Self-evolutionary energy-saving control method for automobile train based on working condition adaptive hybrid expert dispatching

Abstract

The application provides a working condition adaptive hybrid expert scheduling automobile train self-evolution energy-saving control method, specific domain sub-experts with different working condition tendencies are respectively constructed for a torque control domain, a gear selection domain and a fan control domain, a multi-head strategy gate network is introduced to analyze scene state feature time sequence information of vehicle operation, complex working condition types are implicitly identified and control demand trends are predicted, and the current optimal sub-experts are dynamically scheduled and combined from an independent domain expert pool in parallel; the activated sub-experts of each domain synchronously output corresponding physical control instructions according to the instantaneous state features of the domain; the gate network and each sub-expert are end-to-end jointly trained with a reward function oriented by overall vehicle comprehensive energy efficiency and driving smoothness, and global collaborative optimization control of three actuators is realized. The application can continuously improve the overall vehicle energy efficiency and strategy adaptability through the self-evolution mechanism of cloud aggregation and vehicle end fine tuning under complex and variable real vehicle operation conditions.

Description

Technical Field

[0001] This invention belongs to the field of vehicle control technology, specifically relating to a self-evolving energy-saving control method for automobile trains with adaptive operating conditions and hybrid expert scheduling. Background Technology

[0002] With the in-depth development of energy-saving technologies for commercial vehicles, the application of data-driven artificial intelligence methods in the field of vehicle energy management has become an important trend. Existing technologies mainly have the following two architectures.

[0003] One approach is a global optimization method for a single deep reinforcement learning model, aiming to train a single agent to handle all driving scenarios. Specifically, it uses various real-time vehicle information (such as vehicle speed, engine speed, torque control depth, road gradient, etc.) and historical information over a period of time as input to the model; the model's output is the combined command of the actions of the various actuators that need to be controlled. In a simulated vehicle environment, the model repeatedly tries and learns according to a set goal, ultimately forming a fixed, universal control strategy. After deployment, the vehicle uses real-time sensor data, and the model directly calculates and outputs all control commands. However, this approach has the following drawbacks: 1) Curse of dimensionality and policy trivialization: Because it needs to output precise control values ​​from multiple actuators simultaneously, the model faces a high-dimensional action combination space, making the learning process difficult and hindering the stable acquisition of the optimal strategy. Furthermore, the capabilities of a single model will be averaged and compromised across all different and even conflicting control requirements, resulting in an inability to achieve expert-level optimal performance under any single operating condition; 2) Opaque decision-making process and difficulty in online updates: The model's decision-making logic is complex and opaque, making it difficult to analyze and intervene when abnormal control occurs. Furthermore, this approach typically only allows for offline training during the development phase. Once installed in a vehicle, the model remains unchanged, making it difficult to utilize new data generated during actual vehicle operation for safe and effective continuous optimization and upgrades.

[0004] Secondly, existing multi-controller switching systems based on deterministic rules and threshold judgments are modular solutions reliant on human experience. They divide the vehicle's operating conditions into a limited number of typical scenarios, and pre-design or calibrate a set of fixed control parameters or controllers for each scenario. During actual driving, the system judges a series of preset conditions in real time, categorizes the vehicle's state into a predefined scenario category, and directly calls and executes the fixed control strategy for that category. However, this approach has the following drawbacks: 1) Rigid scenario matching and switching oscillations: Real-world road conditions are complex and varied, and strict rules cannot accurately describe many transitional or mixed scenarios, leading to matching errors. Near the boundary between two preset scenarios, small fluctuations in sensor signals can cause frequent system transitions between different control strategies, resulting in control oscillations such as jerky power output and frequent gear shifts, affecting driving smoothness; 2) High development and maintenance costs and inability to adapt to changes: The system's performance depends entirely on the precision of the initial manual classification and the quality of parameter calibration, requiring significant manpower and resources for extensive road testing and manual calibration, resulting in a long development cycle. Once deployed, the system cannot be changed and lacks adaptive and self-evolving capabilities. Summary of the Invention

[0005] To address the shortcomings of existing single-intelligent models, such as mediocre control performance and black-box decision-making, as well as the rigidity of operating condition adaptation and high maintenance costs of multi-controller systems with fixed rules, and the problems of heavy high-dimensional state space computation, insufficient real-time policy performance, and difficulty in online continuous evolution faced by traditional reinforcement learning algorithms in the collaborative energy-saving control of multiple actuators in vehicles, this invention provides a self-evolving energy-saving control method for vehicle trains based on operating condition adaptation and hybrid expert scheduling. By constructing a cluster of expert intelligent agents capable of division of labor and cooperation, designing an understandable and trainable gating network decision-making mechanism, and establishing a closed-loop self-evolving learning loop, this method aims to achieve refined collaborative control and continuous optimization throughout the entire lifecycle of multiple actuators in vehicle trains, including torque control, gear shifting, and fan operation. While ensuring control real-time performance and vehicle dynamics, it significantly improves the overall energy efficiency of the vehicle under various complex operating conditions, while enabling the system to possess the intelligent capability of self-learning and self-evolving from data.

[0006] This application is implemented using the following technical solution: This invention provides a self-evolving energy-saving control method for vehicle-trains based on adaptive hybrid expert scheduling under specific operating conditions, comprising the following steps: S1, Constructing a self-evolving energy-saving control system with adaptive operating conditions and hybrid expert scheduling. S11. Construct deep reinforcement learning sub-experts for specific domains. Construct torque control sub-experts, gear selection sub-experts, and fan control sub-experts for the torque control domain, gear selection domain, and fan control domain, respectively. Each sub-expert takes the instantaneous features of the corresponding domain as input and outputs the corresponding physical control quantity. Each sub-expert constructs sub-expert experience data based on the instantaneous features, physical control quantity, and domain evaluation function value generated by its executed action, and conducts continuous training on the vehicle and cloud respectively. S12, a multi-head policy gating network is constructed based on the policy gradient method and convolutional neural network. The gating network takes vehicle and environment data as input and selects one activated torque control sub-expert, gear selection sub-expert, and fan control sub-expert in parallel from the independent domain expert pool to form a three-domain control combination of torque control, gear selection, and fan control adapted to the current scenario. The gating network constructs gating experience data based on the vehicle and environment data generated by its selection action, the joint policy decision action, and the multi-head policy gating network reward, and performs continuous training on the vehicle and cloud respectively. S2, Layered training in simulation environment: In the AMESim simulation environment, each domain sub-expert is independently pre-trained based on different road spectrums under different working conditions to form an independent domain expert pool; the gated network is integrated with the pre-trained domain sub-experts for collaborative training, and the initial policy network and parameters are obtained through iterative training; S3, Real-vehicle environment fine-tuning training and control: Using real data collected from the actual vehicle as input, the gating network outputs a combination of three-domain expert indices in real time. Each activated sub-expert synchronously outputs and executes corresponding physical control commands. After execution, the rewards for each sub-expert are calculated based on the global reward function and the domain evaluation function, generating single-vehicle experience data. This cumulatively forms the vehicle-side experience pool. Individual vehicles undergo incremental fine-tuning training based on the vehicle-side experience pool. S4, Cloud-based experience aggregation and continuous learning: Vehicles with the same configuration upload their respective vehicle-side experience pool data to the cloud, which is then aggregated to form a cloud-based historical experience database. The cloud uses this database to perform offline experience backtracking optimization training on the gating network and expert pools in each domain, and sends the updated strategy parameters to the vehicle via OTA when the driver authorizes the vehicle.

