AI-based intelligent starting and accelerating control method and system for electric vehicles

Abstract

The present application relates to the technical field of electric vehicle control, and more particularly to an AI-based intelligent starting and accelerating control method and system for electric vehicles. The method comprises the following steps: obtaining the driving log of the vehicle owner, optimizing the intelligent vehicle after starting, and constructing an intelligent starting mode selection strategy; obtaining physiological state data of the driver, analyzing the state change evolution, and constructing a personalized driving portrait; performing holographic state perception and vehicle virtual simulation, and constructing a vehicle synchronous digital simulation model; collecting radar feedback sensing information and real-time monitoring images in front of the vehicle, performing traffic situation prediction and vehicle scene prediction perception, and constructing a vehicle scene prediction perception model. The present application intelligently selects the driving starting mode of the vehicle, matches the demand of the vehicle owner and the environmental available power constraint, and realizes the maximization of vehicle acceleration efficiency and the optimization of safety.

Description

Technical Field

[0001] This invention relates to the field of electric vehicle control technology, and in particular to an AI-based intelligent start-up and acceleration control method and system for electric vehicles. Background Technology

[0002] With the continuous breakthroughs in artificial intelligence technology, especially the widespread application of deep learning, reinforcement learning, and big data analytics, new solutions have been provided for intelligent start-up and acceleration control of electric vehicles. By integrating multi-source sensor data, driver behavior characteristics, and dynamic environmental information, AI-based intelligent control methods can achieve precise modeling and real-time optimization of the vehicle's powertrain system, effectively improving power response speed and energy efficiency management. Simultaneously, these intelligent control methods possess strong adaptive capabilities, dynamically adjusting start-up and acceleration strategies based on individual driver needs and road condition changes, greatly enhancing driving comfort and safety. However, current electric vehicle start-up and acceleration control technology still faces numerous challenges. On one hand, the complex and ever-changing vehicle operating environment makes it difficult for traditional control algorithms to capture and process rich dynamic information in real time, resulting in insufficient adaptability of control strategies. On the other hand, significant individual differences among drivers make it difficult for existing methods to achieve truly personalized power demand matching, limiting further improvements in system performance. Furthermore, the joint analysis and real-time coordination of complex multi-dimensional parameters such as battery pack status, motor performance, and overall vehicle dynamics still lack efficient intelligent solutions, affecting the overall optimization of the start-up and acceleration process. Therefore, a more intelligent electric vehicle control method is needed. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention proposes an AI-based intelligent start-up and acceleration control method and system for electric vehicles, thereby resolving at least one of the aforementioned technical issues.

[0004] To achieve the above objectives, the present invention provides an AI-based intelligent start-up and acceleration control method for electric vehicles, comprising the following steps: Step S1: Obtain the driver's driving log, optimize the intelligent vehicle after startup, and build an intelligent startup mode selection strategy; Step S2: Obtain driver's physiological state data, analyze state changes and evolution, and construct a personalized driving profile; Step S3: Collect battery pack monitoring parameters, perform holographic status perception and vehicle virtual simulation, and construct a synchronous digital simulation model of the vehicle; Step S4: Based on the vehicle's synchronous digital simulation model and personalized driving profile, predict power demand and redistribute battery pack power based on the intelligent start-up mode selection strategy.

[0005] This specification provides an AI-based intelligent start-up and acceleration control system for electric vehicles, used to execute the AI-based intelligent start-up and acceleration control method for electric vehicles as described above, including: The intelligent start module is used to obtain the driver's driving log, optimize the intelligent vehicle after startup, and build an intelligent start mode selection strategy. The driving profile module is used to acquire driver physiological state data, analyze state changes and evolution, and build a personalized driving profile. The synchronous simulation module is used to collect battery pack monitoring parameters, perform holographic state perception and vehicle virtual simulation, and build a synchronous digital simulation model of the vehicle. The scene prediction and perception module is used to collect radar feedback sensor information and real-time monitoring images in front of the vehicle, perform traffic situation prediction and vehicle scene prediction and perception, and build a vehicle scene prediction and perception model. The adaptive power distribution module is used to predict power demand based on the vehicle's synchronous digital simulation model and personalized driving profile, and to redistribute battery pack power based on the intelligent start-up mode selection strategy.

[0006] The beneficial effects of this invention are specifically as follows: Through historical driving behavior data analysis, the system can identify typical driving patterns after a user starts the vehicle (such as rapid acceleration, slow driving, warm-up, etc.), and then formulate a personalized start-up response curve to achieve a personalized intelligent start-up experience. Historical data analysis helps the system predict the user's start-up rhythm, analyze the user's driving scenarios, such as commuting or traveling, identify the user's urgency level, intelligently select the vehicle's driving mode, and preload the driving mode to better match the owner's travel, optimize the vehicle's pre-start response time and energy pre-allocation, reduce the abruptness of starting, and improve ride comfort. By constructing a real-time driver profile through physiological signals such as heart rate, pupils, and electromyography, the system can determine the driver's current emotional and stress state, providing contextual auxiliary decision-making for start-up and acceleration control. The system can delay start-up and limit the initial acceleration response when it detects that the driver is in an unoptimal state, which helps prevent misoperation and accidents. The profile model will provide a data foundation for subsequent driving style learning and power response adjustment, promoting the evolution of acceleration control towards "context-aware driving". By collecting multiple parameters such as SOC, voltage equalization, and temperature gradient, the system assesses battery health and power delivery capabilities in real time, improving the safety and accuracy of the drive system. Combined with digital twin simulation, the system can simulate the impact of different start-up and acceleration strategies on battery load, thereby developing more reasonable energy allocation and output schemes. Holographic perception helps detect battery anomalies early, reducing wear and sudden failures, maximizing battery life and minimizing maintenance costs. The system can proactively adjust acceleration strategies to match current road conditions by perceiving and predicting factors such as traffic flow, vehicle behavior ahead, and obstacles. Introducing forward vision and radar information into the acceleration control logic effectively avoids over-acceleration or delayed response due to unclear road conditions. By accurately identifying scene types (such as congestion, highways, and open roads), the system can select appropriate acceleration responses to maximize efficiency and optimize safety. The system performs comprehensive calculations based on factors such as battery status, vehicle load, and environmental resistance, dynamically adjusting motor output power to achieve real-time optimal energy utilization. Through model-driven real-time calculations, the system can quickly respond to complex traffic environments such as inclines, declines, congestion, and steep road changes, improving maneuverability and energy efficiency. By combining virtual simulation models with real-world scenario predictions, a three-dimensional control foundation integrating data, physics, and behavior is constructed to achieve optimal matching between global power and acceleration strategies. The system predicts the driver's intended actions (such as rapid acceleration or a steady start) based on a user profile model, pre-adjusting battery power distribution and output modes to ensure consistent response speed and user experience. Through a combination of user profile analysis and real-world environmental judgment, the system dynamically adjusts the acceleration curve slope and output speed to achieve "hand-feel" control. Personalized acceleration strategies help reduce unnecessary high-power output, extend battery life, and lower overall energy consumption, achieving green and energy-saving control goals. Attached Figure Description

[0007] Figure 1This is a flowchart illustrating the steps of an AI-based intelligent start-up and acceleration control method for electric vehicles according to the present invention. Detailed Implementation It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0008] This application provides an AI-based intelligent start-up and acceleration control method and system for electric vehicles. The executing entities of the AI-based intelligent start-up and acceleration control method and system for electric vehicles include, but are not limited to, mechanical equipment, data processing platforms, cloud server nodes, and network upload devices that can be considered general computing nodes in this application. The data processing platform includes, but is not limited to, at least one of an audio-visual management system, an information management system, and a cloud-based data management system.

