Method for optimizing vehicle-mounted energy efficiency based on driver behaviors
By optimizing vehicle energy efficiency through multimodal data fusion and reinforcement learning, the problems of driver behavior not being included in decision-making and insufficient personalized preferences in existing technologies are solved, realizing intelligent and personalized management of vehicle energy efficiency, extending driving range and improving driving experience.
Patent Information
- Application Number
- CN202511782205.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-29
- Publication Date
- 2026-01-16
AI Technical Summary
Existing vehicle energy efficiency optimization technologies fail to incorporate the driver's real-time behavior into decision-making and lack personalized comfort preference learning, leading to excessive energy consumption and rigid strategies.
By collecting multimodal data, using deep learning models to extract driver behavior features, and combining reinforcement learning to optimize the power output of in-vehicle equipment, the system dynamically adjusts to adapt to driver preferences, and builds personalized profiles for energy efficiency optimization.
It enables intelligent and personalized management of vehicle energy, extends driving range, and improves driving comfort and energy efficiency.
Smart Images
Figure CN121341189A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of automobile auxiliary driving, in particular to a method for optimizing vehicle energy efficiency based on driver behavior. BACKGROUND
[0002] With the popularization of electric vehicles, vehicle energy management has become a key to improving vehicle endurance. Existing vehicle energy efficiency optimization technologies mostly focus on improving hardware efficiency or distributing energy based on fixed rules, and the strategies are relatively rigid. These methods generally have two main defects: first, the real-time behavior state of the driver is not considered in the energy efficiency decision-making, and aggressive or inappropriate driving behavior itself can lead to additional energy consumption; second, there is a lack of learning ability for individual comfort preference of the driver, and fixed temperature or device settings may not meet the needs of different users, and even excessive energy consumption is maintained for comfort.
[0003] In addition, although some studies have attempted to introduce driver monitoring, they are mostly limited to single modal data, with single analysis dimension and insufficient linkage with vehicle control systems, making it difficult to form a closed-loop optimization. Therefore, there is an urgent need in the art for a technical solution that can deeply integrate driver behavior perception and adaptively learn individual preferences, thereby achieving fine and personalized dynamic optimization of vehicle energy efficiency. SUMMARY
[0004] To solve the above problems of the prior art, the present application provides a method for optimizing vehicle energy efficiency based on driver behavior. This method solves the defects of rigid strategy and ignoring individual differences of the driver in the existing vehicle energy efficiency management system. By integrating real-time behavior state of the driver and multi-dimensional data of the vehicle, and based on reinforcement learning and continuous learning of individual preferences, dynamic and personalized optimization control of vehicle energy is achieved, thereby significantly improving energy efficiency while ensuring driving safety and comfort.
[0005] The present application provides the following solutions:
[0006] A method for optimizing vehicle energy efficiency based on driver behavior, comprising:
[0007] collecting multi-modal data streams;
[0008] extracting visual spatiotemporal features and vehicle sensor time series features;
[0009] fusing visual spatiotemporal features and vehicle sensor time series features to obtain a quantitative driving behavior score;
[0010] constructing a state vector based on the driving behavior score, vehicle state data, and outside temperature, and inputting the state vector into a PPO decision engine to determine an action vector;
[0011] The control instruction is executed based on the action vector, and feedback data is collected. The reward value is calculated based on the feedback data, and the driver portrait is updated.
[0012] Further, the driver's facial video stream is taken as a basic analysis unit of 5 consecutive frames of images. A pre-trained three-dimensional convolutional neural network is used to extract the spatial and temporal features between consecutive frames. The structure of the three-dimensional convolutional neural network includes 4 convolutional layers and 2 fully connected layers, and finally outputs a fixed dimension of 512 feature vectors, which are taken as visual spatial and temporal features, denoted as Fv.
[0013] For acceleration and angular velocity data, the acceleration and angular velocity sequence in a 1-second time window is input into a long short-term memory network (LSTM). The LSTM hidden layer dimension is 128, and the encoding is a fixed dimension of 256 feature vectors, which are taken as vehicle sensor time sequence features, denoted as Fs.