[0007] According to a further explanation of the present invention, preferably, step S11, which constructs the torque control sub-expert, gear selection sub-expert, and fan control sub-expert, specifically includes the following steps: S111, Establishing a torque control sub-expert based on a deep deterministic strategy gradient algorithm. 1) Environmental conditions: Torque control pedal, brake pedal, steering wheel information, vehicle speed, acceleration, motor status, engine status, altitude, and slope information; 2) Control action: The reference engine torque obtained by mapping the torque control pedal is proportionally adjusted, and the adjustment range is ±5% of the reference engine torque; 3) Torque control sub-expert architecture: Construct an Actor-Critic network and its corresponding target network to establish the mapping relationship between vehicle state and torque control actions; 4) Design the reward function for the torque control sub-expert: (1); in, Indicates torque control; For a moment The environmental conditions; For a moment Torque control action; The environmental state at the next moment; From state to state The driving distance; This represents the cumulative fuel consumption during the state transition. This represents the average rate of change of torque control during the state transition process. The positive weighting constant for torque control, which is used to balance the goals of encouraging vehicle movement and reducing fuel consumption, is an energy consumption incentive. A positive weighting constant is used to reward torque control smoothness and to ensure the target of control smoothness; S112, Establishing a gear selection sub-expert based on a deep Q-network algorithm 1) Environmental conditions: vehicle speed, engine speed, current gear, torque control pedal, brake pedal, steering wheel information, torque adjustment command; 2) Control actions: Discrete gear adjustment decisions, including upshifting two gears, upshifting one gear, maintaining the current gear, downshifting one gear, and downshifting two gears; 3) Gear selection sub-expert architecture: master Q network and target Q network, used to establish the mapping relationship between vehicle operating status and gear selection; 4) Design the expert reward function for gear selection: (2); in, Indicates gear selection; For a moment The environmental conditions; For a moment The gear selection action; The environmental state at the next moment; This refers to the engine output power at the next moment; A positive weighting constant is selected for the energy consumption reward level to balance the goals of encouraging vehicle movement and reducing fuel consumption; It is a positive weighting constant for motivational rewards, used to ensure the goals of motivation; S113, Establishing a fan control sub-expert based on a deep deterministic policy gradient algorithm. 1) Environmental conditions: vehicle speed, acceleration, motor status, engine status, cooling system status information, ambient temperature, altitude, torque adjustment command, gear command; 2) Control Action: Continuous fan speed commands within the range of 0 to the maximum fan speed; 3) Fan control sub-expert architecture: Construct an Actor-Critic network and its corresponding target network to establish the mapping relationship between the overall thermal state, operating state and fan control actions; 4) Design the expert reward function for the fan control sub-system: (3); in, Indicates fan control; For a moment The environmental conditions; For a moment Fan control actions; The environmental state at the next moment; This refers to the current coolant temperature. The preset optimal operating coolant temperature for the engine; This is the upper limit of the coolant temperature, used to ensure thermal safety; This is a positive weighting constant for the optimal temperature control target reward. A positive weighting constant for thermal safety rewards, used to control the coolant temperature within the optimal range while ensuring thermal safety.

[0008] According to a further description of the present invention, preferably, step S21 specifically involves: constructing a multi-head policy gating network based on the policy gradient method and convolutional neural network, acquiring vehicle and environmental data and inputting it into the gating network; the gating network uses a convolutional neural network to encode the time-series vehicle and environmental data, extract high-dimensional state features, and the multi-head policy gating network generates three sub-expert selection action probability distributions for torque control, gear selection, and fan control from the expert pool, and then makes sub-expert control combination decisions based on the action probability distribution results, and executes joint control actions to collect reward feedback; The vehicle and environmental data includes driver commands, vehicle operating status, and road and environmental information; the driver commands include torque control pedal, brake pedal, and steering wheel information; the vehicle operating status includes vehicle speed, acceleration, motor status, engine status, transmission status, and cooling system status information; the road and environmental information includes GPS signal, ambient temperature, altitude, and slope information.

[0009] According to a further explanation of the present invention, preferably, step S11, which constructs a multi-head policy gating network, specifically includes the following steps: S121, Input vehicle and environmental data characteristics (4); in, The input is a time series matrix; For a moment Vehicle and environmental data feature vectors; T The set time window length; F The number of feature dimensions, including the vehicle and environment data features; for A dimensional matrix space; This represents the feature vector of vehicle and environmental data at the initial moment; S122, Feature extraction based on CNN (5); in, This represents the state at time t after processing by the CNN convolutional neural network. For A convolutional neural network with parameters; The internal parameters of the convolutional neural network (including the number of neural network layers, the number of neurons, the weights of the convolutional kernel, and the bias) are obtained through pre-training. This patent does not limit its specific structure; any CNN structure that can achieve feature extraction is acceptable.

[0010] S123, Constructing a Multi-Head Policy Gated Network (6); in, For the system at time The policy probability distribution of a multi-head policy gating network; For the system at time The joint strategy decision-making action represents a set of sub-expert control combinations for decision-making; For the system at time Choose expert Chi Qizi. The policy probability distribution of the action; For Sub-experts as parameters The strategy head corresponds to the fully connected layer neural network; The internal parameters of the fully connected neural network (including the number of neural network layers, the number of neurons, weights, and biases) are obtained through pre-training, and this patent does not limit its specific structure. This refers to the expert pool, which consists of torque control, gear selection, and fan control. This refers to any one of the three sub-experts: torque control, gear selection, or fan control. S124, Design the reward function for a multi-head policy gated network. (7); in, For a moment Torque control expert reward function; For a moment Gear selection expert reward function; For a moment The expert reward function for the fan control sub-system; The sum of the rewards for the three sub-experts represents the energy efficiency reward; The reward is used to constrain the smoothness of system operation, representing the number of expert strategy switching per unit travel distance; Deadline The cumulative driving distance; Switch indicator variables for experts; For a moment The selected sub-expert control group number; , These are the weighting constants for energy efficiency rewards and smoothness rewards, respectively.

