decision model parameters of the Actor-Critic architecture are initialized, including the weights of the policy generation layer of the Actor network and the weights of the value evaluation layer of the Critic network. The initialization range is set to [-0.01, 0.01]. Preset safe distance threshold Collision distance threshold Expected speed Preset maximum driving time Preset maximum energy consumption These basic parameters are determined based on the performance parameters of heavy-duty vehicles and the requirements of driving scenarios. Create an experience replay pool and a priority experience replay pool to store experience data (state-action-reward-next state) generated by the interaction between the model and the environment.

[0044] 2. Scene recognition and dynamic weight adjustment steps: During training (simulation environment or real vehicle test environment), scene feature data is collected at a frequency of 50-100ms / time, including traffic density (number of vehicles per unit area), average vehicle speed, and road type (highway / ordinary / congested road section), which is obtained synchronously with the simulation environment parameters through sensor data. The scene is determined based on the collected feature data. The conditions for determining a congested scene are traffic density ≥ 0.7 and average vehicle speed ≤ 30km / h; the conditions for determining a highway scene are that the road type is highway, the average vehicle speed is ≥ 80km / h, and the traffic density is ≤ 0.3; and the ordinary road scene is the case that does not meet the above two conditions. Adjust the weight of the comprehensive reward function based on the judgment results: ① Congested scenarios: =0.2 0.4 (security reward weight) =0.1 0.3 (Efficiency Reward Weight) =0.5 0.7 (Energy efficiency bonus weight); ② High-speed scenario: =0.1 0.3 =0.6 0.8 =0.1 0.3; ③ Ordinary road scenario: =0.4、 =0.4、 =0.2, and always satisfies + + =1; When switching scene types, linear interpolation is used to achieve a smooth transition of weights, with the transition period set to 3-5 training steps to avoid model training oscillations caused by sudden weight changes.

[0045] 3. Model interaction and experience storage steps: The Actor network generates hierarchical actions (high-level desired path point sequence, low-level steering wheel angle and acceleration commands) based on the current multidimensional state space vector and the implicit target with adjusted reward weights. These actions are then executed in a simulation or real-vehicle testing environment. After execution, a new state vector and an immediate comprehensive reward calculated based on the current scene weights are collected. The system also includes a flag indicating whether the training termination condition has been met (such as reaching the target position or a collision has occurred); the complete experience data of the current state, generated action, immediate comprehensive reward, new state, and termination flag is stored in the experience replay pool; at the same time, the experience priority is determined based on the absolute value of the immediate comprehensive reward, and high-value experience (reward absolute value ≥ preset threshold) is stored in the priority experience replay pool.

[0046] 4. Model parameter iterative update steps: Randomly sample batch data from the experience replay pool (batch size set to 32-128), and simultaneously sample 10%-20% of high-value data from the priority experience replay pool according to priority, and combine them to form training batch data; Based on the training batch data, through the advantage function Calculation, where Let the action value function be... Let be the state value function. Calculate the value estimation error of the Critic network, minimize the error using the gradient descent algorithm, update the Critic network parameters, and set the learning rate to 1e-4-5e-4; Based on the state value evaluation of the updated Critic network output, the policy performance gradient of the Actor network is calculated, and the network parameters are updated along the gradient direction to make the policy tend to choose actions that can obtain higher long-term cumulative comprehensive rewards. The learning rate is set to 1e-5-5e-5. During parameter updates, L2 regularization constraints are applied to the network weights to avoid overfitting; adversarial example training is introduced to add small perturbations to the state vector and improve the model's robustness.

[0047] 5. Training termination and model solidification steps: Every 1000 iterations, the average comprehensive reward value and policy variance of the most recent 100 iterations are calculated. When the average comprehensive reward value fluctuates by ≤1% for three consecutive iterations and the policy variance is ≤ a preset threshold, the model training is considered to have converged. To validate the model's performance in a mixed test scenario that includes congestion, highways, and ordinary roads, the following conditions must be met: safe collision rate ≤ 0.1%, average driving efficiency ≥ 85% of the desired speed, and average energy consumption ≤ 80% of the baseline strategy. Otherwise, return to step 3 to continue training. After the performance validation is passed, solidify the network parameters of the current decision model and generate a deployable model file.

