A vehicle driving mode active adjustment method and system based on deep reinforcement learning

Abstract

The application discloses a kind of vehicle driving mode active adjustment method and system based on deep reinforcement learning, it is related to new energy vehicle intelligent control technical field.Pass through real-time acquisition driver operation data and vehicle state data, construct multivariate time series state space;Utilize the time sequence feature of driver behavior and vehicle state extracted by LSTM network;Based on PPO algorithm, construct reinforcement learning intelligent agent, output the collaborative combination action of driving mode and drive mode;Through multi-objective weighted reward function, guide intelligent agent to learn the comprehensive optimal strategy of energy saving, power, comfort and personalized preference.System uses the closed-loop architecture of "real-time acquisition-model inference-safety check-collaborative adjustment-feedback optimization", supports online fine-tuning and adaptive learning.The application realizes the active adjustment and personalized adaptation of driving mode, improves driving experience, energy utilization efficiency and system adaptability, and can be conveniently integrated into the vehicle control system of plug-in hybrid electric vehicle.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for new energy vehicles, specifically to a vehicle driving mode active adjustment strategy and system based on deep reinforcement learning. It achieves driver behavior modeling and dynamic adaptation of driving modes by integrating the PPO (Proximal Policy Optimization) algorithm with LSTM (Long Short-Term Memory Network). Background Technology

[0002] The production and sales of new energy vehicles have been increasing year by year, and they are becoming increasingly popular with consumers. Compared with traditional gasoline vehicles, new energy vehicles have significant advantages in multiple dimensions such as energy consumption costs, environmental attributes, and driving experience. Among them, plug-in hybrid electric vehicles (PHEVs) combine the advantages of pure electric drive and fuel drive, offering low electricity costs for short distances and no anxiety about fuel consumption for long distances. Charging methods are flexible, eliminating the need to rely on public charging stations. The driving experience is superior to that of traditional gasoline vehicles, balancing smoothness and power.

[0003] Plug-in hybrid electric vehicles (PHEVs) integrate a dual-power architecture consisting of a battery and electric motor drive system, and an internal combustion engine and mechanical transmission system, resulting in multiple operating modes such as pure electric drive, engine direct drive, hybrid drive, range extender, and on-the-go charging. There is still significant room for technological optimization in the existing methods of controlling vehicle driving modes.

[0004] The driving mode cannot be switched automatically: it mainly relies on the driver to manually switch (such as the vehicle's economy mode, sport mode, standard mode, etc.), which makes it difficult to achieve precise matching with dynamically changing driving scenarios and driver styles, and thus cannot fully realize the energy consumption optimization potential and power performance advantages of PHEV.

[0005] Insufficient adaptability of drive mode strategies: Although drive modes (such as pure electric drive, engine direct drive, hybrid drive, etc.) can be initially adapted based on single or a few characteristic parameters such as vehicle speed and throttle opening, they lack the ability to deeply perceive and model the driver's personalized driving habits and dynamic driving intentions, resulting in a significant lack of targetedness and environmental adaptability of the control strategy.

[0006] Lack of adaptive adjustment capability: In actual driving, a driver's behavior is significantly personalized and temporally correlated. For example, different drivers have different accelerator pedal control force and braking frequency, and the driving intentions of the same driver will also change dynamically under different road conditions. Existing technologies cannot effectively model the temporal behavioral characteristics of drivers, making it difficult to predict driving intentions in advance and actively adjust driving modes. This often leads to a mismatch between driving modes and actual driving needs, thereby affecting the driving experience and energy efficiency.

[0007] Current technologies generally suffer from limitations in fully leveraging the temporal characteristics of driver behavior to accurately predict driving intentions, thus hindering the achievement of proactive, real-time, and personalized driving mode adaptation. Therefore, developing a vehicle driving mode proactive adjustment strategy and system based on deep reinforcement learning to achieve accurate modeling of driver behavior and advanced prediction of driving intentions, thereby enabling adaptive control of driving modes, has become a critical technical challenge urgently needing to be addressed in the field of intelligent vehicle control. Summary of the Invention