[0011] According to a further description of the present invention, preferably, step S2 specifically includes the following steps: S21, Independent Pre-training of Intelligent Agents: Simulation models of the entire vehicle, engine, transmission, and thermal management system are built in the AMESim simulation environment. Road spectrum data for different user operating conditions is extracted based on vehicle network big data. This road spectrum data includes time-vehicle speed, mileage-altitude, and time-torque control timing feature data. Different road spectrum data are used as input to the simulation model. Based on the deep learning framework in step S12, independent pre-training is performed on the torque control, gear selection, and fan control sub-experts to form sub-experts adapted to the operating condition features, which are then combined to form an expert pool. For the torque control and fan control sub-experts, whose action spaces are continuous variables, the DDPG algorithm is used for training. For the gear selection sub-expert, whose action space is discrete variables, the DQN algorithm is used for training. S22, a two-stage strategy is adopted to train the multi-head policy gating network: First, while keeping the parameters of each sub-expert fixed, the parameters of the gating network are quickly trained based on the multi-head policy gating network in step S1 to obtain an initial policy model; then, the joint collaborative training stage is entered, in which the parameters of the gating network and all sub-experts are updated synchronously at the same time, and the initial policy network and parameters are formed through iterative training.

[0012] According to a further explanation of the present invention, preferably, step S3 specifically involves: in a real vehicle operating environment, using real vehicle data as input, and making real-time decisions on torque control, gear selection, and fan control sub-expert control combined actions through a multi-head strategy gating network; The torque control sub-expert generates the optimal engine torque requirement based on the environmental conditions described in step S111, establishing a basic control benchmark for the entire control system. The gear selection expert combines the optimal engine torque requirement with the environmental conditions described in step S112 to select the most economical gear, so that the engine always operates in the high-efficiency range. Finally, the fan control sub-expert integrates the control actions output by the torque control sub-expert and the gear selection sub-expert, as well as the environmental state described in step S113, to finely adjust the fan speed, minimizing accessory power consumption while ensuring thermal safety. Each sub-expert focuses on different control objectives, but they form a closed-loop collaborative expert execution network through their dependencies. The current environmental state, actions, rewards, and the next environmental state are each vector-combined to serve as the empirical data generated by a single vehicle performing an action. Based on single-bike experience data The accumulated vehicle-side experience pool Experience-based fine-tuning training is conducted on individual vehicles.

[0013] According to a further description of the present invention, preferably, step S4 specifically involves: pooling the vehicle-side experience generated during actual operation of the market and configured vehicles. Securely uploaded to the cloud to form a cloud-based historical experience database, enabling large-scale aggregation of operational data from the fleet. Periodically retrained and fine-tuned torque control, gear selection, fan control sub-experts, and multi-head strategy gating networks are then performed in the cloud, allowing the model to learn richer driving conditions and adapt to a wider range of usage patterns and operating environments. Subsequently, the updated model and parameters are redeployed to each vehicle terminal via OTA remote online upgrades, achieving continuous optimization and adaptive improvement of the entire fleet's control performance.

[0014] The second aspect of this invention provides a self-evolving energy-saving control system for vehicle trains with adaptive operating conditions and hybrid expert scheduling, including a multi-head strategy gating network, a simulation environment hierarchical training module, a real vehicle environment fine-tuning training and control module, and a cloud-based experience aggregation and continuous learning module. The multi-head strategy gating network is used to decide on the appropriate sub-expert control combination from the torque control, gear selection, and fan expert pools based on the current operating condition characteristics. It establishes the input operating condition state, action, and reward function corresponding to the gating network for iterative training of the gating network model parameters. Based on a deep reinforcement learning framework, expert pools are constructed for torque control, gear selection, and fan control respectively. Each expert pool contains multiple sub-expert models associated with vehicle operating condition characteristics. The environmental state, action, and reward function corresponding to each sub-expert are established for iterative training of each sub-expert. The input layer signal acquisition module completes the selection and control decision of each sub-expert, and the actuators adjust the torque, gear, and fan, while simultaneously fine-tuning the model based on action feedback. The simulation environment hierarchical training module is used to implement the initial version of torque control, gear selection and fan control sub-expert pre-training in the simulation environment, as well as to perform collaborative training on the multi-head strategy gating network. The real-vehicle environment fine-tuning training and control module is used to complete the selection and control decisions of each sub-expert in a real-vehicle environment, and at the same time generate single-vehicle experience data. This is used for continuous iterative upgrades of the model; The cloud-based experience aggregation and continuous learning module includes a cloud server for collecting vehicle-side experience pools generated by multiple vehicles during actual operation. The data is uploaded to the cloud to form a cloud-based historical experience database. Based on this database, the torque control, gear selection, and fan control sub-experts, as well as the multi-head strategy gating network, are periodically retrained and fine-tuned in the cloud. The updated model is then deployed to the vehicle via OTA (Over-The-Air) remote online upgrade when requested by the driver.

[0015] Preferably, the actuator accepts control decisions from a multi-head strategy gating network and an expert pool, completes control command actions, provides feedback on the execution results, and generates single-vehicle experience data. ; The input layer signal acquisition module is used to acquire vehicle and environmental data; the vehicle and environmental data includes driver commands, vehicle operating status, road and environmental information; the driver commands include torque control pedal, brake pedal, and steering wheel information; the vehicle operating status includes vehicle speed, acceleration, motor status, engine status, transmission status, and cooling system status information; the road and environmental information includes GPS signal, ambient temperature, altitude, and slope information.

[0016] Preferably, the torque control sub-expert is used to generate the optimal engine torque demand and establish a basic control benchmark for the entire control system; The gear selection expert is used to select the most economical gear so that the engine always operates in the high-efficiency range. The fan control sub-expert is used to finely adjust the fan speed and minimize the power consumption of accessories while ensuring thermal safety.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention aims to address the core problems of traditional reinforcement learning algorithms in the high-dimensional computational environment of vehicle energy-saving control, such as the curse of dimensionality, insufficient real-time policy implementation, and inability to continuously learn and evolve online. Specifically addressing the problem of collaborative energy-saving control of vehicle actuators, it proposes a self-evolving energy-saving control method and system for vehicle trains based on operating conditions and hybrid expert scheduling. This system assigns specific domain sub-experts with different policy preferences to each actuator, including torque, fan, and gear, and introduces a trainable gating network. This network analyzes the temporal information of the vehicle's state, implicitly identifying complex operating conditions and predicting demand trends, thereby dynamically scheduling and combining the optimal dedicated sub-experts for each actuator. The gating network and each sub-expert are jointly trained end-to-end through a reward function targeting overall vehicle energy efficiency and drivability, ultimately achieving global collaborative optimization of control strategies for multiple actuators. This enables global collaborative real-time control of multiple actuators such as torque control, gear, and fan, and continuously improves overall vehicle energy efficiency and policy adaptability under complex and changing real-world operating conditions through a self-evolving mechanism of cloud aggregation and on-board fine-tuning. Attached Figure Description

[0018] Figure 1 The flowchart shows the self-evolving energy-saving control method for vehicle trains based on adaptive hybrid expert scheduling according to the present invention. Figure 2 This is a framework diagram of the self-evolving energy-saving control system for automobile trains based on the adaptive hybrid expert scheduling of the present invention. Detailed Implementation

[0019] Terminology Explanation: Policy Gradient (PG): A method that uses a neural network to map policy actions.