[0014] Further, the visual spatial and temporal features Fv and the vehicle sensor time sequence features Fs are input into a cross-modal Transformer fusion module. The module uses 8 attention heads to perform feature interaction and alignment through multi-head attention mechanism. The mathematical representation of the fusion process is:
[0015] Attention(Q,K,V)=Softmax(QKT / dk0.5)V;
[0016] Where Q is the query vector, which is obtained by linear projection of the visual spatial and temporal features Fv. K is the key vector, which is obtained by linear projection of the vehicle sensor time sequence features Fs. V is the value vector, which is also obtained by linear projection of the vehicle sensor time sequence features Fs. QKT is the dot product between the transpose of the query vector Q and the key vector K. Softmax(...) is a probability distribution that converts all attention scores (similarities with each key vector) corresponding to each query vector into a probability distribution. dk is the projection dimension, which is set to 32.
[0017] Finally, the fusion module outputs a 256-dimensional fusion feature vector Ff.
[0018] Further, the fusion feature vector Ff is input into a scorer, which is composed of two fully connected layers. The first fully connected layer has an output dimension of 128, and the second fully connected layer has an output dimension of 1. The Sigmoid activation function is used after the second layer to map the output value to between 0 and 1, which is taken as the driving behavior score Sd at the current time.
[0019] Further, the driving behavior score Sd,t at the current time t, the battery state of charge SOC,t, the outdoor temperature Tout,t, the air conditioning set temperature Tac,t, the audio output power Pa,t, the seat power gear Pseat,t, and the comfort score Ct-1 calculated in the last decision period are collectively constructed into a state vector St.
[0020] The state vector St is input into the PPO decision engine, and the policy network of the engine outputs a three-dimensional action vector At, the definition of which is as follows:
[0021] At=[ΔTac, Raudio, Lseat];
[0022] Wherein, ΔTac is the air conditioning temperature adjustment amount, the optional value of which comes from the discrete action set {-2.0, -1.5, -1.0, -0.5, 0}, and the unit is Celsius;
[0023] Raudio is the audio volume adjustment coefficient, the optional value of which comes from the set {0.7, 0.8, 0.9, 1.0};
[0024] Lseat is the seat heating power gear, the optional value of which comes from the set {0, 1, 2}, corresponding to off, low gear, and high gear, respectively.
[0025] Further, the control instructions are executed based on the action vector, and feedback data is collected;
[0026] The reward is calculated based on the feedback data, and the PPO decision engine is driven to optimize;
[0027] The driver portrait is updated.
[0028] Further, the action instruction At is issued to the corresponding body domain controller through the CAN FD bus, and after the instruction is executed, the battery state of charge change ΔSOC, the driver's galvanic skin response and heart rate variability signal are synchronized in the next decision period, and the driver's galvanic skin response and heart rate variability signal are input into the pre-trained comfort assessment model to calculate the instantaneous comfort score Ct.
[0029] Further, the comprehensive reward Rt is calculated, Rt=0.6×(△SOC×1000)-0.4×|Ct-4.2|;
[0030] The complete data of this decision {state vector St, action vector At, comprehensive reward Rt, next state vector St+1} is stored as an experience in the experience replay buffer; wherein the next state vector St+1 is a new state vector obtained by re-executing steps S101 to S104 after executing the action At and delaying for a decision cycle, and the comfort score contained therein is Ct calculated in the current cycle;
[0031] Periodically sample a batch of data from the buffer, perform a PPO algorithm iteration, and update the parameters of the policy network and value network in the PPO decision engine.
[0032] Further, the state vector St, action vector At and comfort score Ct of this time are stored after being bound to the ID of the driver;
[0033] Initialize the driver profile: when the system first serves a driver, the temperature preference Taco is initialized to the standard comfortable temperature 22°C, and the operation sensitivity coefficient Ss is initialized to 1.0;
[0034] Run the profile update algorithm every 24 hours, which recalculates the personalized parameters of the driver based on the data stored in the past 72 hours using a weighted linear regression model, said driver personalized parameters including temperature preference Taco and operation sensitivity coefficient Ss.