[0009] Please see Figure 1 This invention provides an AI-based intelligent start-up and acceleration control method for electric vehicles, comprising the following steps: Step S1: Obtain the driver's driving log, optimize the intelligent vehicle after startup, and build an intelligent startup mode selection strategy; Step S2: Obtain driver's physiological state data, analyze state changes and evolution, and construct a personalized driving profile; Step S3: Collect battery pack monitoring parameters, perform holographic status perception and vehicle virtual simulation, and construct a synchronous digital simulation model of the vehicle; Step S4: Based on the vehicle's synchronous digital simulation model and personalized driving profile, predict power demand and redistribute battery pack power based on the intelligent start-up mode selection strategy.

[0010] In the embodiments of the present invention, see Figure 1 This is a flowchart illustrating the steps of an AI-based intelligent start-up and acceleration control method for electric vehicles according to the present invention. In this example, the steps of the AI-based intelligent start-up and acceleration control method for electric vehicles include: Step S1: Obtain the driver's driving log, optimize the intelligent vehicle after startup, and build an intelligent startup mode selection strategy; In this embodiment, a complete driving monitoring log is continuously collected from each vehicle startup to the steady-state driving stage via an in-vehicle T-BOX, central gateway controller, or Advanced Driver Assistance Systems (ADAS) platform. The collected data dimensions include, but are not limited to: vehicle startup time, initial accelerator / electric pedal opening, pedal duration, acceleration command response time, initial acceleration curve, vehicle speed change trajectory, torque response data, battery output power, motor speed, tire slip ratio, vehicle attitude parameters (such as pitch angle), whether startup failure or response delay occurred, and the driver's waiting time before starting the vehicle. These logs are segmented according to the startup cycle, with an average of 80-150 startup behavior records collected per driver per month, forming a high-density behavior sample library. Each startup behavior is analyzed item by item. The first step is to classify startup responsiveness, categorizing behaviors into "rapid response type," "smooth start type," and "hesitant delay type" based on acceleration curves and response times. The second step is to classify the acceleration trajectory after startup using time series clustering (DTW dynamic time warping) technology to identify similar behavior patterns. Taking a real-world experiment as an example, six mainstream startup behavior patterns were identified from 30,000 startup actions by 200 users. The "short response - medium acceleration" pattern accounted for 27.6%, and the "slow start - rapid acceleration" pattern accounted for 21.3%. The third step involved introducing a behavior evaluation model to comprehensively score the energy efficiency, comfort, and safety margin of each startup behavior. Evaluation indicators included acceleration jitter (jerk variation), peak energy consumption, battery temperature rise change, and peak tire slip rate.

[0011] After completing the behavioral analysis, unsupervised learning algorithms (such as K-means++ and Gaussian Mixture Model) are used to cluster and model the driver's starting behavior patterns, classifying drivers into different behavioral preference types, such as "conservative starter," "responsive starter," "energy-saving starter," and "aggressive starter." Simultaneously, the consistency of each type of starting behavior under different environmental conditions (such as sunny days, rainy days, hill starts, and congested traffic) is analyzed, thereby constructing a "starting behavior stability feature vector" for the driver. Based on the above analysis results, an intelligent starting mode selection strategy engine is constructed. In this strategy, the vehicle automatically retrieves the driver's behavioral preference model and current environmental perception data each time it is about to start, and selects the most suitable starting control mode through multi-factor rule matching and weighted fusion decision-making. Available starting modes generally include: energy-saving starting mode, responsive priority mode, comfort and smooth mode, anti-slip control mode, and self-learning optimization mode. For example, when the driver is judged to have a "responsive" driving style and the current environment is a hot and high-adhesion road surface with smooth traffic, the "responsive priority mode" will be activated first; while in a downhill and rainy environment, even if the driver prefers quick start, the "anti-slip control mode" will be activated first to ensure safety.

[0012] Step S2: Obtain driver's physiological state data, analyze state changes and evolution, and construct a personalized driving profile; In this embodiment, driver physiological state data is acquired through multimodal sensing. In practical applications, vehicles can continuously collect driver physiological characteristic information through sensors integrated into seats, steering wheels, dashboards, and even wearable devices (such as PPG photoelectric sensors, EDA conductivity detectors, ECG modules, eye trackers, EEG headbands, temperature / humidity detection modules, etc.). Key data dimensions include heart rate (BPM), heart rate variability (HRV), skin conductance (GSR), palm temperature, eye movement frequency and duration of eye closure, facial expression tension, electromyographic activity, and palm sweating levels. This data is typically sampled at 500ms to 1s intervals and preliminarily processed and cached in the vehicle controller to form time-series data of the driver's physiological state during driving. For example, during a long driving session, the driver's heart rate, skin resistance, and pupil dilation status are recorded once per second, forming a data vector set of over 3000 continuous state nodes. The process then proceeds to the state change evolution analysis stage. In this process, Dynamic Time Warping (DTW), Principal Component Analysis (PCA), and variability detection algorithms are used to perform pattern recognition and trend extraction on continuous physiological state vectors. This allows for the identification of physiological state evolution patterns in drivers under specific driving tasks or situations (such as high speed, nighttime driving, urban congestion, and frequent stop-and-go driving). For example, by clustering the skin conductivity and heart rate fluctuations of 50 drivers during rush hour, it was found that 32 of them exhibited a clear "emotional tension-high responsiveness" pattern, while the other 18 showed a "emotional stability-low intervention" trend. This analysis process further compares the correlation with driving behavior data (such as accelerator pedal pressure, frequency of rapid acceleration events, and changes in vehicle yaw rate) to clarify the degree of coupling between physiological state changes and actual behavior. Based on these behavior-physiological state correlations, a personalized driving profile modeling engine is constructed. This model uses a Graph Neural Network (GNN) or Transformer structure, treating each type of physiological state node as a node entity in the graph, with different behavioral triggers or state transitions constituting the edge weight relationships in the graph. The model training process is based on supervised label learning and partially self-supervised learning mechanisms. By continuously mapping feature vectors of individual drivers' data in multiple different scenarios, a personalized driver state profile that can be updated in real time is constructed. This profile includes not only behavioral style classifications (such as conservative, tense, and aggressive) but also response patterns in different states (e.g., increased acceleration response when heart rate increases, and increased steering frequency when GSR increases), providing important references for subsequent control strategy selection. The profile model is not a static template but has adaptive update capabilities. After each driving session, the deviation between the profile's predicted behavior and the actual behavior is compared. If a long-term deviation exists, the model update mechanism is automatically triggered to retrain some neural network weights to maintain consistency between the profile and the driver's state.This mechanism enables dynamic modeling that is "personalized and time-dependent," ensuring that core driving characteristics and behavioral logic can be accurately captured even when the driver's state is affected by factors such as changes in lifestyle, mental state, or age.