[0020] Convolutional Neural Networks (CNN): A deep learning architecture specifically designed for processing grid-like data.

[0021] OTA (Over-the-Air): A remote online upgrade method.

[0022] Multi-head policy network: A neural network architecture in which the output layer has multiple parallel branches.

[0023] Deep Deterministic Policy Gradient (DDPG): A model-free reinforcement learning algorithm for continuous action spaces.

[0024] Deep Q-Network (DQN) algorithm: a model-free reinforcement learning algorithm that combines deep learning with Q-learning.

[0025] AMESim Model: A modeling and simulation platform for complex systems across multiple disciplines.

[0026] Example 1: like Figure 1 The flowchart of the self-evolving energy-saving control method for vehicle-trains based on adaptive hybrid expert scheduling under operating conditions is shown. The self-evolving energy-saving control method for vehicle-trains based on adaptive hybrid expert scheduling under operating conditions of the present invention specifically includes the following steps: S1, Construct a self-evolving energy-saving control system with condition-adaptive hybrid expert scheduling for outputting vehicle train power system control strategies; S11, Constructing Domain-Specific Deep Reinforcement Learning Sub-Experts. For the three domains of torque control, gear selection, and fan control, separate sub-experts are constructed for torque control, gear selection, and fan control, respectively. Each sub-expert receives the instantaneous features required for its domain, i.e., high-refresh-rate vehicle dynamics parameters, such as the instantaneous engine speed, actual torque, and rate of change of required torque input in the torque domain. During training, the sub-expert's state space consists of instantaneous features, its action space consists of physical control quantities such as throttle adjustment coefficients, target gear, or fan speed, and its reward function is the evaluation of local energy consumption and ride comfort within its domain. Through interactive iteration in a simulation environment, each sub-expert learns the optimal control strategy for specific operating conditions, forming a domain skill library that can be scheduled. S12, Construct a multi-head strategy gating network. This network receives vehicle scene features, including driver instructions, vehicle operating status, road and environmental information within a preset time window. Based on the scene features, the gating network selects an activation index from each of the torque, gear, and fan expert pools in parallel, forming a three-domain control combination adapted to the current scene. During training, its state space is the scene features, its action space is the expert index vector, and its reward function is the overall vehicle energy consumption and timeliness evaluation. Through interactive iteration, the gating network gradually learns the scheduling capability of selecting the optimal sub-expert for each domain under different scenarios. S2, Layered Training in Simulation Environment: In the AMESim simulation platform, firstly, for different typical operating conditions, sub-experts in torque control, gear selection, and fan control are independently pre-trained. In this stage, each sub-expert takes its domain-specific instantaneous characteristics as input and its domain-specific physical control quantity as output. Through interactive iteration with the local energy consumption and ride comfort evaluation functions of its domain, specific domain expert pools are gradually formed, each covering different operating condition characteristics. Subsequently, the multi-head strategy gating network is integrated with the pre-trained three-domain expert pools for end-to-end collaborative training. In this stage, the parameters of each sub-expert are frozen or updated with a low learning rate. The gating network takes vehicle scene state characteristics as input and outputs an expert index selection vector. Through interactive iteration with the vehicle's comprehensive energy consumption and timeliness evaluation functions, the gating network's selection and scheduling capabilities are optimized. Through the above layered training and simulation verification iterations, a preliminary strategy network and model parameters are obtained that can achieve effective combination decision-making of specific domain expert pools under comprehensive operating conditions, providing an initial control strategy for subsequent real-vehicle deployment and testing.

[0027] S3, Real-world environment fine-tuning training and control: The initial policy network and model parameters obtained from the simulation layered training are deployed to the real-world vehicle controller. During real-world vehicle operation, the multi-head policy gating network takes the real-time collected vehicle scene state features as input and outputs three-domain expert index vectors in parallel, activating corresponding sub-experts from the expert pool of a specific domain. Each activated sub-expert synchronously receives the instantaneous state features required by its domain and outputs physical control quantities such as throttle adjustment coefficient, target gear command, and fan speed command, which are then executed by the vehicle actuators. After the action is executed, the system calculates the reward values ​​of the global evaluation function (used for the gating network) and the domain evaluation function (used for the corresponding sub-experts) according to the reward mechanism defined in step S1, and obtains the scene state features and instantaneous state features of each domain at the next moment. The empirical data generated by a single vehicle executing one action is recorded as follows: During actual vehicle operation, the accumulated experience data generated by a single vehicle constitutes the vehicle-side experience pool. The vehicle utilizes idle time periods or online background processes to sample experience data from the vehicle-side experience pool for incremental fine-tuning training, and updates the parameters of the gating network scheduling strategy and the control strategies of each sub-expert, so that the control strategy can continuously evolve under real vehicle operating conditions, and gradually improve the reliability and environmental adaptability of energy-saving control.

[0028] S4, Cloud-based Experience Aggregation and Continuous Learning: A pool of on-vehicle experience generated during actual operation of vehicles with similar market configurations. The historical experience database is created by synchronously uploading data to a cloud server via the vehicle communication module. This database possesses diverse characteristics reflecting the needs of different individual drivers, providing rich data sources for further optimization training and offering substantial data support for the generalization and optimization of strategies. Experience-based backtracking optimization training is performed based on this cloud-based historical experience database. The updated multi-head policy gating network and the sub-experts of each domain in the specific domain expert pool are then distributed to the vehicle via OTA (Over-The-Air) remote online upgrades, enabling continuous online optimization of the control strategy. Upgrades and optimizations are then implemented on the vehicle upon driver request.