[0035] Further, the updated personalized driver profile parameters, i.e. temperature preference Taco and operation sensitivity coefficient Ss, will jointly constitute the enhanced state vector Ste for decision-making with real-time state data;
[0036] Combine the operation sensitivity coefficient in the enhanced state vector with the pre-defined action amplitude weight vector, generate a personalized modulation mask through power operation, multiply the mask with the original action probability distribution output by the policy network element by element, and normalize the processing to achieve personalized modulation of action decision.
[0037] Through the above scheme, the following beneficial technical effects are obtained:
[0038] The application realizes the intelligentization and personalization of vehicle-mounted energy management. The traditional method often ignores the key variable of the driver state, and adopts a static or single energy efficiency strategy. The application fuses multi-modal data including vision, inertial measurement unit, vehicle bus and physiological signals, accurately quantifies driving behavior by using a deep learning model, and dynamically adjusts the power output of vehicle-mounted loads such as air conditioners and stereos by using a reinforcement learning engine. The system can sensitively perceive the state of the driver's distraction and fatigue, and actively implement energy efficiency optimization strategies under the premise of ensuring driving safety and comfort, effectively extending the cruising range of electric vehicles. More importantly, the system builds and dynamically updates the driver portrait by continuously learning the personal preferences and operation sensitivity of the driver, significantly enhancing the personalization and comfort of the driving experience while improving energy utilization efficiency, and achieving dual optimization of energy efficiency and experience. BRIEF DESCRIPTION OF DRAWINGS
[0039] Figure 1 is a flowchart of a method for optimizing vehicle-mounted energy efficiency based on driver behavior provided by one or more embodiments of the application.
[0040] Figure 2 is a flowchart of an analysis method for quantifying driving behavior scores of the application.
[0041] Figure 3 is a flowchart of a strategy optimization method of the application. DETAILED DESCRIPTION
[0042] The technical solutions of the application will be described clearly and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.
[0043] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the application described herein can be implemented. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not necessarily limit to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.
[0044] Please refer to Figure 1 shown, which is a flowchart of a method for optimizing vehicle-mounted energy efficiency based on driver behavior according to the present embodiment, the method comprising:
[0045] Step S101, collect multi-modal data stream.
[0046] Specifically, a 1280x720 resolution infrared camera deployed in the cockpit is used to continuously collect the driver's facial video stream at a sampling rate of 60 frames / second; a nine-axis inertial measurement unit integrated in the vehicle is used to synchronously collect three-axis acceleration and three-axis angular velocity at a sampling rate of 100Hz; a controller area network bus is used to obtain vehicle state data at a fixed period of 100ms, the vehicle state data including battery state of charge, air conditioning set temperature, audio output power, and seat heating power gear; an outdoor environment temperature and real-time geographic location information of the vehicle are obtained through a temperature and humidity sensor and a global positioning system module, and at the same time, a contact sensor integrated in the steering wheel or seat is used to collect the driver's skin electricity response and heart rate variability signal at a sampling rate of 10Hz.
[0047] Specifically, by comprehensively collecting the driver's facial video, vehicle motion data, internal system state, and external environment information, and combining physiological signal monitoring, the system can construct a multi-dimensional real-time data basis, and such multi-modal data collection ensures that the subsequent analysis can cover various aspects of the driver state, vehicle operation, and external environment, providing rich and reliable input for energy efficiency optimization, thereby improving the accuracy and response capability of the overall system.
[0048] Please continue to refer to Figure 1 As shown, the method for optimizing vehicle energy efficiency based on driver behavior further includes:
[0049] Step S102, extract visual spatio-temporal features and vehicle sensor time series features.
[0050] Specifically, for the driver's facial video stream, 5 consecutive frames of images are used as a basic analysis unit, and a pre-trained three-dimensional convolutional neural network is used to extract spatio-temporal features between consecutive frames, the structure of the three-dimensional convolutional neural network including 4 convolutional layers and 2 fully connected layers, and finally outputting a fixed dimension of 512 feature vectors, which are used as visual spatio-temporal features, denoted as Fv.