[0013] Step S3: Collect battery pack monitoring parameters, perform holographic status perception and vehicle virtual simulation, and construct a synchronous digital simulation model of the vehicle; In this embodiment, the vehicle battery management system (BMS) is at the core. A high-precision sensor network distributed inside and around the battery pack continuously acquires multi-dimensional monitoring parameters during battery pack operation. These parameters mainly include individual cell voltage (typically with a sampling accuracy of ±5mV), battery module temperature (sampled by thermistors or thermocouples, with an accuracy of ±0.5℃), battery SOC (State of Charge), SOH (State of Health), battery internal resistance, charging and discharging current, battery pack heat flow changes, insulation status, and external cooling conditions. Data is refreshed at a frequency of 1Hz to 10Hz and can be synchronized and cached with the central controller via the vehicle's CAN bus, Ethernet, or OTA connection. Taking a mass-produced electric vehicle as an example, approximately 200,000 battery operating status data points can be collected per hour, covering more than 50 physical quantities, ensuring sufficient timeliness and resolution for dynamic modeling. The process then enters the holographic state perception stage. This process first uses multi-source data fusion algorithms (such as Kalman filtering and multi-dimensional time series fitting) to denoise, normalize, and remove outliers from the aforementioned sensor data, ensuring the reliability of the data input to the model. Building upon this foundation, local clustering and time-series aggregation methods are used to identify temperature gradients, voltage differences, and impedance evolution trends between different regions and modules within the battery pack. For example, during high-power acceleration, it can be identified that the temperature rise rate of the module in the lower right corner of the battery pack is faster than that of other regions, and the internal resistance growth rate of the corresponding module is 1.8 times the average, initially indicating a risk of local thermal imbalance. This type of analysis is organized using a graphical structure, with each cell or module as a node and the thermal and electrical behavior coupling relationships between them as edges. By constructing a "battery holographic state map," a comprehensive perception, dynamic prediction, and safety warning of the battery's operating state are achieved. Furthermore, a SOX dynamic evolution model (including SOC, SOH, SOP, etc.) will be introduced to predict the remaining battery capacity under different usage conditions, thereby achieving a dynamic extension from "current observation" to "future prediction." After completing the holographic perception, the vehicle virtual simulation stage will be entered based on these multi-dimensional perception data and evolution maps. In this phase, battery thermal management models, motor control models, chassis dynamics models, and external environment models will be integrated to construct a vehicle-level synchronous simulation structure on a digital twin platform (such as one based on MATLAB / Simulink, CarSim, or a self-developed physical simulation platform). The focus is on the battery pack's response changes under different operating conditions (such as 0–60 km / h acceleration, continuous uphill driving, frequent start-stop, and energy recovery), including output capacity, temperature rise rate, pressure differential limit, and range degradation. For example, under the conditions of an ambient temperature of 35℃, a battery SOC of 65%, and a fully loaded vehicle, three simulated start-up accelerations were performed. The simulation results show that when the third start-up reached 4 seconds, the battery thermal response significantly increased, with a single-cell temperature difference reaching 6.2℃, while the maximum output power decreased by approximately 12%, indicating a performance degradation problem under continuous load.Such simulations will be repeated to build a multi-scenario, multi-variable battery response database for the entire vehicle.

[0014] Step S4: Based on the vehicle's synchronous digital simulation model and personalized driving profile, predict power demand and redistribute battery pack power based on the intelligent start-up mode selection strategy. In this embodiment, a vehicle synchronous digital simulation model is used as input to quantitatively analyze the current operating state and potential of the vehicle, thereby calculating a "Vehicle Comprehensive Performance Index" (VCPI) representing the current performance limit of the vehicle. This index is mainly calculated based on the real-time output capability of the battery pack, motor efficiency, electric drive response delay, chassis load status, and tire adhesion. Its core calculation relies on a multi-dimensional dynamic weight fusion model, which standardizes performance parameters from different physical domains and assigns them weights. For example, in a simulation run, 10 acceleration simulations of a certain model of electric vehicle were performed from 0 to 60 km / h. The battery SOC was maintained at 60%, the average maximum motor output torque was 180 N·m, and the tire sideslip angle remained stable within 2.5°. The measured VCPI value under this condition was 0.84 (out of 1.0), indicating that the vehicle still has high acceleration potential, but the battery temperature is close to the safety threshold (42.3°C), and subsequent power release needs to be controlled. Environmental constraint parameters obtained from the vehicle scene prediction and perception model are introduced, especially key factors such as road slope, adhesion coefficient, wind resistance parameters, traffic density, and distribution of dynamic objects ahead. The goal of this stage is to transform environmental constraints into power release boundaries, that is, to determine the safe and reasonable acceleration magnitude and power limit that the vehicle can release under the current traffic and road conditions. By simulating the vehicle response and environmental interaction under different accelerations, and combining scene simulation tools (such as SUMO or a self-developed traffic flow prediction engine), a "power release envelope" is constructed. For example, in a real-world scenario, if the current road slope is 3.5%, the adhesion coefficient is 0.65, and there are two pedestrians and a slow-moving vehicle 15 meters ahead, and it is predicted that there is a 75% probability of needing to actively brake within the next 2 seconds, the maximum release acceleration is limited to 1.2 m / s², and a limiting window (0~140 N·m) is applied to the motor torque.

[0015] The system employs multi-objective optimization algorithms to simultaneously satisfy three control objectives: power responsiveness, energy economy, and safety boundaries. Common methods include reinforcement learning-based policy optimization (such as DDPG and PPO), mixed-integer programming (MILP), and nonlinear constraint optimization (such as SQP). The Vehicle Comprehensive Performance Index (VCPI) is used as the core parameter of the current power capability, and the power release parameters in the scenario prediction model are used as boundary constraints to construct an objective function for optimization. The optimization process updates on a second-level basis, adjusting the power distribution curve in real time to ensure that the vehicle can achieve smooth and efficient starting under actual operating conditions without triggering energy consumption peaks or exceeding safety thresholds. For example, in one simulation, the optimal power release path under a certain acceleration condition was evaluated as follows: initial stage controlled at 90 N·m, mid-stage peak release at 130 N·m, and final stage converged to 70 N·m, with overall power consumption controlled at 1.8 kWh / 100 km, acceleration time of 6.3 seconds, Jerk index controlled within 1.4 m / s³, and user satisfaction score better than 80%. A state-space model will be used to construct the power distribution mapping relationship, mapping specific driving needs, vehicle states, and environmental characteristics to the corresponding torque-current distribution matrix, forming a state-driven responsive power control logic. This logic supports dynamic adjustment and feedback correction mechanisms for each acceleration request, enabling rapid responses to sudden conditions and complex traffic situations. For example, when it detects that a sudden lane change by a vehicle ahead reduces the available acceleration space, it can immediately reduce the acceleration torque output and activate energy recovery logic to achieve a stable control transition.