[0008] The purpose of this invention is to provide an intelligent active adjustment vehicle driving mode control algorithm, addressing the urgent need to construct a vehicle driving mode active adjustment strategy and system based on deep reinforcement learning. By learning from the vehicle's real-time operating conditions (vehicle speed, acceleration, accelerator pedal opening, steering angle, etc.), the algorithm learns the driver's historical driving behavior characteristics to construct an active adjustment strategy capable of accurately predicting driving intentions and adapting to individual driving habits. This strategy is then combined with the driving mode to build a corresponding driving driving mode system, achieving active adjustment of the driving driving mode and improving driving safety, comfort, and economy. This solves the problems of passivity, poor adaptability, and insufficient personalization in existing vehicle driving mode adjustment technologies.

[0009] The technical solution adopted in this invention is a method for actively adjusting vehicle driving modes based on deep reinforcement learning, comprising the following steps:

[0010] Step S1: Real-time collection of driver operation data and vehicle status data to form a multivariate time series state space;

[0011] Step S2: Use an LSTM network to extract temporal features from the state space to generate a hidden state vector containing historical information;

[0012] Step S3: Input the hidden state vector into the reinforcement learning agent based on the PPO algorithm, and the Actor network outputs the optimal combination of actions at the current moment. The combination of actions includes driving mode adjustment actions and driving mode suggestion actions.

[0013] Step S4: After the combined action is verified by the safety verification layer, it is executed by the execution control module to realize the coordinated adjustment of driving mode and drive mode;

[0014] Step S5: Evaluate the effect of the action based on the reward function and update the agent network parameters using the PPO algorithm;

[0015] Step S6: Collect interaction data in the background in shadow mode or after driver confirmation. When the driver manually corrects the driving mode output by the model three times in a row, the online fine-tuning process is triggered to update the model parameters using the latest 1,000 time-series samples.

[0016] Furthermore, the state space includes:

[0017] Driver operation sequence: accelerator pedal opening, brake pedal opening, steering wheel angle and its rate of change; and vehicle status sequence: vehicle speed, acceleration, battery SOC, engine speed, current driving mode, and current drive mode.

[0018] Furthermore, the LSTM network contains two hidden layers, each with 64 LSTM units. It employs a gating mechanism consisting of a forget gate, an input gate, and an output gate to output a high-dimensional temporal feature vector with a dimension of 128, which is used to characterize driver style, driving intention, and driving scenario.

[0019] Furthermore, the combined action space includes:

[0020] Driving modes: ECO, Normal, Sport; Drive modes: EV, HEV; Combined actions include: EVECO, EVNormal, EVSport, HEVECO, HEVNormal, HEVSport.

[0021] Furthermore, the reward function is a multi-objective weighted sum, including:

[0022] Energy efficiency bonus: negatively correlated with instantaneous fuel consumption and electricity consumption; Power response bonus: positive bonus when matching Sport mode under rapid acceleration conditions; Comfort penalty: negative penalty for frequent switching of driving mode or drastic changes in power source; Preference matching bonus: assesses the degree of matching between actions and the driver's historical preferences through cosine similarity; Rule constraint bonus: ensures that actions comply with safety and physical rules.

[0023] A vehicle driving mode active adjustment system based on deep reinforcement learning, comprising:

[0024] The data acquisition module is used to collect driver operation and vehicle status data in real time;

[0025] The model inference module has a built-in PPO+LSTM intelligent agent model, which is used to output driving mode and drive mode adjustment commands.

[0026] The safety verification layer connects the model inference module and the execution control module, and is used to prevent dangerous or damaging vehicle operations before the action is executed.

[0027] The execution control module is used to receive verified instructions and perform coordinated switching between driving mode and drive mode;

[0028] The human-machine interface module is used to display the adjustment results and receive manual correction signals from the driver;

[0029] The online fine-tuning module is connected to the data acquisition module, the human-computer interaction module, and the model inference module, respectively, and is used to trigger online updates of model parameters based on correction signals.

[0030] Furthermore, the PPO+LSTM agent includes:

[0031] LSTM Temporal Encoder: Contains 2 hidden layers, each with 64 LSTM units, used to extract temporal features; Actor Network: Used to output action probability distribution; Critic Network: Used to output state value function V(s).