[0051] For acceleration and angular velocity data, the acceleration and angular velocity sequence within a 1-second time window is input into a long short-term memory network LSTM, the LSTM hidden layer dimension is 128, and the encoding is a fixed dimension of 256 feature vectors, which are used as vehicle sensor time series features, denoted as Fs.
[0052] Specifically, the three-dimensional convolutional neural network is pre-trained using a dataset containing 10,000 driving video segments, each labeled with attention distraction, fatigue, and other driving behavior labels, and trained to convergence using a cross-entropy loss function and an Adam optimizer; the LSTM is trained using simultaneously collected IMU data and corresponding driving event labels such as sudden acceleration and sharp turning, also using a cross-entropy loss function.
[0053] Specifically, key features are extracted from video and sensor data using a deep learning model, and the system can capture the dynamic changes of driver behavior and the subtle patterns of vehicle motion. This step helps to accurately identify the driver's state of attention distraction, fatigue, and other driving events such as sudden acceleration and sharp turning, providing high-quality feature representation for subsequent fusion and scoring, enhancing the depth and robustness of behavior analysis.
[0054] Please continue to refer to Figure 1 As shown, the method for optimizing vehicle energy efficiency based on driver behavior further comprises:
[0055] Step S103, fusion of spatio-temporal features and vehicle sensor time series features to obtain a quantitative driving behavior score.
[0056] Specifically, by integrating multi-modal features and generating a quantitative score, the system can convert complex driver states into simple numerical indicators, which not only simplifies the subsequent decision-making process, but also enables the energy efficiency optimization strategy to directly respond to driving behavior changes, achieving more precise and adaptive control.
[0057] Please refer to Figure 2 As shown, the analysis method of the quantitative driving behavior score comprises:
[0058] Step S201, fusion of spatio-temporal features and vehicle sensor time series features to obtain a fusion feature vector.
[0059] Specifically, the spatio-temporal features Fv and the vehicle sensor time series features Fs are input into a cross-modal Transformer fusion module, which adopts 8 attention heads and performs feature interaction and alignment through a multi-head attention mechanism. The mathematical representation of the fusion process is:
[0060] Attention(Q,K,V)=Softmax(QKT / dk0.5)V;
[0061] where Q is the query vector, which is a matrix obtained by linear projection of the visual spatio-temporal feature Fv, K is the key vector, which is a matrix obtained by linear projection of the vehicle sensor time-series feature Fs, V is the value vector, which is also a matrix obtained by linear projection of the vehicle sensor time-series feature Fs, QKT is the dot product between the query vector Q and the transpose of the key vector K, Softmax(...) is a function that converts all the attention scores (similarities with each key vector) corresponding to each query vector into a probability distribution, dk is the projection dimension, which is set to 32;
[0062] Finally, the fusion module outputs a 256-dimensional fusion feature vector Ff.
[0063] Please continue to refer to Figure 1 As shown in the figure, the analysis method for quantifying the driving behavior score comprises:
[0064] Step S202, determining a quantified driving behavior score based on the fusion feature vector.
[0065] Specifically, the fusion feature vector Ff is input into a scorer, which is composed of two fully connected layers, wherein the output dimension of the first fully connected layer is 128, the output dimension of the second fully connected layer is 1, and a Sigmoid activation function is used after the second layer to map the output value to between 0 and 1, as the driving behavior score Sd at the current time.
[0066] Specifically, the training of the scorer adopts an end-to-end manner and is jointly optimized with the PPO algorithm, and is not pre-trained separately; the output value is calculated with the mean square error of the expert score (annotated by experienced driving instructors according to video and sensor data) at the initial stage of training as an auxiliary supervision signal.
[0067] Please continue to refer to Figure 1 As shown in the figure, the method for optimizing the vehicle energy efficiency based on the driver behavior further comprises:
[0068] Step S104, constructing a state vector based on the driving behavior score, vehicle state data, and outside temperature, and inputting the state vector into a PPO decision engine to determine an action vector.