[0029] Furthermore, step S11 constructs the torque control sub-expert, gear selection sub-expert, and fan control sub-expert, specifically including the following steps: S111, Establishing a torque control sub-expert based on a deep deterministic strategy gradient algorithm. 1) Environmental conditions: Torque control pedal, brake pedal, steering wheel information, vehicle speed, acceleration, motor status, engine status, altitude, and slope information; 2) Control action: The reference engine torque obtained by mapping the torque control pedal is proportionally adjusted, and the adjustment range is ±5% of the reference engine torque; 3) Torque control sub-expert architecture: Construct an Actor-Critic network and its corresponding target network to establish the mapping relationship between vehicle state and torque control actions; 4) Design the reward function for the torque control sub-expert: (1); in, Indicates torque control; For a moment The environmental conditions; For a moment Torque control action; The environmental state at the next moment; From state to state The driving distance; This represents the cumulative fuel consumption during the state transition. This represents the average rate of change of torque control during the state transition process. The positive weighting constant for torque control, which is used to balance the goals of encouraging vehicle movement and reducing fuel consumption, is an energy consumption incentive. A positive weighting constant is used to reward torque control smoothness and to ensure the target of control smoothness; S112, Establishing a gear selection sub-expert based on a deep Q-network algorithm 1) Environmental conditions: vehicle speed, engine speed, current gear, torque control pedal, brake pedal, steering wheel information, torque adjustment command; 2) Control actions: Discrete gear adjustment decisions, including upshifting two gears, upshifting one gear, maintaining the current gear, downshifting one gear, and downshifting two gears; 3) Gear selection sub-expert architecture: master Q network and target Q network, used to establish the mapping relationship between vehicle operating status and gear selection; 4) Design the expert reward function for gear selection: (2); in, Indicates gear selection; For a moment The environmental conditions; For a moment The gear selection action; The environmental state at the next moment; This refers to the engine output power at the next moment; A positive weighting constant is selected for the energy consumption reward level to balance the goals of encouraging vehicle movement and reducing fuel consumption; This is a positive weighting constant for motivational rewards, used to ensure the goals of motivation. and Although both weight parameters are derived from the positive weight constants of energy consumption reward, they belong to two completely independent energy consumption reward functions: gear selection and torque control. They are independent of each other and can be labeled with different values ​​according to their respective control objectives (such as the emphasis on energy consumption and power performance in gear selection, and the emphasis on energy consumption and smoothness in torque control). Therefore, different codes must be used to distinguish them.

[0030] S113: Establishing a fan control sub-expert based on a deep deterministic policy gradient algorithm 1) Environmental conditions: vehicle speed, acceleration, motor status, engine status, cooling system status information, ambient temperature, altitude, torque adjustment command, gear command; 2) Control Action: Continuous fan speed commands within the range of 0 to the maximum fan speed; 3) Fan control sub-expert architecture: Construct an Actor-Critic network and its corresponding target network to establish the mapping relationship between the overall thermal state, operating state and fan control actions; 4) Design the expert reward function for the fan control sub-system: (3); in, Indicates fan control; For a moment The environmental conditions; For a moment Fan control actions; The environmental state at the next moment; This refers to the current coolant temperature. The preset optimal operating coolant temperature for the engine; This is the upper limit of the coolant temperature, used to ensure thermal safety; This is a positive weighting constant for the optimal temperature control target reward. A positive weighting constant for thermal safety rewards, used to control the coolant temperature within the optimal range while ensuring thermal safety.

[0031] Further, step S12 specifically involves: constructing a multi-head policy gating network based on the policy gradient method and convolutional neural network, acquiring vehicle and environmental data, and inputting it into the gating network; the gating network uses a convolutional neural network to encode the time-series vehicle and environmental data, extracting high-dimensional state features, and generating three sub-expert action probability distributions for torque control, gear selection, and fan control from the expert pool; subsequently, based on the action probability distribution results, it makes sub-expert control combination decisions and executes joint control actions to collect reward feedback; the vehicle and environmental data includes driver instructions, vehicle operating status, road and environmental information; the driver instructions include torque control pedal, brake pedal, and steering wheel information; the vehicle operating status includes vehicle speed, acceleration, motor status, engine status, transmission status, and cooling system status information; the road and environmental information includes GPS signals, ambient temperature, altitude, and slope information. Specifically, it includes the following steps: S121, Input vehicle and environmental data characteristics (4); in, The input is a time series matrix; For a moment Vehicle and environmental data feature vectors; T The set time window length; F The number of feature dimensions, including the vehicle and environment data features; for A dimensional matrix space; This represents the feature vector of vehicle and environmental data at the initial moment; S122, Feature extraction based on CNN (5); in, This represents the state at time t after processing by the CNN convolutional neural network. For A convolutional neural network with parameters; The internal parameters of the convolutional neural network (including the number of neural network layers, the number of neurons, the weights of the convolutional kernel, and the bias) are obtained through pre-training. This patent does not limit its specific structure; any CNN structure that can achieve feature extraction is acceptable.

[0032] S123, Constructing a Multi-Head Policy Gated Network (6); in, For the system at time The policy probability distribution of a multi-head policy gating network; For the system at time The joint strategy decision-making action represents a set of sub-expert control combinations for decision-making; For the system at time Choose expert Chi Qizi. The policy probability distribution of the action; For Sub-experts as parameters The strategy head corresponds to the fully connected layer neural network; The internal parameters of the fully connected neural network (including the number of neural network layers, the number of neurons, weights, and biases) are obtained through pre-training, and this patent does not limit its specific structure. This refers to the expert pool, which consists of torque control, gear selection, and fan control. This refers to any one of the three sub-experts: torque control, gear selection, or fan control. Formula (6) is derived from formulas (8) and (9): The probability distributions of the strategies for the torque control, gear selection, and fan control sub-experts are as follows: (8); in, For the system at time of expert The strategy probability distribution; For the system at time Pair expert Select an action; The joint strategy probability distribution is expressed as: (9); in, For the system at time of Strategy probability distribution; S124, Design the reward function for a multi-head policy gated network. (7); in, For a moment Torque control expert reward function; For a moment Gear selection expert reward function; For a moment The expert reward function for the fan control sub-system; The sum of the rewards for the three sub-experts represents the energy efficiency reward; The reward is used to constrain the smoothness of system operation, representing the number of expert strategy switching per unit travel distance; Deadline The cumulative driving distance; Switch indicator variables for experts; For a moment The selected sub-expert control group number; , These are the weighting constants for energy efficiency rewards and smoothness rewards, respectively, used to balance the trade-offs between multiple objectives.

[0033] Formula (7) is derived from formulas (10), (11), (12) and (13): (10); (11); (12); (13); in, The energy efficiency bonus is equal to the sum of the bonuses for the three sub-experts; , , The calculations are shown in formulas (1), (2), and (3), respectively; Used to constrain the smoothness of system operation.