[0016] In this embodiment, historical driving data is used for time-series modeling, employing recurrent neural networks (RNNs) or long short-term memory networks (LSTMs) to predict driver actions in specific situations. For example, it is identified that a driver in urban traffic during rush hour frequently uses a "short-duration high-intensity acceleration + rapid deceleration" pattern, with an average starting torque demand of approximately 120 N·m and an average acceleration of 1.6 m / s² within 3 seconds of starting. Simultaneously, behavior can be linked to physiological state maps; for instance, when a driver is under high stress, their power demand pattern tends to be more aggressive. Finally, a "target power demand prediction value" is output, including the predicted torque value, acceleration value, duration, and acceptable delay range. The current power distribution strategy is dynamically adjusted based on the predicted target power demand, with a focus on battery pack power reallocation. This not only needs to meet the driver's power request but also needs to consider the battery pack's current output capacity, temperature control status, health condition, and safety boundaries. During power redistribution, the system utilizes internal battery pack state information (such as SOC, SOH, voltage consistency, and thermal distribution for each module) and employs a multi-constraint optimization algorithm to adjust the power output weights of each module. For example, when the predicted acceleration request requires 150 N·m of torque, lasts 2.8 seconds, and has an estimated peak power demand of 52 kW, the system assesses the load capacity of multiple battery modules. It finds that module A's current temperature has reached 45°C, nearing its thermal management limit, while module C is under lower load at 32°C. Therefore, the current path is redistributed, increasing the output ratio of module C and reducing the power release of module A, ensuring the battery maintains thermal balance and health while meeting the demand. This process is completed collaboratively by a dynamic current controller and a model predictive control (MPC) algorithm, with real-time corrections made dozens of times per second. This generates a personalized "intelligent matching acceleration control engine" with customized drive characteristics. This engine is not a single control model but an intelligent control framework with multiple input parameters, dynamic response capabilities, and individual adaptability. It integrates driver profile parameters, environmental perception parameters, vehicle status parameters, and battery pack limitation parameters, enabling it to make the optimal response choice before each acceleration request is triggered. The control engine uses reinforcement learning (such as PPO or DDPG) combined with rule guidance to fine-tune the strategy: for example, for the same acceleration request, different drivers will trigger different torque release paths and energy consumption management strategies. For Type A drivers (conservative), the energy-optimal path is prioritized; for Type B drivers (aggressive), response speed and acceleration are prioritized; and for Type C drivers (conservative), the upper limit of power release is further compressed, and more safety buffers are introduced.

[0017] In this embodiment, the detailed implementation steps of step S1 include: Obtain the vehicle owner's driving log; Define a time-series behavior period and extract the vehicle startup behavior from the owner's driving log to obtain the vehicle startup behavior data stream within the period; The driving behavior is analyzed sequentially after each vehicle startup from the vehicle startup behavior data stream to generate driving behavior features; Extract the vehicle startup timestamp from the vehicle startup behavior data stream; Perform periodic time series fitting on the timestamps to construct a periodic timestamp marker map; Intelligent vehicle startup optimization is performed based on driving behavior characteristics and periodic timestamp markers to construct an intelligent startup mode selection strategy.

[0018] In this embodiment, the owner's historical driving data logs are collected and summarized through intelligent in-vehicle systems (such as OBD data acquisition modules, CAN bus access modules, BMS, T-Box remote communication modules, etc.) to establish a basic data pool for subsequent intelligent start-up mode strategies. The collected data mainly covers the following five categories: 1) vehicle start-up time and motor activation time; 2) vehicle speed changes, acceleration curves, and braking force within the first 10 minutes after start-up; 3) battery SOC (state of charge) and voltage fluctuations during start-up; 4) environmental data, such as temperature, humidity, and traffic density; 5) driving operation behaviors, such as accelerator pedal response and throttle response curves. The data sampling frequency is recommended to be set to 1Hz~10Hz to ensure sufficient capture of vehicle state changes. The collection period is recommended to cover at least 30 days of continuous driving behavior to reflect the owner's stable usage habits. After collection, the data needs to be cleaned and normalized, including outlier removal (such as GPS drift, current surges), timestamp standardization, and data null value interpolation, ultimately outputting a unified format driving monitoring log sequence. This study extracts the "vehicle start-up behavior cycle" from typical driving behaviors of car owners and constructs a periodic start-up behavior data stream. First, by performing time-series frequency statistics on historical data, high-frequency daily start-up periods (such as 7:30, 17:40, etc.) are identified. A start-stop clustering algorithm based on a sliding time window (such as DBSCAN + time density clustering) is then used to identify periodic behaviors. Each "vehicle ignition and start-up to stable acceleration phase" is defined as a "start-up behavior unit," thus dividing the starting point of each cycle. All relevant data within this cycle (such as start-up current, voltage, ambient temperature, acceleration changes, etc.) are extracted and combined to form a start-up behavior data stream within the cycle.

[0019] To improve the accuracy of cycle identification, a minimum number of cycle initiation behaviors (e.g., no less than 3 times / week) can be set as a filtering condition. Each cycle data stream consists of multiple fields, including initiation timestamp, vehicle speed change rate, battery voltage response, acceleration change curve, etc., providing a complete input structure for subsequent behavior pattern recognition. Quantitative analysis is conducted on driving performance "after initiation behavior" to extract the driver's driving behavior characteristics. Methodologically, a segmented time window analysis technique is used, setting the first 1 minute, 3 minutes, and 5 minutes after initiation behavior as different time windows to extract key driving behavior indicators, such as average acceleration, maximum vehicle speed, throttle response curve slope, frequency of rapid acceleration, and energy consumption growth rate, forming a multi-dimensional feature vector of behavior. To model the stability and individual characteristics of driving behavior, change rate and fluctuation amplitude are introduced as supplementary features, such as acceleration standard deviation, voltage fluctuation amplitude, and number of braking abrupt change points. Arranging these behavioral features in chronological order can construct a "driving behavior evolution map". Simultaneously, a Dynamic Time Warping (DTW) algorithm is introduced to compare the similarity of behavioral patterns across different periods, thereby identifying the consistency of the driver's behavior and typical driving patterns. These behavioral features serve as the input basis for the intelligent start-up optimization algorithm, used to predict the energy output curve and acceleration expectation model required by the vehicle after start-up, improving the accuracy and comfort of the start-up response. In the periodic data stream, the start-up timestamp is the starting marker for each "ignition + motor activation," directly related to the temporal distribution pattern of driving behavior and daily habits. In this step, the motor start-up signal points recorded by the vehicle control unit (VCU), combined with GPS data and the current activation signal output by the BMS, are first used to extract the start-up timestamps (precisely recorded in the form of year, month, day, hour, minute, and second) from all historical data. Subsequently, these timestamps are organized to form a continuous "start-up time sequence."