[0048] 6. Advanced training steps for course learning: The learning strategy employs a course-based approach, first training the model to convergence in simple scenarios (straight roads without obstacles, fixed-speed cruising), and then gradually introducing complex scenarios (multiple obstacles, changes in road curvature, frequent scene switching) to improve the model's adaptability to complex environments. Generalization training is conducted in scenarios with different road conditions (slope, road surface friction coefficient) and different traffic flow intensities. For each new scenario, steps 2-5 are repeated for targeted iteration to ensure that the model can maximize comprehensive rewards in diverse scenarios.

[0049] S3. Deploy the trained decision model to the on-board computing unit for real-time path planning and vehicle control; S4. During actual vehicle operation, driving data is continuously collected, and the parameters of the decision-making model are updated based on the incremental learning mechanism; at the same time, an energy consumption-road condition correlation database is built and continuously updated based on the collected data. In this step, the steps for updating the decision model parameters based on the incremental learning mechanism include: (a) During actual vehicle operation, real-time data on vehicle status, environmental perception, control commands, energy consumption and reward feedback are collected synchronously at a preset sampling frequency. The data is then processed by removing outliers, filling in missing values ​​and aligning time sequences to generate standardized data samples. (b) The standardized data samples are temporarily stored in the first experience pool according to the driving scenario classification; the data in the first experience pool are periodically evaluated and ranked based on the potential value of each data point for improving the current decision-making model, and high-value data is selected and transferred to the second experience pool using a priority storage strategy. (c) When the preset triggering conditions are met, the incremental learning process is started, and data is extracted from the first experience pool and the second experience pool in a preset ratio to form a training set. The parameters of the currently deployed decision model are used as the initial values ​​for small-batch training, and the convergence of the model's comprehensive reward loss function is used as the training termination condition to obtain the updated model parameters. The triggering conditions include at least the new data volume reaching a threshold, the model performance continuously declining, or the detection of new unrecorded scenarios. (d) Perform offline playback verification and online restricted trial operation on the updated model parameters; only when the verification and trial operation results meet the preset performance and safety standards will the updated model parameters be deployed to the vehicle computing unit to replace the original model parameters; if they do not meet the standards, the original model parameters will be rolled back and the corresponding data will be added to re-trigger incremental learning.

[0050] In this step, the establishment of the energy consumption-road condition correlation database includes: The database storage structure is defined, using a combination of road slope intervals, traffic flow state levels, and vehicle speed ranges as the scenario index key, and the corresponding optimal energy consumption strategy parameters and execution effect evaluation values ​​as the associated data values, and an index is established. Specifically, the road slope intervals are divided according to a preset gradient, the traffic flow state levels are divided into three levels—smooth, slow, and congested—based on real-time vehicle density, and the vehicle speed ranges are divided according to preset speed intervals. The optimal energy consumption strategy parameters include the target acceleration range and engine speed range; the execution effect evaluation values ​​include the actual energy consumption value and the tracking accuracy error. During vehicle operation, real-time road gradient, traffic flow status, vehicle speed, and corresponding vehicle energy consumption data and execution strategy parameters are collected under different driving scenarios, and the data is cleaned and classified according to the scenarios. For each type of scenario, the optimal energy consumption strategy is selected from multiple sets of historical data, with energy consumption as the core optimization target and driving safety and tracking accuracy as constraints. The optimal energy consumption strategy is associated with the corresponding scenario features and execution effect evaluation values ​​and stored in the energy consumption-road condition association database.

[0051] S5. During the real-time decision-making process, the energy consumption-road condition correlation database is queried based on the current road condition characteristics, and the output of the decision-making model is guided or fine-tuned using the retrieved historical energy-saving strategies, so as to achieve rapid scenario adaptation of energy-saving strategies.