[0069] Specifically, the driving behavior score Sd,t at the current time t, the battery state of charge SOC,t, the outside temperature Tout,t, the air conditioner set temperature Tac,t, the audio output power Pa,t, the seat power gear Pseat,t, and the comfort score Ct-1 calculated in the last decision period are jointly constructed into a state vector St.
[0070] The state vector St is input into the PPO decision engine, and the policy network of the engine outputs a three-dimensional action vector At, wherein the definition of the action vector At and the discrete action set are as follows:
[0071] At=[ΔTac, Raudio, Lseat];
[0072] wherein ΔTac is the air conditioning temperature adjustment, whose optional values come from the discrete action set {-2.0, -1.5, -1.0, -0.5, 0}, with the unit of Celsius degree;
[0073] Raudio is the audio volume adjustment coefficient, whose optional values come from the set {0.7, 0.8, 0.9, 1.0};
[0074] Lseat is the seat heating power gear, whose optional values come from the set {0, 1, 2}, corresponding to off, low gear, and high gear, respectively.
[0075] Specifically, at system initialization, the default comfort score Ct=4.2 is set, and this value is taken as Ct-1 in the first decision-making; in subsequent decision-making, the comfort score Ct-1 calculated in the last decision-making cycle is combined with other state data to construct a state vector St.
[0076] For example, in this embodiment, the low gear can be 20w, and the high gear can be 40w.
[0077] Specifically, the policy network and the value network of the PPO decision engine are both two-layer fully connected networks, with a hidden layer dimension of 128 and using a ReLU activation function; specifically, the policy network and the value network of the PPO decision engine are both double-layer fully connected networks with a variable input layer dimension, with a hidden layer dimension of 128 and using a ReLU activation function; when the input is the base state St, the input layer is 7-dimensional; when the input is the enhanced state Ste, the input layer is 9-dimensional; the training hyperparameters are set as follows: learning rate 0.001, discount factor γ=0.99, generalized advantage estimation (GAE) parameter λ=0.95, experience replay buffer capacity 10,000 data, and 256 experience samples are randomly sampled from the buffer for each 10 decisions to perform a round of policy update.
[0078] Specifically, by constructing a state vector containing driving behavior, vehicle state, and environmental factors, the system can comprehensively capture the current driving situation, output optimized actions using a reinforcement learning decision engine, and realize intelligent adjustment of the vehicle-mounted device, thereby maintaining the comfort of the driver while effectively reducing energy consumption and improving the dynamicity and intelligent level of energy efficiency management.
[0079] Please continue to refer to Figure 1 As shown, the method for optimizing vehicle-mounted energy efficiency based on driver behavior further includes:
[0080] Step S105, execute the control instruction based on the action vector and collect feedback data, calculate the reward value based on the feedback data and update the driver portrait.
[0081] Please refer to Figure 3 As shown in the figure, the strategy optimization method comprises:
[0082] Step S301, execute the control instruction based on the action vector and collect feedback data.
[0083] Specifically, the action instruction At is issued to the corresponding body domain controller through the CAN FD bus, and after the instruction is executed, the battery state of charge change amount ΔSOC, the driver's galvanic skin response and heart rate variability signal are synchronized in the next decision cycle, and the driver's galvanic skin response and heart rate variability signal are input into the pre-trained comfort evaluation model to calculate the instantaneous comfort score Ct.
[0084] Specifically, the collected GSR and HRV signals are input into a pre-trained comfort evaluation model, which is a three-layer fully connected neural network, the input layer dimension is 2 (corresponding to GSR and HRV two features), the output layer dimension is 1, using Sigmoid activation function and scaling to 0-5 points range, calculating the instantaneous comfort score Ct, the comfort evaluation model uses a data set containing 5,000 groups of GSR, HRV data and its corresponding artificial subjective score (0-5 points), the input features are the standard deviation of GSR signal and the two key indicators of HRV RMSSD (root mean square difference of adjacent heart interval), the network structure is a three-layer fully connected network with input layer 2 dimensions, hidden layer 64 dimensions and output layer 1 dimension, using mean square error loss and Adam optimizer for training.