[0032] Furthermore, the execution control module includes dual-dimensional collaborative control logic for driving mode and drive mode:

[0033] The driving mode is ECO, Normal, or Sport; the drive mode is EV or HEV. The EV mode includes EVECO, EVMormal, EVSport, and range extender mode, while the HEV mode includes HEVECO, HEVNormal, and HEVSport. Each mode corresponds to a different drive sub-mode switching logic.

[0034] Furthermore, in the HEVECO mode, the engine operates within the optimal fuel consumption curve range in the hybrid drive mode, and the target SOC of the power battery is set to a low threshold.

[0035] In the HEVNormal mode, the engine operates at the maximum torque curve corresponding to the optimal fuel consumption range in the hybrid drive mode, and the target SOC of the power battery is slightly increased.

[0036] In the HEVSport mode, the core objective is to optimize power performance, allowing switching between engine direct drive and hybrid drive, thus reducing fuel economy constraints.

[0037] Furthermore, the module connection relationships of the system are as follows:

[0038] The data acquisition module is connected to the model inference module via CAN bus and Ethernet;

[0039] The output of the model inference module is connected to the input of the security verification layer;

[0040] The output of the security verification layer is connected to the input of the execution control module;

[0041] The first output of the execution control module is connected to the vehicle actuator via a CAN bus, and the second output is connected to the human-machine interface module via an Ethernet.

[0042] The output of the human-computer interaction module is connected to the first input of the online fine-tuning module;

[0043] The output of the data acquisition module is connected to the second input of the online fine-tuning module;

[0044] A bidirectional data interface is provided between the online fine-tuning module and the model inference module for reading interactive data and writing back updated model parameters.

[0045] The system was pre-trained in the CARLA simulation environment based on the CLTC-P standard operating condition, and the vehicle model was a plug-in hybrid electric vehicle with a P1+P3 motor architecture.

[0046] The coordinated adjustment of the driving mode and drive mode operates according to a closed-loop logic of "real-time data acquisition - model inference - safety verification - coordinated adjustment - feedback optimization", achieving adaptive continuous optimization in all scenarios.

[0047] Beneficial results:

[0048] 1. Enable proactive adjustment of driving modes to enhance the driving experience: By capturing the temporal characteristics of the driver's driving behavior through the LSTM network and combining it with the PPO algorithm to accurately predict the driving intention, the driving mode can be proactively adjusted before the driver manually switches, avoiding the problem of adjustment lag and improving the smoothness and comfort of driving.

[0049] 2. Adapt to personalized driving habits and enhance adaptability: By learning from the historical driving data of different drivers, the model can accurately identify the driving style of drivers and adapt to the needs of different drivers, thus solving the shortcomings of the fixed adjustment logic of existing technologies.

[0050] 3. Improve adjustment accuracy and reduce erroneous switching: By integrating driver operation and vehicle status, and through LSTM temporal feature extraction and PPO stabilization decision, the accuracy of driving mode adjustment is significantly improved, reducing erroneous switching or frequent switching caused by a single parameter trigger.

[0051] 4. Possesses online learning capabilities for continuous performance optimization: The system supports online learning functions, which can update model parameters based on real-time feedback from drivers. As usage time increases, the adjustment decisions become more and more in line with driver needs, achieving continuous performance optimization.

[0052] 5. High engineering applicability: The control logic of this invention is designed and developed based on standard driving conditions and real vehicle dynamics models. The hardware carrier uses general-purpose processors and storage media, which can be easily integrated into the existing vehicle control system architecture of plug-in hybrid electric vehicles without large-scale reconstruction of the original hardware, effectively reducing the industrial application cost and implementation threshold of the technical solution. Attached Figure Description

[0053] Figure 1 This is a model diagram of a plug-in hybrid electric vehicle.

[0054] Figure 2 This is a diagram of the CARLA simulation environment.

[0055] Figure 3 This is the block diagram of the EV electric mode state machine design.

[0056] Figure 4 This is a schematic diagram of the EV electric mode switching process.

[0057] Figure 5 This is the state machine design block diagram for the HEV ECO economic mode.

[0058] Figure 6 This is a schematic diagram of the HEV ECO economic mode switching process.