[0052] In this step, the process of implementing scenario-adaptive energy-saving strategies based on the energy consumption-road condition correlation database includes: Define the core fields of the database, including road slope range (divided into gradient ranges in ±1° increments), traffic flow status level (divided into three levels: smooth, slow, and congested based on vehicle density), optimal energy consumption strategy parameters (including target acceleration range, engine speed range, and transmission shift logic threshold), applicable vehicle speed range for the strategy, and strategy execution effect evaluation value (including actual energy consumption value and tracking accuracy error). A key-value pair storage structure is adopted, with road slope range-traffic flow status level-vehicle speed range as the composite primary key, to construct a three-dimensional index system. During the initial deployment of the actual vehicle and during online operation, a complete data chain is continuously collected under different driving scenarios, including real-time road slope data obtained through onboard sensors, vehicle density and average vehicle speed data obtained through the traffic flow detection module, energy consumption-related data such as engine speed, fuel consumption rate, and transmission gear obtained through the vehicle CAN bus, as well as the current energy consumption strategy parameters output by the reinforcement learning model; the collected data is cleaned, outliers are removed, and the data is classified into corresponding gradient intervals and state levels according to preset scenario classification rules; For each scenario category, a multi-objective optimization algorithm is used to evaluate and rank multiple energy consumption strategies under the same scenario. With the lowest energy consumption as the core objective and the tracking accuracy meeting the safety threshold as the constraint, the optimal energy consumption strategy is selected. The parameters, scenario features and execution effect evaluation values ​​of the optimal strategy are associated and stored in the energy consumption-road condition association database, and the strategy generation time and cumulative application times are recorded. During vehicle operation, real-time data on current road slope, traffic flow status, and vehicle speed are collected. Scene matching is performed in the energy consumption-road condition association database using a three-dimensional indexing system. If a historical scene with a similarity of ≥90% exists, the optimal energy consumption strategy corresponding to that scene is directly called as the initial control strategy. If the similarity is between 70% and 90%, the optimal energy consumption strategy corresponding to that scene is extracted as the basic strategy. If the similarity is <70%, the initial energy consumption strategy is generated in real-time based on a reinforcement learning model. For the initial / basic strategy that is invoked or extracted, the strategy parameters are fine-tuned by an incremental learning algorithm, taking into account the real-time dynamic information of the current scenario (including changes in obstacle motion state, traffic light state switching, sudden changes in road curvature, etc.). During the fine-tuning process, the real-time feedback of the comprehensive reward function is used as the optimization target, and key parameters such as the target acceleration range and engine speed range are adjusted to ensure that the strategy maintains the optimal energy-saving effect while meeting the safety and efficiency requirements of the current scenario. The fine-tuned optimal strategy, current scenario data, and execution effect evaluation values ​​are added to the energy consumption-road condition association database. If the scenario already has historical records, the database content is updated according to the principle of overwriting when the new strategy has better execution effect. At the same time, the database is periodically clustered to merge scenario categories with high similarity, delete inefficient strategy data that has not been called for a long time, and optimize database storage efficiency and query speed.

[0053] like Figure 2 As shown, this embodiment of the invention also provides a heavy-duty vehicle route planning and tracking system that balances efficiency and energy saving indicators, including: The state space construction module is used to construct the multi-dimensional state space required for the decision model based on reinforcement learning. The multi-dimensional state space integrates the vehicle's own state, environmental dynamic information, traffic environment semantic information, and real-time energy consumption indicators. The model training module is used to construct and train a decision model based on the multidimensional state space. The training process includes: The output of the decision model is defined as a hierarchical action space, which includes a high-level action space for path planning and a low-level action space for vehicle control. Define a comprehensive reward function consisting of a weighted average of security rewards, efficiency rewards, and energy-saving rewards; A scene recognition-based dynamic adjustment mechanism for reward weights is introduced, namely: during model training, the weights of each reward item in the comprehensive reward function are dynamically adjusted according to the real-time identified driving scene type; The decision model is trained and learned with the goal of maximizing the comprehensive reward function. The model deployment and execution module is used to deploy the trained decision model to the on-board computing unit for real-time path planning and vehicle control. The online learning and database construction module is used to continuously collect driving data during real vehicle operation and update the parameters of the decision model based on the incremental learning mechanism; at the same time, it builds and continuously updates the energy consumption-road condition correlation database based on the collected data. The scenario-adaptive decision module is used to query the energy consumption-road condition association database based on the current road condition characteristics during the real-time decision-making process, and use the retrieved historical energy-saving strategies to guide or fine-tune the output of the decision model, so as to achieve rapid scenario adaptation of energy-saving strategies.