[0034] Furthermore, step S2 specifically includes the following steps: S21, Independent Pre-training of Intelligent Agents: Simulation models of the entire vehicle, engine, transmission, and thermal management system are built in the AMESim simulation environment. Road spectrum data for different user operating conditions is extracted based on vehicle network big data. This road spectrum data includes time-vehicle speed, mileage-altitude, and time-torque control timing feature data. Different road spectrum data are used as input to the simulation model. Based on the deep learning framework in step S12, independent pre-training is performed on the torque control, gear selection, and fan control sub-experts to form sub-experts adapted to the operating condition features, which are then combined to form an expert pool. For the torque control and fan control sub-experts, whose action spaces are continuous variables, the DDPG algorithm is used for training. For the gear selection sub-expert, whose action space is discrete variables, the DQN algorithm is used for training. S22, a two-stage strategy is adopted to train the multi-head policy gating network: First, while keeping the parameters of each sub-expert fixed, the parameters of the gating network are quickly trained based on the multi-head policy gating network in step S1 to obtain an initial policy model; then, the joint collaborative training stage is entered, in which the parameters of the gating network and all sub-experts are updated synchronously at the same time, and the initial policy network and parameters are formed through iterative training.

[0035] Furthermore, step S3 specifically involves: in a real-world vehicle operating environment, using real-world vehicle data as input, and through a multi-head strategy gating network, making real-time decisions on torque control, gear selection, and fan control sub-expert control combined actions; The torque control sub-expert generates the optimal engine torque requirement based on the environmental conditions described in step S111, establishing a basic control benchmark for the entire control system. The gear selection expert combines the optimal engine torque requirement with the environmental conditions described in step S112 to select the most economical gear, so that the engine always operates in the high-efficiency range. Finally, the fan control sub-expert integrates the control actions output by the torque control sub-expert and the gear selection sub-expert, as well as the environmental state described in step S113, to finely adjust the fan speed, minimizing accessory power consumption while ensuring thermal safety. Each sub-expert focuses on different control objectives, but they form a closed-loop collaborative expert execution network through their dependencies. The current environmental state, actions, rewards, and the next environmental state are vector-combined, and the empirical data generated by a single vehicle performing an action is used. Based on single-bike experience data The accumulated vehicle-side experience pool Experience-based fine-tuning training is conducted on individual vehicles. Furthermore, step S4 specifically involves: pooling the vehicle-side experience generated during the actual operation of the market-configured vehicles. Securely uploaded to the cloud to form a cloud-based historical experience database, enabling large-scale aggregation of operational data from the fleet. This cloud-based historical experience database possesses diverse characteristics reflecting the needs of different individual drivers, and is used to train more comprehensive and robust models in the cloud. Periodic retraining and fine-tuning are performed on torque control, gear selection, fan control sub-experts, and multi-head strategy gating networks, allowing the model to learn richer driving conditions and adapt to a wider range of usage patterns and operating environments. Subsequently, the updated model and parameters are redeployed to each vehicle terminal via OTA remote online upgrades, achieving continuous optimization and adaptive improvement of the entire fleet's control performance.

[0036] Example 2: like Figure 2 The framework diagram of the self-evolving energy-saving control system for vehicle trains with adaptive operating conditions and hybrid expert scheduling is shown in this embodiment. This embodiment provides a self-evolving energy-saving control system for vehicle trains with adaptive operating conditions and hybrid expert scheduling, including a simulation environment layered training module, a real vehicle environment fine-tuning training and control module, and a cloud-based experience aggregation and continuous learning module. The simulation environment hierarchical training module is used to implement the initial version of torque control, gear selection and fan control sub-expert pre-training in the simulation environment to form an independent expert pool, as well as to perform collaborative training on the multi-head strategy gating network. The real-vehicle environment fine-tuning training and control module is used to acquire vehicle and environmental data, including driver commands, vehicle operating status, road and environmental information, in a real-vehicle environment, complete the selection of sub-experts and control decisions, and generate single-vehicle experience data. This is used for continuous iterative upgrades of the model; The cloud-based experience aggregation and continuous learning module includes a cloud server for collecting vehicle-side experience pools generated by multiple vehicles during actual operation. This forms a diverse cloud-based historical experience database; the torque control, gear selection, and fan control sub-experts, as well as the multi-head strategy gating network, are periodically retrained and fine-tuned in the cloud, and the updated model is deployed to the vehicle when requested by the driver via OTA remote online upgrade; The multi-head strategy gating network is used to decide on the appropriate sub-expert control combination from the torque control, gear selection, and fan expert pools based on the current operating condition characteristics. It establishes the input operating condition state, action, and reward function corresponding to the gating network for iterative training of the gating network model parameters. Based on a deep reinforcement learning framework, expert pools are constructed for torque control, gear selection, and fan control respectively. Each expert pool contains multiple sub-expert models associated with vehicle operating condition characteristics. The environmental state, action, and reward function corresponding to each sub-expert are established for iterative training of each sub-expert. The input layer signal acquisition module completes the selection and control decision of each sub-expert, and the actuators adjust the torque, gear, and fan, while simultaneously fine-tuning the model based on action feedback. The input layer signal acquisition module is used to acquire vehicle and environmental data, including driver commands, vehicle operating status, road and environmental information; The actuator receives control decisions, executes control commands, provides feedback on the execution results, and generates single-vehicle experience data. ; The torque control sub-expert is used to generate the optimal engine torque demand and establish a basic control benchmark for the entire control system. The gear selection expert is used to select the most economical gear so that the engine always operates in the high-efficiency range. The fan control sub-expert is used to finely adjust the fan speed and minimize the power consumption of accessories while ensuring thermal safety.

[0037] This invention fundamentally resolves the dimensionality curse and real-time challenges faced by single models through the structure of an expert database and a gated network scheduling; through a time-series analysis-based gated network, it achieves smooth and interpretable dynamic adaptation to complex continuous operating conditions, overcoming the rigidity defects of rule-based systems; by establishing a closed-loop self-evolution mechanism, the system gains the ability to continuously learn and optimize using real data throughout its lifespan, significantly reducing long-term maintenance costs.

[0038] The above description is merely an example of embodiments of the present invention and is not intended to limit the present invention in any way. The scope of protection of the present invention is defined by the claims and is not limited to the specific embodiments described above. Any simple modifications or equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention shall fall within the scope of protection of the present invention.