[0059] Figure 7 This is the block diagram of the HEV Normal standard mode state machine design.

[0060] Figure 8 This is a schematic diagram of the HEV Normal standard mode switching process.

[0061] Figure 9 This is the state machine design block diagram for HEV Sport mode.

[0062] Figure 10 This is a schematic diagram of the HEV Sport mode switching process.

[0063] Figure 11 This is a framework diagram of the active adjustment strategy for driving drive mode based on PPO+LSTM.

[0064] Figure 12 This is a flowchart of a driving mode adjustment method based on deep reinforcement learning.

[0065] Figure 13 This is a diagram of the LSTM algorithm structure. Detailed Implementation

[0066] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0067] This invention proposes a deep reinforcement learning-based active driving mode adjustment strategy based on PPO+LSTM. The system implements this strategy by combining different driving modes, and specifically includes the following steps:

[0068] Step 1:

[0069] Because driver behavior (such as continuous accelerator and brake pedal control, and cornering maneuvers) and vehicle operating status are highly time-dependent, traditional fully connected networks cannot effectively capture long-term temporal correlation features. This invention employs an LSTM network as the core for feature extraction. Each LSTM layer contains two hidden layers, with 64 LSTM units per layer. A "forget gate-input gate-output gate" gating mechanism is used: the forget gate filters out invalid temporal information (such as accidental misoperations), the input gate retains key driving behavior features (such as sustained high throttle operation during aggressive driving), and the output gate outputs a high-dimensional temporal feature vector with a dimension of 128. This feature vector can accurately represent the driver's driving style (aggressive / normal / conservative), real-time driving intentions (overtaking / cruising / deceleration and avoidance), and the core features of the current driving scenario, providing data support for subsequent PPO (Plan-Off-Purpose) strategy decisions.

[0070] Step 2:

[0071] The state space (State, s) is defined as a multivariate time series, including: (1) Driver operation time series: historical window data of accelerator pedal opening, brake pedal opening, steering wheel angle and its rate of change. (2) Vehicle state time series: vehicle speed, acceleration, battery charge (SOC), engine speed, current driving mode, and current drive mode.

[0072] Action space (Action, a): Defined as discrete combination actions, including driving mode adjustment actions (switching to one of conservative ECO, normal, or aggressive Sport) and drive mode suggestion actions (suggesting switching to one of EV or HEV). The final action is a combination of the two (EV ECO, EV Normal, EV Sport, HEV ECO, HEV Normal, HEV Sport).

[0073] The reward function (Reward, r) is designed as a multi-objective weighted sum and is key to guiding the agent's learning. It includes: an energy-saving reward, negatively correlated with instantaneous fuel / electricity consumption, encouraging economical driving; a power response reward, given when the driver accelerates rapidly, the current mode is Sport, and the power response is rapid; a comfort penalty, given for frequent switching of driving modes or drastic changes in power source; a preference-matching reward, which, through implicit learning, matches long-term decision-making styles with the driver's historical habits; and a rule-constrained reward, ensuring actions comply with safety and physical rules, such as prohibiting prolonged EV mode at low SOC.

[0074] Step 3:

[0075] A PPO+LSTM agent network is constructed. Multidimensional temporal data from the state space s are input into a fully connected network for feature extraction. The extracted feature sequences are then input into an LSTM layer. The LSTM layer captures the temporal dependencies between driver actions and vehicle states, generating a latent state vector h_t containing historical information.

[0076] The Actor network takes the LSTM output h_t as input and outputs the probability distribution of all possible action combinations in the current state through the policy network. The PPO algorithm is used to constrain the policy update magnitude, ensuring training stability.

[0077] Critic network: Takes the same h_t as input and outputs a state value function V(s), which evaluates the long-term expected reward of the current state and is used to guide the update of the Actor network.