[0020] Further processing can be performed by aggregating timestamps at a weekly or monthly granularity, and drawing a start-up density map with a 24-hour time period to identify high-frequency start-up periods. In addition, daily and weekly start-up frequencies and standard deviations are statistically analyzed to determine if car owners have highly regular travel habits. This type of timestamp information will be an important component of the time variable in the subsequent construction of a periodic prediction model. After obtaining the start-up timestamp sequence, a "periodic time distribution model" of vehicle starts needs to be constructed. This step uses periodic function fitting techniques to perform time series analysis on the start-up time series, such as extracting the main period frequency based on Fourier transform (FFT), or using weighted least squares fitting to approximate high-frequency start-up periods. The results are presented in the form of a periodic timestamp marker map, which divides a 24-hour day into multiple smaller time periods (e.g., 15 minutes / segment), marking the probability density of start-up events within each segment. For example, if a high probability of start-up occurs consecutively within the time period of 07:15~07:30, it will be marked as a "high-confidence periodic start point" and analyzed in conjunction with corresponding driving behavior characteristics. This timestamp map not only reflects the typical starting rhythm of the car owner, but also provides a temporal basis for the next step of intelligent mode selection. Especially in the comparison of weekdays and rest days, this map can be used to determine the degree of rhythm variation, so as to achieve more granular scene recognition. A driving mode classification model is constructed (using unsupervised clustering algorithms such as K-Means and HDBSCAN) to classify different combinations of driving behavior features into modes, such as "steady-state cruise", "frequent rapid acceleration", and "urban commuting". Based on the periodic timestamp map, the current time is matched to predict the type of starting cycle that the user is about to enter, and joint reasoning is performed with the behavioral feature model. Finally, intelligent recommendation of starting mode based on "time + behavior" is realized. For example, if it is identified as "commuting + 07:30 high frequency cycle", the starting current output curve, optimal energy release path, and driving torque response model will be intelligently set at the moment of ignition, and the BMS current preheating mechanism will be activated in advance to shorten the acceleration response delay and improve the overall driving experience. The implementation of this strategy is recommended to combine hardware and software deployment: the strategy is invoked by the onboard intelligent control unit, and executed through closed-loop regulation by the VCU-motor controller, in conjunction with continuous optimization using a cloud-based periodic behavior model. Ultimately, this forms a personalized, scenario-adaptive intelligent start-up mode selection mechanism for vehicle owners, driving electric vehicles towards greater intelligence, energy efficiency, and high performance.

[0021] In this embodiment, the specific steps for optimizing intelligent vehicle startup based on driving behavior characteristics and periodic timestamp markers, and constructing an intelligent startup mode selection strategy are as follows: Identify the real-time entry time of the vehicle owner's vehicle; The optimal similarity matching calculation is performed on the real-time vehicle entry time of the car owner based on the periodic timestamp marker map to obtain the most similar start timestamp in history. Based on the most similar historical startup timestamp, analyze the car owner's usage scenario to obtain the current car owner's usage scenario; The travel time delay is calculated based on the most similar start timestamp in history to obtain the travel delay parameters for this trip; The vehicle driving mode is intelligently selected based on driving behavior characteristics and the delay parameters of this trip to obtain a pre-loaded driving mode; Intelligent vehicle startup optimization is performed based on pre-loaded driving modes, and an intelligent startup mode selection strategy is constructed.

[0022] In this embodiment, the starting point of the user's actual vehicle use is captured to provide a time reference for subsequent decisions. Vehicle entry time identification primarily relies on interaction signals between the owner and the vehicle, including key sensing, door unlocking records, smartphone app connection, NFC, or Bluetooth near-field communication. For example, when the owner approaches the vehicle with the smart key at a certain distance (usually within 1.5 meters) and simultaneously triggers the unlocking action, this time point is recorded as the starting time of vehicle entry. To improve the accuracy and anti-interference capabilities of identification, multiple signal fusion is typically used to determine the occurrence of the entry event. For example, the owner's phone successfully pairs with the vehicle via Bluetooth, the door unlocking signal is reported, the in-vehicle seat sensors are activated, and multiple signals are cross-verified before triggering the "vehicle entry event" confirmation logic. The timestamp of this event is a high-precision record at the millisecond level, typically using UTC standard time synchronization. During the experiment, the recognition mechanism was deployed on 100 electric vehicles, and behavioral data of vehicle owners entering the vehicles were collected. On average, each person recorded about 2 to 3 valid entry events per day, and the recognition accuracy rate reached over 98%. For users with regular commuting habits, the entry time was highly consistent with their periodic behavior, providing a solid foundation for subsequent time matching and behavior prediction.

[0023] After obtaining the real-time vehicle entry time, this time point is compared with the owner's periodic timestamp marker map for optimal similarity matching. The periodic timestamp marker map is a daily time behavior map constructed based on vehicle start-up data from the past 30 to 90 days, containing the probability distribution of vehicle start-ups at different times of the day. The current entry time is matched with all marked times in this map, using a sliding time window search strategy combined with a Gaussian weighted mechanism or a simple time distance sorting algorithm to find one or more historical start-up time records closest to the current time. For example, if the current time is 07:48, the high-frequency start-up times between 07:45 and 08:00 every day in the past 28 days are found in the periodic map, and the record closest to the current time is selected as the "most similar historical start-up timestamp". The matching process not only considers time distance but also incorporates multiple constraints such as whether the driving behavior corresponding to the time point on that day is complete, whether the battery usage is normal, and whether there is a complete travel route after start-up, to avoid selecting abnormal data as matching samples. In the experiment, a matching calculation could be completed on average within 2 seconds, with matching accuracy maintained at over 92%, and the performance was more stable, especially for users with regular schedules.