[0054] In some embodiments, the state space construction module includes: The data acquisition unit is used to collect raw data about the vehicle's surrounding environment through onboard lidar, millimeter-wave radar, cameras, and ultrasonic sensors. An environmental perception unit is used to filter, spatiotemporally align, and correlate the raw data to obtain dynamic environmental information including obstacle positions, speeds, and motion trends. The semantic fusion unit is used to perform semantic-level fusion of static information such as road network, lane lines and traffic signs from high-precision maps with the dynamic environmental information to generate traffic environment semantic information. The energy consumption calculation unit is used to calculate the instantaneous fuel consumption rate and cumulative energy consumption based on the universal characteristic curve of the engine and the vehicle dynamics model of heavy-duty vehicles, according to the current vehicle status and driving instructions, as a real-time energy consumption indicator. The state fusion unit is used to vectorize and fuse the vehicle's own state, environmental dynamic information, traffic environment semantic information, and real-time energy consumption indicators to form a multi-dimensional state space.

[0055] In some embodiments, the decision model built in the model training module adopts an Actor-Critic reinforcement learning architecture, wherein: Actor networks, as policy networks, are used to generate specific actions in a hierarchical action space based on a multidimensional state space. The Critic network, as a value network, is used to evaluate the expected long-term cumulative reward that can be obtained by performing the action in the current state; The model training module is configured to train the decision model through collaborative iterative optimization of the Actor network and the Critic network.

[0056] In some embodiments, the online learning and database construction module includes: The data collection and preprocessing unit is used to synchronously collect real-time data on vehicle status, environmental perception, control commands, energy consumption and reward feedback at a preset sampling frequency during actual vehicle operation, and to perform outlier removal, missing value filling and time sequence alignment on the data to generate standardized data samples. The experience pool management unit is used to temporarily store the standardized data samples according to driving scenarios in the first experience pool; periodically evaluate and sort the data in the first experience pool based on the potential value of each data point for improving the current decision-making model, and select high-value data to transfer to the second experience pool using a priority storage strategy. The incremental learning triggering and execution unit is used to start the incremental learning process when a preset triggering condition is met. It extracts data from the first experience pool and the second experience pool according to a preset ratio to form a training set, performs small-batch training with the currently deployed decision model parameters as initial values, and uses the convergence of the model's comprehensive reward loss function as the training termination condition to obtain the updated model parameters. The triggering condition includes at least the new data volume reaching a threshold, the model performance continuously declining, or the detection of new unrecorded scenarios. The model validation and update unit is used to perform offline playback validation and online restricted trial runs on the updated model parameters. Only when the validation and trial run results meet the preset performance and safety standards will the updated model parameters be deployed to the vehicle computing unit to replace the original model parameters. If the standards are not met, the system will revert to the original model parameters and supplement the corresponding data to re-trigger incremental learning.