[0078] Mathematical description of LSTM timing encoder:

[0079] Input gate: i_t =σ(W_i·[h_{t-1}, x_t] + b_i)

[0080] Forget gate: f_t =σ(W_f·[h_{t-1}, x_t] + b_f)

[0081] Candidate memory: g_t = tanh(W_g·[h_{t-1}, x_t] + b_g)

[0082] Memory update: c_t = f_t⊙c_{t-1} + i_t⊙g_t

[0083] Output gate: o_t =σ(W_o·[h_{t-1}, x_t] + b_o)

[0084] Implicit state: h_t = o_t⊙tanh(c_t)

[0085] Where: x t : The state feature vector at the current time step; h t−1 : The hidden state of the previous time step; c t−1 : Memory cell state at the previous time step; σ: sigmoid activation function; ⊙: Hadamard product (element-wise multiplication); W ∗ , b ∗ : Trainable parameter matrix and bias terms.

[0086] The agent's objective function takes the PPO-Clip form:

[0087]

[0088] Where: r t (θ) represents the probability ratio between the old and new strategies; t The advantage function estimate is calculated using GAE (Generalized Advantage Estimation):

[0089]

[0090] ϵ is the trimming parameter (usually set to 0.1-0.3); γ is the discount factor; and λ is the GAE parameter.

[0091] Reward function R t The complete calculation formula:

[0092] R t =w1⋅R efficiency +w2⋅R performance +w3⋅R comfort +w4⋅R preference

[0093] The calculations for each component are as follows:

[0094] Energy Efficiency Bonus: R efficiency =−α⋅fuel_rate(t)−β⋅electric_power(t)

[0095] Performance Bonus:

[0096]

[0097] Comfort penalty: R comfort =−λ1⋅1 mode_change (t)−λ2⋅∣ΔTorque(t)∣

[0098] Preference matching reward: R preference =cos_similarity(v action ,v driver_profile ), where v driver_profile It is a driver preference vector learned from historical data.

[0099] Step 4:

[0100] The China Light-duty Vehicle Test Cycle for Passenger Cars (CLTC-P) was used as the standard operating condition for training the agent. Interaction data generated in the CARLA simulation environment was used to pre-train the PPO+LSTM agent, enabling it to learn basic energy-saving and power-performance trade-off strategies. The vehicle model was a plug-in hybrid electric vehicle with a P1+P3 motor architecture, such as... Figure 1As shown. The CARLA simulation environment is as follows. Figure 2 As shown.

[0101] During the online deployment and adaptive phase, the pre-trained model is loaded into the onboard computing unit. The real-time perceived state is encoded by LSTM, and the Actor network outputs the optimal action to adjust the execution mode. All decisions pass through a safety verification layer to prevent actions that are dangerous or damaging to the vehicle.

[0102] In the background, the system runs in "shadow mode" or, after driver confirmation, collects real interaction data (s, a, r, s'). It periodically or under safe conditions, fine-tunes the agent online. When the system detects that the driver has manually corrected the driving mode output by the model three times in a row, it triggers the fine-tuning process and updates the model parameters using the latest 1,000 time-series samples, so that the strategy gradually adapts to the driver's personalized style.

[0103] Step 5

[0104] The driver style (conservative / normal / aggressive) output by the intelligent agent is analyzed by the mode decision execution module, which directly controls the vehicle to switch to the corresponding ECO / Normal / Sport mode, and then enters the drive mode (EV electric mode / HEV hybrid mode) for judgment.

[0105] The EV electric mode includes pure electric EV ECO mode, EV Normal mode, EV Sport mode, and range extender mode. In electric drive mode, the P3 motor is the sole drive source. When the battery charge falls below the SOC target value, the drive mode switches from pure electric drive mode to range extender mode. In pure electric drive mode, different EV drive modes are selected based on the driver's driving style. The control strategy state machine design is as follows: Figure 3 As shown, the switching process conditions are as follows: Figure 4 As shown.

[0106] In a depleted battery state, the range-extended drive mode exhibits better fuel economy than the driving-charging drive mode for low-speed urban driving conditions. However, real-world low-speed urban driving is characterized by frequent dynamic fluctuations in vehicle speed, which can lead to repeated switching between range-extended and driving-charging modes. This can cause the engine speed to enter the driving-charging mode before it has completed steady-state adjustment, resulting in increased fuel consumption. Therefore, this invention limits the use of the range-extended drive mode only in pure electric drive mode.