[0024] After obtaining the most similar historical startup timestamp, the vehicle enters the usage scenario analysis phase. The aim is to infer the owner's current usage intent and scenario type, aiding in decisions such as whether preheating is necessary and choosing between power and energy-saving modes. This analysis is based on a behavioral label classification model. The model's training data comes from driving behavior data within 30 minutes of startup at the most similar historical time point, including dimensions such as driving time, average speed, maximum speed, battery level drop, and whether highways or complex road sections were involved. These behavioral characteristics are used to train and model typical usage scenarios (such as morning rush hour commuting, weekend short shopping trips, picking up children, and nighttime short trips), and inference is performed based on similarity algorithms or supervised classification models (such as decision trees or support vector machines). For example, assuming that at 07:50, most historical trips last around 20 minutes, the route points to the office area, and the average speed is consistently below 30 km / h, it can be determined that the current scenario is most likely a "daily commuting trip." Statistical analysis of a large number of user samples revealed high stability and repeatability in vehicle usage behavior at specific times, especially on weekday mornings and evenings, where the consistency rate of user behavior scenarios reached as high as 87%. Weekend scenarios were more diverse but still showed certain patterns. The system calculates the time difference between the most similar historical start timestamp and the current vehicle entry time to determine if there is a delay or advancement in the current trip. This delay parameter is not just a simple difference in absolute time but also an important indicator reflecting changes in the user's travel rhythm. If the current entry time is 15 minutes later than a similar historical time, it is inferred that the user's travel plan may have been delayed due to some temporary factors, thus requiring a reassessment of the driving response strategy. Experimental data shows that among commuting users, the probability of a travel time delay exceeding 10 minutes is 22%, with delay times typically fluctuating between 5 and 20 minutes. To quantify the impact of this delay on vehicle start response, delay times are categorized into levels: less than 5 minutes is considered normal, 5 to 15 minutes is moderate delay, and more than 15 minutes is high delay, with each level corresponding to a different response strategy. Especially for situations involving delayed travel, it anticipates that drivers may want to reach their destination as soon as possible, thus providing a more proactive acceleration response during subsequent startup optimization.

[0025] Based on the driving behavior characteristics, usage scenario categories, and current travel delay parameters analyzed above, the vehicle enters the intelligent driving mode selection phase. Controlled by a multi-input, multi-factor AI decision-making model, this model integrates periodic behavior maps, scenario labels, driving style, and delay conditions to determine whether to select energy-saving, standard, power, or adaptive driving modes. For example, if the current behavior is a weekday commute, and the driver has historically preferred faster acceleration in the same scenario, and the current travel time delay exceeds 10 minutes, they are more likely to choose the power response mode to provide higher initial acceleration and acceleration response frequency. The model primarily employs ensemble learning structures, such as XGBoost or reinforcement learning strategies based on LSTM time series modeling. The model's training data comes from thousands of past driving behavior records, and the output decision value is the mode label and corresponding control parameter configuration. In actual deployment, the driving mode decision can be completed within 2-3 seconds of vehicle entry, quickly responding to the driver's personalized needs. The model's self-learning capability also allows it to optimize in real time based on driver feedback, continuously adjusting its understanding of driver habits over long-term use.

[0026] In this embodiment, the detailed implementation steps of step S2 include: Based on the real-time acquisition of driver's heart rate variability, eye movement trajectory, blink frequency and closure degree by vehicle-mounted multimodal sensor array, driver physiological state data is constructed. Multi-time point state change evolution analysis is performed on the multi-dimensional feature vector to obtain multi-time point state evolution features; The potential driving intentions and the urgency of driving scenarios are mined from the multi-time point state evolution characteristics to generate potential driving intentions and the urgency of driving scenarios. Personalized real-time profiles are built based on potential driving intentions and the urgency of driving scenarios to create personalized driving profiles.

[0027] In this embodiment, a multimodal sensor array on-board is used to collect real-time data on various physiological indicators of the driver, such as heart rate variability (HRV), eye movement trajectory, blink frequency, and eye closure degree, to construct a multidimensional feature vector of the driver's physiological state. The sensor array typically includes a photoplethysmography (PPG) sensor embedded in the steering wheel, an infrared / near-infrared camera in the cockpit (for eye tracking and eye closure detection), and a micro-vibration sensing module in the seat. These sensors can provide 30 data samples per second, ensuring that changes in physiological state are captured at a high frequency. In the experimental setup, this system was deployed on 50 electric vehicles, and daily driving of 200 users was tracked for up to 3 months. Heart rate variability was extracted by analyzing the standard deviation of the RR interval (SDNN) to extract the autonomic nervous system state; eye movement trajectory was modeled by the movement path of the pupil center within the visual field; blink frequency was based on the number of blinks per minute, combined with the duration of each blink closure to determine the degree of fatigue. These data are uniformly timestamped and standardized, and then formed into a multidimensional physiological state vector (usually including 1015 feature dimensions) according to a fixed time window (e.g., every 5 seconds) to form the basic expression of the driver's physiological state.

[0028] Multi-time-point state change evolution analysis was performed on the aforementioned multi-dimensional feature vectors to uncover the dynamic evolution characteristics of the driver's physiological state within a short period. Specifically, a sliding time window (e.g., the past 60 seconds, with a step size of 5 seconds) was used as the unit to input the feature vector sequence at different times into a time series modeling framework, such as a Long Short-Term Memory (LSTM) network or a Gated Recurrent Unit (GRU) network, thereby capturing the trend changes and sudden patterns of the driver's physiological state within a unit of time. For example, if HRV shows a continuous downward trend, while blinking frequency decreases and eye closure increases, it indicates that the driver is gradually entering a state of fatigue or decreased attention. In the experiment, an LSTM-based time evolution feature extraction model was trained to label and identify five typical state evolution trends (focused → distracted, awake → fatigued, relaxed → anxious, alert → tense, stable → disordered), achieving a classification accuracy of over 88%. This model can continuously output the evolution label and state change rate index of the driver's state at the current time point, providing in-depth behavioral evidence for subsequent driving intention and scenario assessment. The time-point state evolution characteristics are input into the potential driving intention recognition and driving scenario urgency assessment modules to further infer the driver's current possible driving intentions (such as rushing, relaxed driving, inattentiveness, etc.) and determine whether the driving task is in a high-risk, high-stress state. The intention recognition part uses a multi-layer neural network model for classification tasks. Its training samples are based on real driving behavior and physiological state data, and a supervised sample set is established through driver subjective questionnaires and behavioral labels. The identified driving intentions include five typical types: "rushing", "relaxed", "passive fatigue", "alert defensive", and "highly focused". Each intention type corresponds to a specific set of physiological state evolution characteristics. For example, "rushing" is often characterized by a decrease in HRV but an increase in eye movement frequency, reduced blinking, and high concentration. "Relaxed" is characterized by a higher HRV and a relatively smooth eye movement path. At the same time, the urgency of the driving scenario is assessed by combining decision trees with behavioral statistical models, such as the frequency of sudden acceleration, the number of sudden accelerator pedal changes, the degree of synchronization between route congestion and the driver's physiological reactions, and a comprehensive score is used to output the urgency level of the scenario. The assessment results are divided into three categories: "low urgency", "medium urgency" and "high urgency". Experimental data shows that the model maintains an accuracy rate of over 90% in scenario assessment under actual driving conditions and can provide early warning of sudden dangerous situations 1 to 2 minutes in advance.