[0057] In some embodiments, the online learning and database construction module further includes a database construction unit for establishing an energy consumption-road condition correlation database, configured as follows: The database storage structure is defined, using a combination of road slope intervals, traffic flow state levels, and vehicle speed ranges as the scenario index key, and the corresponding optimal energy consumption strategy parameters and execution effect evaluation values ​​as the associated data values, and an index is established. Specifically, the road slope intervals are divided according to a preset gradient, the traffic flow state levels are divided into three levels—smooth, slow, and congested—based on real-time vehicle density, and the vehicle speed ranges are divided according to preset speed intervals. The optimal energy consumption strategy parameters include the target acceleration range and engine speed range; the execution effect evaluation values ​​include the actual energy consumption value and the tracking accuracy error. During vehicle operation, real-time road gradient, traffic flow status, vehicle speed, and corresponding vehicle energy consumption data and execution strategy parameters are collected under different driving scenarios, and the data is cleaned and classified according to the scenarios. For each type of scenario, the optimal energy consumption strategy is selected from multiple sets of historical data, with energy consumption as the core optimization target and driving safety and tracking accuracy as constraints. The optimal energy consumption strategy is associated with the corresponding scenario features and execution effect evaluation values ​​and stored in the energy consumption-road condition association database.

[0058] In some embodiments, the scene adaptive decision-making module includes: The scene matching unit is used to generate current scene features based on the road slope, traffic flow status and vehicle speed collected by the vehicle in real time, and to retrieve similar historical scene records in the energy consumption-road condition association database; The strategy retrieval unit is used to select, call the best historical strategy, or use it as a reference for the basic strategy, or trigger the model to regenerate the initial strategy based on the retrieval results and the preset similarity threshold. The strategy fusion and fine-tuning unit is used to fine-tune the strategy parameters in real time by combining the initial or basic strategy determined based on historical records with the dynamic changes of the current environment, and generating the final execution strategy through an online optimization algorithm with the comprehensive reward function as the objective. The database update unit is used to supplement the energy consumption-road condition association database with the actual execution effect of the final execution strategy as a new record.

[0059] 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. A heavy vehicle path planning and tracking method that balances efficiency and energy saving indicators, characterized in that, The method comprises the following steps: S1, constructing a multi-dimensional state space required for a decision model based on reinforcement learning, the multi-dimensional state space integrating vehicle self-state, environmental dynamic information, traffic environment semantic information and real-time energy consumption indicators; S2, based on the multi-dimensional state space, constructing and training a decision model, the training process comprising: defining the output of the decision model as a hierarchical action space, the hierarchical action space comprising a high-level action space for path planning and a bottom-level action space for vehicle control; defining a comprehensive reward function composed of safety reward, efficiency reward and energy-saving reward weighted; introducing a reward weight dynamic adjustment mechanism based on scene recognition, that is, in the model training process, according to the real-time recognized driving scene type, the weights of each reward item in the comprehensive reward function are dynamically adjusted; maximizing the comprehensive reward function as the goal to drive the decision model to train and learn; S3, deploying the trained decision model to a vehicle-mounted computing unit for real-time path planning and vehicle control; S4, in real vehicle operation, continuously collecting driving data, updating the parameters of the decision model based on an incremental learning mechanism; at the same time, based on the collected data, constructing and continuously updating an energy consumption-road condition association database; S5, in the real-time decision process, querying the energy consumption-road condition association database according to the current road condition characteristics, and guiding or fine-tuning the output of the decision model by using the queried historical energy-saving strategy, so as to realize the rapid scene adaptation of the energy-saving strategy.

2. The path planning and tracking method for heavy-duty vehicles with efficiency and energy saving indicators according to claim 1, characterized in that, S1 specifically comprises: collecting the original data of the vehicle surrounding environment through the laser radar, millimeter wave radar, camera and ultrasonic sensor mounted on the vehicle; filtering, space-time alignment and data association are performed on the original data to obtain environmental dynamic information containing obstacle position, speed and motion trend; performing semantic-level fusion on the road network, lane line and traffic sign static information from the high-precision map and the environmental dynamic information to generate traffic environment semantic information; based on the engine universal characteristic curve and the vehicle dynamics model of the heavy truck, the instantaneous fuel consumption rate and the cumulative energy consumption are calculated as real-time energy consumption indicators according to the current vehicle state and driving instruction; vectorizing and coding the vehicle self-state, environmental dynamic information, traffic environment semantic information and real-time energy consumption indicators, and fusing them to form a multi-dimensional state space.