Claims

1. A self-evolution energy-saving control method of an automobile train with adaptive hybrid expert scheduling, characterized in that, Includes the following steps: S1, Constructing a self-evolving energy-saving control system for vehicle-trains with adaptive operating conditions and hybrid expert scheduling. S11. Construct deep reinforcement learning sub-experts for specific domains. Construct torque control sub-experts, gear selection sub-experts, and fan control sub-experts for the torque control domain, gear selection domain, and fan control domain, respectively. Each sub-expert takes the instantaneous features of the corresponding domain as input and outputs the corresponding physical control quantity. Each sub-expert constructs sub-expert experience data based on the instantaneous features, physical control quantity, and domain evaluation function value generated by its executed action, and conducts continuous training on the vehicle and cloud respectively. S12, a multi-head policy gating network is constructed based on the policy gradient method and convolutional neural network. The gating network takes vehicle and environment data as input and selects one activated torque control sub-expert, gear selection sub-expert, and fan control sub-expert in parallel from the independent domain expert pool to form a three-domain control combination of torque control, gear selection, and fan control adapted to the current scenario. The gating network constructs gating experience data based on the vehicle and environment data generated by its selection action, the joint policy decision action, and the multi-head policy gating network reward, and performs continuous training on the vehicle and cloud respectively. S2, Layered training in simulation environment: In the AMESim simulation environment, each domain sub-expert is independently pre-trained based on different road spectrums under different working conditions to form an independent domain expert pool; the gated network is integrated with the pre-trained domain sub-experts for collaborative training, and the initial policy network and parameters are obtained through iterative training; S3, Real-vehicle environment fine-tuning training and control: Using real data collected from the actual vehicle as input, the gating network outputs a combination of three-domain expert indices in real time. Each activated sub-expert synchronously outputs and executes corresponding physical control commands. After execution, the rewards for each sub-expert are calculated based on the global reward function and the domain evaluation function, generating single-vehicle experience data. This cumulatively forms the vehicle-side experience pool. Individual vehicles undergo incremental fine-tuning training based on the vehicle-side experience pool. S4, Cloud-based experience aggregation and continuous learning: Vehicles with the same configuration upload their respective vehicle-side experience pool data to the cloud, which is then aggregated to form a cloud-based historical experience database. Based on the database, the cloud performs offline experience backtracking optimization training on the gating network and expert pools in each domain, and sends the updated strategy parameters to the vehicle via OTA when the driver authorizes the vehicle.

2. The self-evolving energy-saving control method for automobile trains according to claim 1, characterized in that... Step S11 involves constructing the torque control sub-expert, gear selection sub-expert, and fan control sub-expert, specifically including the following steps: S111, Establishing a torque control sub-expert based on a deep deterministic strategy gradient algorithm. 1) Environmental conditions: Torque control pedal, brake pedal, steering wheel information, vehicle speed, acceleration, motor status, engine status, altitude, and slope information; 2) Control action: The reference engine torque obtained by mapping the torque control pedal is proportionally adjusted, and the adjustment range is ±5% of the reference engine torque; 3) Torque control sub-expert architecture: Construct an Actor-Critic network and its corresponding target network to establish the mapping relationship between vehicle state and torque control actions; 4) Design the reward function for the torque control sub-expert: ； in, Indicates torque control; For a moment The environmental conditions; For a moment Torque control action; The environmental state at the next moment; From state to state The driving distance; This represents the cumulative fuel consumption during the state transition. This represents the average rate of change of torque control during the state transition process. The positive weighting constant for torque control, which is used to balance the goals of encouraging vehicle movement and reducing fuel consumption, is an energy consumption incentive. A positive weighting constant is used to reward torque control smoothness and to ensure the target of control smoothness; S112, Establishing a gear selection sub-expert based on a deep Q-network algorithm 1) Environmental conditions: vehicle speed, engine speed, current gear, torque control pedal, brake pedal, steering wheel information, torque adjustment command; 2) Control actions: Discrete gear adjustment decisions, including upshifting two gears, upshifting one gear, maintaining the current gear, downshifting one gear, and downshifting two gears; 3) Gear selection sub-expert architecture: master Q network and target Q network, used to establish the mapping relationship between vehicle operating status and gear selection; 4) Design the expert reward function for gear selection: ； in, Indicates gear selection; For a moment The environmental conditions; For a moment The gear selection action; The environmental state at the next moment; This refers to the engine output power at the next moment; A positive weighting constant is selected for the energy consumption reward level to balance the goals of encouraging vehicle movement and reducing fuel consumption; It is a positive weighting constant for motivational rewards, used to ensure the goals of motivation; S113, Establishing a fan control sub-expert based on a deep deterministic policy gradient algorithm. 1) Environmental conditions: vehicle speed, acceleration, motor status, engine status, cooling system status information, ambient temperature, altitude, torque adjustment command, gear command; 2) Control Action: Continuous fan speed commands within the range of 0 to the maximum fan speed; 3) Fan control sub-expert architecture: Construct an Actor-Critic network and its corresponding target network to establish the mapping relationship between the overall thermal state, operating state and fan control actions; 4) Design the expert reward function for the fan control sub-system: ； in, Indicates fan control; For a moment The environmental conditions; For a moment Fan control actions; The environmental state at the next moment; This refers to the current coolant temperature. The preset optimal operating coolant temperature for the engine; This is the upper limit of the coolant temperature, used to ensure thermal safety; This is a positive weighting constant for the optimal temperature control target reward. A positive weighting constant for thermal safety rewards, used to control the coolant temperature within the optimal range while ensuring thermal safety.

3. The self-evolving energy-saving control method for automobile trains according to claim 1, characterized in that... Step S12, which involves constructing a multi-head policy gating network based on the policy gradient method and convolutional neural network, includes: acquiring vehicle and environmental data and inputting it into the gating network; the gating network uses a convolutional neural network to encode the time-series vehicle and environmental data, extracts high-dimensional state features, and generates action probability distributions for torque control sub-experts, gear selection sub-experts, and fan control sub-experts from an independent domain expert pool; subsequently, it makes sub-expert control combination decisions based on the action probability distribution results and executes joint control actions to collect reward feedback; the vehicle and environmental data includes driver instructions, vehicle operating status, road and environmental information; the driver instructions include torque control pedal, brake pedal, and steering wheel information; the vehicle operating status includes vehicle speed, acceleration, motor status, engine status, transmission status, and cooling system status information; the road and environmental information includes GPS signals, ambient temperature, altitude, and slope information.

4. The self-evolving energy-saving control method for automobile trains according to claim 1, characterized in that, Step S12 specifically includes the following steps: S121, Input vehicle and environmental data characteristics ； in, The input is a time series matrix; For a moment Vehicle and environmental data feature vectors; T The set time window length; F The number of feature dimensions, including the vehicle and environment data features; for A dimensional matrix space; This represents the feature vector of vehicle and environmental data at the initial moment; S122, Feature extraction based on CNN ； Wherein, cnn represents a convolutional neural network; This represents the state at time t after processing by the CNN convolutional neural network. For A convolutional neural network with parameters, These are the internal parameters of the convolutional neural network, including the number of neural network layers, the number of neurons, the weights of the convolutional kernel, and the bias. Their specific values ​​are obtained through pre-training. S123, Constructing a Multi-Head Policy Gated Network ； in, For the system at time The policy probability distribution of a multi-head policy gating network; For the system at time The joint strategy decision-making action represents a set of sub-expert control combinations for decision-making; For the system at time Choose expert Chi Qizi. The policy probability distribution of the action; For Sub-experts as parameters The strategy head corresponds to the fully connected layer neural network; These are the internal parameters of a fully connected neural network, including the number of layers, the number of neurons, the weights, and the biases. Their specific values ​​are obtained through pre-training. This represents an independent domain expert pool consisting of torque control sub-experts, gear selection sub-experts, and fan control sub-experts. It refers to any one of the torque control sub-expert, gear selection sub-expert, or fan control sub-expert; S124, Design the reward function for a multi-head policy gated network. ； in, For a moment Torque control expert reward function; For a moment Gear selection expert reward function; For a moment The expert reward function for the fan control sub-system; The sum of the rewards for the three sub-experts represents the energy efficiency reward; The reward is used to constrain the smoothness of system operation, representing the number of expert strategy switching per unit travel distance; Deadline The cumulative driving distance; Switch indicator variables for experts; For a moment The selected sub-expert control group number; , These are the weighting constants for energy efficiency rewards and smoothness rewards, respectively.