[0107] HEV ECO economic mode includes pure electric drive mode, charging drive mode, engine direct drive mode, and hybrid drive mode, minimizing fuel consumption while meeting the vehicle's basic power requirements. In hybrid drive mode, the engine output power is actively reduced to keep the engine operating at the optimal fuel consumption curve T.en_op The range; at the same time, pure electric drive is given priority, and the target state of charge (SOC) of the power battery is set. target A lower threshold is set to improve energy efficiency and reduce fuel consumption. The control strategy state machine design is as follows: Figure 5 As shown, the switching process conditions are as follows: Figure 6 As shown.

[0108] HEV Normal mode includes pure electric drive mode, engine direct drive mode, charging drive mode, and hybrid drive mode. In Normal mode, engine output power is unrestricted. In hybrid drive mode, the engine operates within the optimal fuel consumption range, corresponding to the maximum torque curve T. en_op_max At the same time, the target state of charge (SOC) of the power battery will be... target Slightly increase the power output to suit the performance requirements of urban driving conditions. The control strategy state machine design is as follows: Figure 7 As shown, the switching process conditions are as follows: Figure 8 As shown.

[0109] The HEV Sport mode includes an engine direct drive mode and a hybrid drive mode. This mode prioritizes optimal power performance while moderately reducing fuel economy constraints. It is activated when the battery charge level is greater than 20%, the vehicle speed is less than or equal to 140 km / h, and the vehicle's required torque T... en_req Greater than the engine's maximum torque T within its optimal operating range en_op_max Under these three conditions, the drive system will switch from engine direct drive mode to hybrid drive mode. The control strategy state machine design is as follows: Figure 9 As shown, the switching process conditions are as follows: Figure 10 As shown.

[0110] Step 6:

[0111] The strategy execution logic employs a two-dimensional adjustment mechanism, coordinating the control of driving mode and drive mode. After system startup, it operates according to a closed-loop logic of "real-time data acquisition - model inference - collaborative adjustment - feedback optimization," with the specific execution steps as follows:

[0112] Step 1: The data acquisition module collects driver operations and vehicle status in real time and transmits them to the model inference module via CAN bus and Ethernet.

[0113] Step 2: The model inference module preprocesses the data, generates time-series samples to input into the PPO+LSTM model, and outputs driving mode adjustment instructions (ECO / Normal / Sport) and adaptive driving mode suggestions.

[0114] Step 3: The control module receives instructions, enters the drive mode (EV electric mode / HEV hybrid mode) judgment, and actively completes the adjustment of the driving drive mode;

[0115] Step 4: The human-computer interaction module displays the adjustment results. If the driver does not make manual corrections, the model continues to infer based on real-time data. If the driver manually corrects the mode, the correction signal is fed back to the model inference module, triggering the online fine-tuning process.

[0116] Step 5: Repeat Steps 1-4 to achieve continuous adaptive and coordinated adjustment of driving mode and drive mode, ensuring driving experience and economy in all scenarios.

[0117] A deep reinforcement learning-based driving mode proactive adjustment strategy framework based on PPO+LSTM is as follows: Figure 11 As shown.

Claims

1. A method for actively adjusting vehicle driving modes based on deep reinforcement learning, characterized in that, Includes the following steps: Step S1: Real-time collection of driver operation data and vehicle status data to form a multivariate time series state space; Step S2: Use an LSTM network to extract temporal features from the state space to generate a hidden state vector containing historical information; Step S3: Input the hidden state vector into the reinforcement learning agent based on the PPO algorithm, and the Actor network outputs the optimal combination of actions at the current moment. The combination of actions includes driving mode adjustment actions and driving mode suggestion actions. Step S4: After the combined action is verified by the safety verification layer, it is executed by the execution control module to realize the coordinated adjustment of driving mode and drive mode; Step S5: Evaluate the effect of the action based on the reward function and update the agent network parameters using the PPO algorithm; Step S6: Collect interaction data in the background in shadow mode or after driver confirmation. When the driver manually corrects the driving mode output by the model three times in a row, the online fine-tuning process is triggered to update the model parameters using the latest 1,000 time-series samples.

2. The method according to claim 1, characterized in that, The state space includes: Driver operation sequence: accelerator pedal opening, brake pedal opening, steering wheel angle and its rate of change; and vehicle status sequence: vehicle speed, acceleration, battery SOC, engine speed, current driving mode, and current drive mode.