3. The path planning and tracking method for heavy-duty vehicles that takes into account efficiency and energy saving indicators according to claim 1, characterized in that, The specific calculation formula of the comprehensive reward function is: Wherein, respectively are weight coefficients of security reward, high efficiency reward and energy saving reward, and satisfy ; Security rewards The formula for calculating is: wherein, is a distance of the vehicle to the nearest obstacle; is a safety distance threshold; is a collision distance threshold; is a reward value when driving safely; is a penalty value when a collision occurs; is a distance penalty coefficient; Efficiency reward The calculation formula is: wherein, is the desired speed; is the elapsed time; is the preset maximum travel time; and are the reward weight coefficients for speed and time, respectively. Energy saving reward The calculation formula is: wherein, is the energy consumption at the current time; is the preset maximum energy consumption; is the energy-saving reward weight coefficient.

4. The path planning and tracking method for heavy-duty vehicles that takes into account efficiency and energy saving indicators according to claim 1, characterized in that, In S2, the decision model constructed adopts an Actor-Critic reinforcement learning architecture, in which: the Actor network acts as a policy network for generating specific actions in the hierarchical action space according to the multi-dimensional state space; the Critic network acts as a value network for evaluating the long-term cumulative reward expectation that can be obtained by executing the action in the current state; the decision model is trained through the collaborative iterative optimization of the Actor network and the Critic network; the steps of the collaborative iterative optimization comprise: (a) in a simulation or actual environment, the decision model executes an action based on the Actor network, obtains a new state and an immediate reward after interacting with the environment, and stores the experience data of this interaction into an experience replay pool; (b) sampling a batch of experience data from the experience replay pool, calculating the value estimation error of the current Critic network using the advantage function, and minimizing the error by gradient descent algorithm to update the parameters of the Critic network; (c) based on the value evaluation provided by the updated Critic network, calculating the performance gradient of the Actor network, and updating the parameters of the Actor network in the gradient direction to make the policy tend to select actions that can obtain higher long-term cumulative rewards; (d) repeating steps (a) to (c) until the policy performance of the decision model converges or reaches the preset training target.

5. The path planning and tracking method for heavy-duty vehicles with efficiency and energy saving indicators according to claim 4, characterized in that, using a merit function computing a value estimate error, the merit function by the formula where is an action value function, is a state value function.

6. The path planning and tracking method for heavy-duty vehicles that takes into account efficiency and energy saving indicators according to claim 5, characterized in that, In S2, the reward weight dynamic adjustment mechanism based on scene recognition specifically includes: The real-time recognized driving scene types at least include congestion scenes, high-speed scenes and ordinary road scenes; When the driving scene is identified as a congestion scene, the weight of the energy-saving reward in the comprehensive reward function is increased, and the weight of the high-efficiency reward is decreased; When the driving scene is identified as a high-speed scene, the weight of the high-efficiency reward in the comprehensive reward function is increased, and the weight of the energy-saving reward is decreased.

7. The path planning and tracking method for heavy-duty vehicles that takes into account efficiency and energy saving indicators according to claim 1, characterized in that, In S4, the step of updating the parameters of the decision model based on the incremental learning mechanism includes: (a) During the running of the real vehicle, real-time data of vehicle state, environmental perception, control command, energy consumption and reward feedback are synchronously collected at a preset sampling frequency, and the data are processed for outlier rejection, missing value filling and time alignment to generate standardized data samples; (b) The standardized data samples are temporarily stored in a first experience pool according to driving scene classification; the data in the first experience pool are periodically evaluated and sorted based on the potential value of each data for improving the current decision model, and the high-value data are filtered out and transferred to a second experience pool adopting a priority storage strategy; (c) When a preset trigger condition is met, an incremental learning process is started, data are extracted from the first experience pool and the second experience pool at a preset proportion to form a training set, small batch training is performed with the current deployed decision model parameters as the initial value, and the convergence of the comprehensive reward loss function of the model is taken as the training termination condition to obtain updated model parameters; wherein the trigger condition at least includes that the amount of new data reaches a threshold, the model performance continuously decreases, or a new unrecorded scene is detected; (d) The updated model parameters are verified offline and tested online; only when the verification and test results meet the preset performance and safety standards, the updated model parameters are deployed to the vehicle-mounted computing unit to replace the original model parameters; if not, the original model parameters are rolled back, and the corresponding data are supplemented to trigger the incremental learning again.