5. The self-evolving energy-saving control method for automobile trains according to claim 1, characterized in that, Step S2 specifically includes the following steps: S21, Independent pre-training of intelligent agents: Build simulation models of the whole vehicle, engine, transmission and thermal management system in the AMESim simulation environment, and extract road spectrum data of different user working conditions based on the Internet of Vehicles big data. The road spectrum data includes time-vehicle speed, mileage-altitude and time-torque control timing feature data. Using different road spectrums under different working conditions as input to the simulation model, based on the deep learning framework in step S12, the torque control sub-expert, gear selection sub-expert, and fan control sub-expert are independently pre-trained to form sub-experts that are adapted to the working condition characteristics, and combined to form an independent domain expert pool. For the torque control sub-expert and the fan control sub-expert, whose action spaces are continuous variables, the DDPG algorithm is used for training; for the gear selection sub-expert, whose action space is discrete variables, the DQN algorithm is used for training. S22, a two-stage strategy is adopted to train the multi-head policy gating network: First, while keeping the parameters of each sub-expert fixed, the parameters of the gating network are quickly trained based on the multi-head policy gating network in step S1 to obtain an initial policy model; then, the joint collaborative training stage is entered, in which the parameters of the gating network and all sub-experts are updated synchronously at the same time, and the initial policy network and parameters are formed through iterative training.

6. The self-evolving energy-saving control method for automobile trains according to claim 2, characterized in that, Step S3 specifically involves: in a real-world vehicle operating environment, using real-world vehicle data as input, and through a multi-head strategy gating network, making real-time decisions on torque control sub-experts, gear selection sub-experts, and fan control sub-experts to control combined actions. The torque control sub-expert generates the optimal engine torque requirement based on the environmental conditions described in step S111, establishing a basic control benchmark for the entire control system. The gear selection expert combines the optimal engine torque requirement with the environmental conditions described in step S112 to select the most economical gear, so that the engine always operates in the high-efficiency range. Finally, the fan control sub-expert integrates the control actions output by the torque control sub-expert and the gear selection sub-expert, as well as the environmental state described in step S113, to finely adjust the fan speed, minimizing accessory power consumption while ensuring thermal safety. Each sub-expert focuses on different control objectives, but they form a closed-loop collaborative expert execution network through their dependencies. The current environmental state, actions, rewards, and the next environmental state are each vector-combined to serve as the empirical data generated by a single vehicle performing an action. Based on single-bike experience data The accumulated vehicle-side experience pool Experience-based fine-tuning training is conducted on individual vehicles.

7. The self-evolving energy-saving control method for automobile trains according to claim 1, characterized in that... Step S4 specifically involves: accumulating vehicle-side experience pools generated during actual operation of vehicles with the same market configuration. Securely uploaded to the cloud to form a cloud-based historical experience database, enabling large-scale aggregation of operational data from the fleet. Periodically retrained and fine-tuned torque control, gear selection, fan control sub-experts, and multi-head strategy gating networks are then performed in the cloud, allowing the model to learn richer driving conditions and adapt to a wider range of usage patterns and operating environments. Subsequently, the updated model and parameters are redeployed to each vehicle terminal via OTA remote online upgrades, achieving continuous optimization and adaptive improvement of the entire fleet's control performance.

8. The self-evolving energy-saving control method for vehicle trains with adaptive hybrid expert scheduling according to any one of claims 1-7, characterized in that: The self-evolving energy-saving control system for automobile trains includes a multi-head strategy gating network, a simulation environment hierarchical training module, a real vehicle environment fine-tuning training and control module, and a cloud-based experience aggregation and continuous learning module. The multi-head strategy gating network is used to decide on the appropriate sub-expert control combination from the torque control, gear selection, and fan expert pools based on the current operating condition characteristics. It establishes the input operating condition state, action, and reward function corresponding to the gating network for iterative training of the gating network model parameters. Based on a deep reinforcement learning framework, expert pools are constructed for torque control, gear selection, and fan control respectively. Each expert pool contains multiple sub-expert models associated with vehicle operating condition characteristics. The environmental state, action, and reward function corresponding to each sub-expert are established for iterative training of each sub-expert. The input layer signal acquisition module completes the selection and control decision of each sub-expert, and the actuators adjust the torque, gear, and fan, while simultaneously fine-tuning the model based on action feedback. The simulation environment hierarchical training module is used to implement the initial version of torque control, gear selection and fan control sub-expert pre-training in the simulation environment, as well as to perform collaborative training on the multi-head strategy gating network. The real-vehicle environment fine-tuning training and control module is used to complete the selection and control decisions of each sub-expert in a real-vehicle environment, and at the same time generate single-vehicle experience data. Used for continuous iterative upgrades of the model; The cloud-based experience aggregation and continuous learning module includes a cloud server for collecting vehicle-side experience pools generated by multiple vehicles during actual operation. The data is uploaded to the cloud simultaneously to form a cloud-based historical experience database. Based on this database, the torque control, gear selection, and fan control sub-experts, as well as the multi-head strategy gating network, are periodically retrained and fine-tuned in the cloud. The updated model is then deployed to the vehicle via OTA (Over-The-Air) remote online upgrade when requested by the driver.

9. The self-evolving energy-saving control method for vehicle trains with adaptive hybrid expert scheduling according to claim 8, characterized in that: The actuator receives control decisions from a multi-head strategy gating network and an expert pool, executes control commands, provides feedback on the execution results, and generates single-vehicle experience data. ; The input layer signal acquisition module is used to acquire vehicle and environmental data.

10. The self-evolving energy-saving control method for vehicle trains with adaptive hybrid expert scheduling according to claim 8, characterized in that: The torque control sub-expert is used to generate the optimal engine torque demand and establish a basic control benchmark for the entire control system. The gear selection expert is used to select the most economical gear so that the engine always operates in the high-efficiency range. The fan control sub-expert is used to finely adjust the fan speed and minimize the power consumption of accessories while ensuring thermal safety.

Images