3. The method according to claim 1, characterized in that, The LSTM network contains two hidden layers, each with 64 LSTM units. It employs a gating mechanism consisting of a forget gate, an input gate, and an output gate, and outputs a high-dimensional temporal feature vector with a dimension of 128, which is used to characterize driver style, driving intention, and driving scenario.

4. The method according to claim 1, characterized in that, The combined action space includes: Driving modes: ECO, Normal, Sport; Drive modes: EV, HEV; Combined actions include: EVECO, EVNormal, EVSport, HEVECO, HEVNormal, HEVSport.

5. The method according to claim 1, characterized in that, The reward function is a multi-objective weighted sum, including: Energy efficiency bonus: negatively correlated with instantaneous fuel consumption and electricity consumption; Power response bonus: positive bonus when matching Sport mode under rapid acceleration conditions; Comfort penalty: negative penalty for frequent switching of driving mode or drastic changes in power source; Preference matching bonus: assesses the degree of matching between actions and the driver's historical preferences through cosine similarity; Rule constraint bonus: ensures that actions comply with safety and physical rules.

6. A vehicle driving mode active adjustment system based on deep reinforcement learning, characterized in that, include: The data acquisition module is used to collect driver operation and vehicle status data in real time; The model inference module has a built-in PPO+LSTM intelligent agent model, which is used to output driving mode and drive mode adjustment commands. The safety verification layer connects the model inference module and the execution control module, and is used to prevent dangerous or damaging vehicle operations before the action is executed. The execution control module is used to receive verified instructions and perform coordinated switching between driving mode and drive mode; The human-machine interface module is used to display the adjustment results and receive manual correction signals from the driver; The online fine-tuning module is connected to the data acquisition module, the human-computer interaction module, and the model inference module, respectively, and is used to trigger online updates of model parameters based on correction signals.

7. The system according to claim 6, characterized in that, The PPO+LSTM agent includes: LSTM Temporal Encoder: Contains 2 hidden layers, each with 64 LSTM units, used to extract temporal features; Actor Network: Used to output action probability distribution; Critic Network: Used to output state value function V(s).

8. The system according to claim 6, characterized in that, The execution control module includes dual-dimensional collaborative control logic for driving mode and drive mode: The driving mode is ECO, Normal, or Sport; the drive mode is EV or HEV. The EV mode includes EVECO, EVMormal, EVSport, and range extender mode, while the HEV mode includes HEVECO, HEVNormal, and HEVSport. Each mode corresponds to a different drive sub-mode switching logic.

9. The system according to claim 8, characterized in that: In the HEVECO mode, the engine operates in the optimal fuel consumption curve range under hybrid drive mode, and the target SOC of the power battery is set to a low threshold. In the HEVNormal mode, the engine operates at the maximum torque curve corresponding to the optimal fuel consumption range in the hybrid drive mode, and the target SOC of the power battery is slightly increased. In the HEVSport mode, the core objective is to optimize power performance, allowing switching between engine direct drive and hybrid drive, thus reducing fuel economy constraints.

10. The system according to claim 6, characterized in that, The module connection relationships of the system are as follows: The data acquisition module is connected to the model inference module via CAN bus and Ethernet; The output of the model inference module is connected to the input of the security verification layer; The output of the security verification layer is connected to the input of the execution control module; The first output of the execution control module is connected to the vehicle actuator via a CAN bus, and the second output is connected to the human-machine interface module via an Ethernet. The output of the human-computer interaction module is connected to the first input of the online fine-tuning module; The output of the data acquisition module is connected to the second input of the online fine-tuning module; A bidirectional data interface is provided between the online fine-tuning module and the model inference module for reading interactive data and writing back the updated model parameters; The system was pre-trained in the CARLA simulation environment based on the CLTC-P standard operating condition, and the vehicle model was a plug-in hybrid electric vehicle with a P1+P3 motor architecture. The coordinated adjustment of the driving mode and drive mode operates according to a closed-loop logic of "real-time data acquisition - model inference - safety verification - coordinated adjustment - feedback optimization", achieving adaptive continuous optimization in all scenarios.

Images