8. The path planning and tracking method for heavy-duty vehicles that takes into account efficiency and energy saving indicators according to claim 1, characterized in that, The steps of establishing the energy-consumption-road condition association database include: The storage structure of the database is defined, a combination of a road slope interval, a traffic flow state level and a vehicle speed range is taken as a scene index key, corresponding optimal energy consumption strategy parameters and an execution effect evaluation value are taken as associated data values, and an index is established; the road slope interval is divided according to a preset gradient, the traffic flow state level is divided into three levels of smooth, slow and congested according to real-time vehicle density, and the vehicle speed range is divided according to a preset speed interval; the optimal energy consumption strategy parameters include a target acceleration range and an engine speed interval; and the execution effect evaluation value includes an actual energy consumption value and a tracking accuracy error; In the vehicle operation, real-time road slopes, traffic flow states, vehicle speeds and corresponding vehicle energy consumption data and execution strategy parameters under different driving scenes are collected, and the data is cleaned and classified by scenes; For each type of scene, the optimal energy consumption strategy under the scene is selected from multiple groups of historical data, taking the energy consumption index as the core optimization target and taking driving safety and tracking accuracy as the constraint condition; The optimal energy consumption strategy is stored in association with the corresponding scene characteristics and the execution effect evaluation value in the energy consumption-road condition association database.

9. The path planning and tracking method for heavy-duty vehicles with efficiency and energy saving indicators according to claim 8, characterized in that, In S5, the steps of realizing scene adaptation of the energy-saving strategy based on the energy consumption-road condition association database include: According to the real-time collected road slope, traffic flow state and speed of the vehicle, the current scene characteristics are generated, and similar historical scene records are retrieved in the energy consumption-road condition association database; According to the retrieval result and a preset similarity threshold, the historical optimal strategy is selected and called, or is used as a reference for a basic strategy, or triggers the model to regenerate an initial strategy; For the initial strategy or the basic strategy determined based on the historical records, the strategy parameters are real-time fine-tuned by an online optimization algorithm with a comprehensive reward function as the target, to generate a final execution strategy; The actual execution effect of the final execution strategy is supplemented as a new record to the energy consumption-road condition association database.

10. A heavy vehicle path planning and tracking system that balances efficiency and energy saving indicators, characterized in that, It includes: A state space construction module for constructing a multi-dimensional state space required by a decision-making model based on reinforcement learning, which integrates vehicle state, environmental dynamic information, traffic environment semantic information and real-time energy consumption index; A model training module for constructing and training a decision-making model based on the multi-dimensional state space, and the training process includes: Defining the output of the decision-making model as a hierarchical action space, which includes a high-level action space for path planning and a bottom-level action space for vehicle control; Defining a comprehensive reward function composed of safety reward, efficiency reward and energy-saving reward weighting; Introducing a reward weight dynamic adjustment mechanism based on scene recognition, that is, in the model training process, the weights of each reward item in the comprehensive reward function are dynamically adjusted according to the real-time recognized driving scene type; Taking maximizing the comprehensive reward function as the target, the decision-making model is driven to train and learn; A model deployment and execution module for deploying the trained decision-making model to a vehicle-mounted computing unit for real-time path planning and vehicle control. An online learning and database construction module is configured to continuously collect driving data in real vehicle operation, and update parameters of the decision model based on an incremental learning mechanism; meanwhile, an energy consumption-road condition correlation database is constructed and continuously updated based on the collected data; A scenario adaptive decision module is configured to query the energy consumption-road condition correlation database according to current road condition characteristics in real-time decision making, and guide or fine-tune the output of the decision model by using the queried historical energy-saving strategies, so as to realize quick scenario adaptation of the energy-saving strategies.

Images