[0029] After obtaining dual assessments of potential driving intentions and the urgency of the driving scenario, a personalized real-time profile model is created for the driver, forming a dynamic driver state profile. This profile not only includes the current physiological state and behavioral trends but also records changes in similarity to historical behavioral profiles, forming a dynamic driver adaptation model. Profile modeling is constructed using vector graphs; each driver has a "psychological-behavioral graph" that evolves over time. High-dimensional features are compressed using an autoencoder, and the profile relationship structure is modeled using a graph neural network (GNN). Before each driving start, the most suitable driving mode and control parameters are dynamically recommended based on the profile state combined with intention and scenario level. For example, if the system detects that the user is currently in a "rushing intention + high urgency state," and the historical profile indicates a high acceleration tendency in similar situations, the "power enhancement + response optimization" control mode will be loaded into the startup strategy. The profile model also supports real-time iterative updates, proactively fine-tuning the next startup and control strategy based on feedback from the differences between the most recent driving behavior and the profile. Experimental results show that after adopting a profile-driven personalized control strategy, users' driving satisfaction scores increased by 14%, and the accuracy of predicting behavioral deviations increased to over 93%, significantly improving the adaptive capability of intelligent control.

[0030] In this embodiment, the detailed implementation steps of step S3 include: Collect battery pack monitoring parameters; Temperature distribution analysis is performed on the multi-dimensional monitoring parameters, voltage difference calculation and internal resistance change identification of individual cells are performed, and holographic state perception is performed to construct a holographic state perception map of the battery pack. Data on motor rotor position, stator temperature, and magnetic field strength are collected, and motor performance is analyzed to obtain a real-time profile of motor performance. The vehicle attitude, tire pressure, and suspension load are acquired, and chassis state information is obtained by fitting. Perform vehicle dynamics evolution analysis on chassis status information to construct a vehicle dynamics feature map; Virtual vehicle simulation is performed on the vehicle dynamics feature map, real-time motor performance profile and battery pack holographic state perception map to build a synchronous digital simulation model of the vehicle. Collect radar feedback sensor information and real-time monitoring images of the front of the vehicle to perform traffic situation prediction and vehicle scene prediction perception, and build a vehicle scene prediction perception model.

[0031] In this embodiment, multi-dimensional monitoring parameters of the battery pack are acquired in real time through the vehicle battery management system (BMS). These parameters include information such as voltage, current, temperature, internal resistance, SOC (state of charge), and SOH (state of health) of each individual cell, with a sampling frequency typically between 15Hz. Taking a 75kWh ternary lithium battery pack as an example, it usually consists of 500 cells distributed across several modules. The operating status of each cell is recorded individually, and a spatial topology between the cells and modules is established for subsequent thermal distribution and voltage consistency analysis. The collected data is transmitted to the central computing module via CAN bus or Ethernet, where data integrity and sampling frequency are verified in real time to ensure that each frame of data can support high-resolution time-series analysis. This process is the foundation for building battery state awareness, effectively identifying potential battery anomalies, aging trends, and thermal management risks.

[0032] After data acquisition, these multi-dimensional monitoring parameters are comprehensively analyzed, including battery temperature distribution modeling, cell voltage difference measurement, and internal resistance change trend identification. The temperature analysis section uses thermal imaging fitting based on the spatial arrangement of modules and cells, simulating temperature diffusion paths under high load or fast charging conditions using a thermal balance model to identify "hot spots" or uneven cooling issues. Cell voltage difference analysis calculates voltage consistency based on the maximum and minimum values ​​between individual cells, while using time-series voltage response curves to determine if there are sluggish cells or individuals with abnormal capacity reduction. For internal resistance analysis, the dynamic change trend of AC internal resistance is deduced from the voltage response changes during charging and discharging, and aging characteristic points are identified by combining historical data. All this information is integrated into a multi-layered "Battery Pack Holographic State Perception Map." This map uses modules as basic units, with each node carrying composite attributes such as temperature, voltage, and internal resistance. The edge structure represents the thermal / electrical coupling relationship between cells. The establishment of this map provides structured support for subsequent battery management optimization, early warning mechanisms, and energy dispatch strategies. Based on 300 hours of actual vehicle operation data, this perception map can accurately capture more than 80% of abnormal cell behavior and temperature control deviations, greatly improving perception accuracy.

[0033] Key operating parameters of the motor, including rotor position, stator temperature, and magnetic field strength, are collected as core data for motor performance analysis. Rotor position is provided in real-time by a high-precision encoder or Hall effect sensor, with a resolution of up to 0.1°, used to assess the dynamic response and angular velocity changes of the motor rotation. Stator temperature is obtained through thermocouples with embedded coils or infrared temperature measuring elements, with a measurement range between -40°C and 180°C, reflecting the motor's thermal load in real time. Magnetic field strength is sensed by a flux sensor or magnetoresistive effect device to detect the uniformity and instantaneous changes in the magnetic field distribution. These parameters are used to model motor efficiency, thermal stability, and magnetic energy conversion efficiency, combined with multiple vehicle operating load characteristics, to form a "real-time motor performance profile." This profile includes not only efficiency curves and thermal efficiency maps, but also indicators such as electromagnetic torque fluctuation rate, response lag time, and high-temperature risk areas. Through a dynamic update mechanism, the profile is refreshed every 5 seconds. In an experiment involving 1,500 kilometers of measured data, the real-time profiling accurately predicted six early signs of motor overheating and two abnormal magnetic flux deviations, significantly enhancing the motor's fault prediction capabilities and adaptive adjustment foundation.

[0034] Further real-time status information of the vehicle chassis is obtained, including vehicle attitude (such as pitch, roll, and yaw rate), tire pressure, and dynamic load distribution of the suspension. Vehicle attitude is typically obtained through a six-axis inertial measurement unit (IMU) with a sampling frequency of over 100Hz per second; tire pressure is continuously monitored via TPMS, reported once per minute; and suspension load relies on force sensors at the four corners of the vehicle or electronically controlled damping feedback parameters, reflecting the suspension compression degree and force transmission state in real time. By fusing these sensor data and utilizing vehicle kinematics models and multibody simulation, the real-time operating state of the chassis is fitted. This state not only expresses the vehicle's stability under different road conditions but also predicts its response capability during extreme handling. For example, during cornering, the roll limit can be determined based on real-time suspension load and attitude changes, and a risk signal can be fed back to the driver assistance module. In experimental testing, the fitted model accurately reconstructed the chassis state under high-dynamic handling and successfully identified 16 attitude imbalance events during high-speed obstacle avoidance, providing predictive support for the dynamic control of the entire vehicle.