[0085] Specifically, by issuing control instructions and synchronously collecting vehicle states and physiological signals after execution, the system can obtain immediate feedback of action effect, which helps to monitor and adjust the influence on driver comfort and energy consumption in real time, provides key data support for subsequent reward calculation and strategy optimization, and ensures that the system responds timely and effectively.
[0086] Please continue to refer to Figure 3 As shown in the figure, the strategy optimization method further comprises:
[0087] Step S302, calculate the reward based on the feedback data and drive the PPO decision engine to optimize.
[0088] Specifically, the comprehensive reward Rt is calculated, Rt=0.6×(△SOC×1000)-0.4×|Ct-4.2|;
[0089] The complete data of this decision {state vector St, action vector At, comprehensive reward Rt, next state vector St+1} is stored as an experience in the experience replay buffer; wherein the next state vector St+1 is a new state vector obtained by re-executing steps S101 to S104 after a delay of one decision cycle after executing the action At, and the comfort score contained therein is Ct calculated in the current cycle;
[0090] Periodically sample a batch of data from the buffer, perform a PPO algorithm iteration, and update the parameters of the policy network and value network in the PPO decision engine.
[0091] Specifically, by calculating the comprehensive reward and updating the policy using experience replay, the system can continuously optimize the decision model to make energy efficiency control more accurate and stable, and the reward mechanism balances energy saving and comfort maintenance, prompting the system to gradually improve its performance in long-term operation and achieve sustainable energy efficiency optimization.
[0092] Please continue to refer to Figure 3 As shown in the figure, the policy optimization method further includes:
[0093] Step S303, update the driver profile.
[0094] Specifically, the state vector St, action vector At and comfort score Ct of this time are stored after being bound to the ID of the driver;
[0095] Initialize the driver profile: when the system first serves a driver, the temperature preference Taco is initialized to the standard comfortable temperature 22°C, and the operation sensitivity coefficient Ss is initialized to 1.0;
[0096] Run the profile update algorithm every 24 hours, which recalculates the driver's personalized parameters, including temperature preference Taco and operation sensitivity coefficient Ss, based on the data stored in the past 72 hours using a weighted linear regression model;
[0097] The calculation process of the temperature preference Taco is as follows: extract the triplets {outdoor temperature Tout, air conditioner set temperature Tac, comfort score Ct} from the historical data. Bin the outdoor temperature by 2°C intervals, for each temperature interval, count the air conditioner set temperature Tac that gives the highest average comfort score Ct, and record this temperature value as the temperature preference Taco of the driver in the current Tout interval; for the temperature intervals not covered, use the Taco of the adjacent interval for linear interpolation;
[0098] The calculation process of the operation sensitivity coefficient Ss is: calculating the weighted change amount of each action dimension in the adjacent two decision periods from the historical data as the action change amount ΔAt, and the calculation formula of the ΔAt is:
[0099] ΔAt=0.5×|(ΔTac,t)-(ΔTac,t-1)| / 2.0+0.3×|(Raudio,t)-(Raudio,t-1)| / 0.3+0.2×|(Lseat,t)-(Lseat,t-1)| / 2;
[0100] Wherein, the weights 0.5, 0.3 and 0.2 reflect the importance of each action dimension, and the denominators 2.0, 0.3 and 2 are the maximum change range of each dimension action value for normalization; ΔTac,t is the air conditioner temperature adjustment amount at the current time (t), ΔTac,t-1 is the air conditioner temperature adjustment amount at the last time (t-1), Raudio,t is the audio volume adjustment coefficient at the current time (t), Raudio,t-1 is the audio volume adjustment coefficient at the last time (t-1), Lseat,t is the seat heating power gear at the current time (t), and Lseat,t-1 is the seat heating power gear at the last time (t-1);
[0101] Taking ΔAt as the independent variable and the comfort change amount ΔCt=|Ct-Ct-1| as the dependent variable, a weighted linear regression model ΔCt=k×ΔAt is fitted. This regression coefficient k is recorded as the operation sensitivity coefficient Ss of the driver;
[0102] The updated personalized driver portrait parameters, i.e. the temperature preference Taco and the operation sensitivity coefficient Ss, will be combined with the real-time state data to form an enhanced state vector Ste for decision-making, and the specific construction method is as follows:
[0103] The original state vector St is spliced with the personalized portrait parameters to form a new enhanced state vector:
[0104] Ste = [Sd, SOC, Tout, Tac, Pa, Pseat, Ct-1, Taco, Ss];
[0105] In order to realize personalized control by using Ss, an action mask mechanism is designed. First, the amplitude weight vector of each action dimension is predefined:
[0106] For the action set {-2.0, -1.5, -1.0, -0.5, 0} of ΔTac, the amplitude weight is defined as Wtemp=[0.1, 0.3, 0.6, 0.9, 1.0];
[0107] For the action set {0.7, 0.8, 0.9, 1.0} of Raudio, define the amplitude weight as Waudio = [0.1, 0.4, 0.7, 1.0];
[0108] For the action set {0, 1, 2} of Lseat, define the amplitude weight as Wseat = [1.0, 0.6, 0.2];
[0109] Scale each weight vector by the operation sensitivity coefficient Ss as an index to generate the final combined mask:
[0110] Mask = [Wtemp^Ss, Waudio^Ss, Wseat^Ss];
[0111] Element-wise multiply the original action probability distribution Po output by the PPO policy network with the mask Mask, and normalize to obtain the adjusted probability distribution Pfinal = Softmax(Po ⊙ Mask), where ⊙ represents element-wise multiplication.
[0112] Specifically, by analyzing historical data and updating personalized parameters, the system can identify the temperature preference and operation sensitivity of the driver, thereby customizing and adjusting the control strategy. This portrait updating mechanism enhances the personalized adaptation ability of the system, enabling energy efficiency optimization to better meet individual needs, improving driving comfort and energy utilization efficiency.
[0113] Obviously, the above embodiments of the present application are only examples for clearly illustrating the present application, and are not a limitation on the embodiments of the present application. For those skilled in the art, on the basis of the above description, other different forms of changes or variations can be made, and it is impossible to enumerate all the embodiments here. Any obvious changes or variations derived from the technical solutions of the present application are still within the protection scope of the present application.
Claims
1. A method for optimizing vehicle energy efficiency based on driver behavior, characterized in that, Comprise: Collecting multi-modal data stream; Extracting visual spatio-temporal features and vehicle sensor time series features; Fusing visual spatio-temporal features and vehicle sensor time series features to obtain quantitative driving behavior score; Based on driving behavior score, vehicle state data and outside temperature to construct state vector, and input state vector into PPO decision engine to determine action vector; Based on the action vector, execute the control instruction and collect the feedback data, calculate the reward value based on the feedback data and update the driver portrait.
2. The method for optimizing on-board energy efficiency based on driver behavior of claim 1, wherein, On the driver's face video stream, use a pre-trained three-dimensional convolutional neural network to extract the spatio-temporal features between consecutive frames, the structure of the three-dimensional convolutional neural network contains 4 convolutional layers and 2 fully connected layers, and finally outputs a fixed dimension of 512 feature vector, which is used as visual spatio-temporal features, denoted as Fv; For acceleration and angular velocity data, input the acceleration and angular velocity sequence in 1 second time window into a long short-term memory network LSTM, the LSTM hidden layer dimension is 128, and the encoding is a fixed dimension of 256 feature vector, which is used as vehicle sensor time series features, denoted as Fs.
3. The method for optimizing on-board energy efficiency based on driver behavior of claim 2, wherein, The visual spatio-temporal features Fv and the vehicle sensor time series features Fs are input into a cross-modal Transformer fusion module, which adopts 8 attention heads to interact and align the features through multi-head attention mechanism, and the mathematical representation of the fusion process is: Attention(Q,K,V)=Softmax(QKT / dk0.5)V; Wherein, Q is the query vector, which is obtained by linear projection of the visual spatio-temporal features Fv, K is the key vector, which is obtained by linear projection of the vehicle sensor time series features Fs, V is the value vector, which is also obtained by linear projection of the vehicle sensor time series features Fs, QKT is the dot product between the query vector Q and the transpose of the key vector K, Softmax(...) is to convert all attention scores corresponding to each query vector into a probability distribution, dk is the projection dimension, which is set to 32; Finally, the fusion module outputs a 256-dimensional fusion feature vector Ff.
4. The method for optimizing on-board energy efficiency based on driver behavior of claim 3, wherein, The fusion feature vector Ff is input into a scorer, which is composed of two fully connected layers, the output dimension of the first fully connected layer is 128, the output dimension of the second fully connected layer is 1, and the Sigmoid activation function is used after the second layer to map the output value to 0 to 1, as the driving behavior score Sd at the current time.
5. The method for optimizing on-board energy efficiency based on driver behavior of claim 4, wherein, The driving behavior score Sd,t at the current time t, the battery state of charge SOC,t, the outside temperature Tout,t, the air conditioning set temperature Tac,t, the audio output power Pa,t, the seat power gear Pseat,t, and the comfort score Ct-1 calculated in the last decision period are jointly constructed into a state vector St. The state vector St is input into the PPO decision engine, and the policy network of the engine outputs a three-dimensional action vector At, the definition of the action vector At and the discrete action set are as follows: At=[ΔTac, Raudio, Lseat]; wherein, ΔTac is the air conditioning temperature adjustment, whose optional values come from the discrete action set {-2.0, -1.5, -1.0, -0.5, 0}, with the unit of Celsius degree; Raudio is the audio volume adjustment coefficient, whose optional values come from the set {0.7, 0.8, 0.9, 1.0}; Lseat is the seat heating power gear, whose optional values come from the set {0, 1, 2}, corresponding to off, low gear, and high gear, respectively.
6. The method for optimizing on-board energy efficiency based on driver behavior of claim 5, wherein, Execute the control instruction based on the action vector and collect feedback data; Calculate the reward based on the feedback data and drive the PPO decision engine to optimize; Update the driver portrait.
7. The method for optimizing on-board energy efficiency based on driver behavior of claim 6, wherein, The action instruction At is sent to the corresponding body domain controller through the CANFD bus. After the instruction is executed, the battery state of charge change ΔSOC, the driver's skin electrical reaction and heart rate variability signal are synchronized in the next decision cycle, and the driver's skin electrical reaction and heart rate variability signal are input into the pre-trained comfort evaluation model to calculate the instantaneous comfort score Ct.
8. The method for optimizing on-board energy efficiency based on driver behavior of claim 7, wherein, Calculate the comprehensive reward Rt, Rt=0.6×(△SOC×1000)-0.4×|Ct-4.2|; Store the complete data of this decision {state vector St, action vector At, comprehensive reward Rt, next state vector St+1} as an experience in the experience replay buffer; wherein, the next state vector St+1 is a new state vector obtained by re-executing steps S101 to S104 after executing the action At and delaying for one decision cycle, and it contains the comfort score Ct calculated in the current cycle; Periodically sample a batch of data from the buffer to perform a PPO algorithm iteration and update the parameters of the policy network and value network in the PPO decision engine.
9. The method for optimizing on-board energy efficiency based on driver behavior of claim 8, wherein, Store the state vector St, action vector At and comfort score Ct of this time after binding them with the ID of the driver; Initialize the driver portrait: when the system first serves a driver, the temperature preference Taco is initialized to the standard comfortable temperature 22°C, and the operation sensitivity coefficient Ss is initialized to 1.0; Run the portrait update algorithm every 24 hours, which recalculates the personalized parameters of the driver, including the temperature preference Taco and the operation sensitivity coefficient Ss, based on the data stored in the past 72 hours using a weighted linear regression model.
10. The method for optimizing on-board energy efficiency based on driver behavior of claim 9, wherein, The updated personalized driver portrait parameters, i.e. the temperature preference Taco and the operation sensitivity coefficient Ss, will form an enhanced state vector Ste together with the real-time state data for decision-making. Combine the operation sensitivity coefficient in the enhanced state vector with the pre-defined action amplitude weight vector, generate a personalized modulation mask through power operation, multiply the mask with the original action probability distribution output by the policy network, and perform normalization processing to realize personalized modulation of the action decision.