[0035] By fusing chassis state information with motor response data, a vehicle dynamics evolution analysis was conducted, resulting in the construction of a "vehicle dynamics feature map." This map, with time as the main axis, constructs the dynamic behavior evolution path through multi-dimensional feature vectors, covering key dimensions such as acceleration changes, yaw rate response, motor torque distribution, suspension stress state, and vehicle attitude change trajectory. The map construction employs methods such as time-series data clustering, principal component analysis, and correlation graph analysis, efficiently extracting the vehicle's dynamic response characteristics under different driving modes. For example, the feature map clearly shows that in "urban congestion mode," the vehicle's dynamic characteristics exhibit a small fluctuation and low-speed, high-frequency state evolution pattern, while in "high-speed cruising" mode, it displays a highly stable and low-energy-consumption output path. This map can be used for driving behavior retrospective analysis and as an important input for AI models to determine the vehicle's dynamic state. Through 200 hours of real-vehicle sampling and map comparison verification, the accuracy rate for dynamic state recognition reached 94%, significantly improving the modeling capability under complex driving scenarios. Using the completed holographic state perception map of the battery pack, real-time performance profile of the motor, and vehicle dynamics feature map as core inputs, the system enters the virtual simulation stage to build a synchronous digital simulation model of the vehicle. This model is not a traditional static simulation, but a "digital twin" that runs in parallel with the vehicle's real-time data stream. It can pre-assess the energy consumption, heat load, motor efficiency fluctuations, and vehicle dynamic response that each control action may bring before AI control decisions occur. The simulation platform is typically deployed on a high-performance edge computing unit at the vehicle end, using a real-time computing engine (such as a Simulink Real-Time or AUTOSAR platform-compatible model). Before each driving action is initiated, the simulation model runs for a rapid calculation of 300ms to 500ms, outputting performance comparison data under multiple control strategies. Experiments show that after adopting this virtual simulation mechanism, it is possible to detect 8% of potential motor overheating risks in advance and effectively reduce peak energy consumption by 5%, while improving the real-time feedback accuracy of the control strategy. Synchronous digital simulation models are a key bridge for AI control to move from "perception-driven" to "prediction-driven", marking the entry of vehicle start-up and acceleration control into a visualized, quantified, and parallelized intelligent stage.

[0036] In this embodiment, the steps of collecting radar feedback sensing information and real-time monitoring images of the vehicle's front, performing traffic situation prediction and vehicle scene prediction perception, and constructing a vehicle scene prediction perception model include the following: The vehicle-mounted LiDAR performs an all-around environmental scan to obtain radar feedback sensing information. Based on radar feedback sensing information, road slope, road surface friction coefficient and wind speed and direction are calculated to obtain environmental parameter characteristics; Acquire real-time monitoring images of the area in front of the vehicle; perform traffic flow status recognition and traffic light change cycle calculation on the real-time monitoring images in front of the vehicle to obtain traffic status information features; Based on real-time monitoring images of the vehicle in front, deep visual recognition is performed to analyze the dynamic distribution of pedestrians and changes in vehicle displacement, thereby obtaining the dynamic target distribution characteristics of the scene. Traffic situation prediction is performed based on traffic condition information features and dynamic target distribution features in the scene, and a traffic situation prediction map is constructed. Vehicle scene prediction and perception are performed based on traffic situation prediction maps and environmental parameter characteristics, and a vehicle scene prediction and perception model is constructed.

[0037] In this embodiment, an all-around environmental scan is performed using an onboard LiDAR to obtain radar feedback sensing information as the raw data source for environmental modeling. LiDAR provides 360-degree high-resolution point cloud data, which is used to capture the spatial structure and obstacle distribution around the vehicle. In actual deployment, 32-line or 64-line rotating LiDARs are used, mounted on the roof or both sides of the front of the vehicle. The point cloud refresh rate is 10Hz to 20Hz, the horizontal angular resolution is 0.2~0.4 degrees, the vertical angle can reach over 30°, and the detection range exceeds 120 meters. Existing intelligent new energy vehicles have various radar types, which can basically meet the hardware parameter requirements of this technology. In dynamic environments, this point cloud data is continuously recorded in a time series, forming a spatiotemporally consistent structure. The point cloud data is preprocessed to filter out static background and high-frequency noise points, while voxel grid downsampling is used to reduce data dimensionality. The preprocessed point cloud information is then used as the core input to the perception engine, providing basic data support for subsequent environmental parameter calculations. After obtaining radar feedback information, physical environmental parameters are calculated from the point cloud data, specifically including road slope, road surface friction coefficient, and wind speed and direction estimation. The radar's three-dimensional point cloud needs to be projected onto the ground area in front of the vehicle. By fitting the angle between the normal vector of the ground point set and the vehicle coordinate system, the longitudinal and transverse slope information of the road ahead is accurately calculated. In the experiment, the least squares plane fitting method was used to estimate the slope, with an accuracy controlled within ±1.2 degrees. Friction coefficient estimation is based on the road surface point cloud reflectivity and historical weather models for classification. For example, in rainy, slippery road surfaces have reduced reflectivity and more uniform point cloud density changes, identifying this as a low-friction coefficient area. Wind speed and direction estimation is derived based on the radar point cloud change rate and the vehicle's own acceleration correction residual model. Especially at high speeds, the windward point cloud offset and speed difference can be used as quantitative indicators of wind disturbance. Combining the above calculation results, a set of environmental parameter features is formed, updated once per second, for dynamically adjusting the driving control strategy. Real-time image data of the area in front of the vehicle is acquired via an onboard camera to identify traffic flow conditions and analyze traffic light cycles, thereby obtaining traffic status information features. Image acquisition utilizes a high-definition RGB camera with a resolution of 1920×1080 and a frame rate of at least 30fps, mounted in the center of the windshield or front bumper area. After image processing via a convolutional neural network (CNN), traffic flow parameters such as vehicle density, average vehicle speed, and vehicle spacing in the lane ahead can be identified, and the flow can be determined as free-flowing, steady-flowing, or congested. For traffic light recognition, YOLOv5 or EfficientDet structures are used to detect the position of traffic lights, and the signal cycle is calculated using inter-frame brightness variation curves to determine the duration and remaining time of the current light phase. By modeling the sequence of traffic light state changes over the past 30-60 seconds, signal change trends can be predicted in advance, enabling adaptive acceleration prediction in dynamic signal environments.Experimental data show that the traffic light recognition accuracy reaches 97%, and the average signal cycle prediction error is within ±1.3 seconds, providing key rhythm information support for the intelligent start strategy.

[0038] After obtaining traffic state features, deep visual recognition is performed based on camera images to analyze dynamic elements in the scene ahead, especially the spatial distribution and movement trends of pedestrians and vehicles. This relies on the combined application of depth estimation networks and target tracking algorithms, such as using models like Monodepth2 or DPT to construct a depth map of the image ahead, and then combining SORT or DeepSORT for temporal target tracking. It not only identifies target types (pedestrians, bicycles, motorcycles, cars, etc.) but also calculates their trajectories and potential spatial locations within the next 13 seconds, constructing a dynamic target distribution feature map of the scene. In specific scenarios such as pedestrian crossings, complex intersections, or school zones, it automatically improves pedestrian detection sensitivity, predicting pedestrian walking speed, directional intentions, and other behavioral tendencies through keypoint recognition and skeleton tracking technologies. In experimental road tests, it can stably track 812 dynamic targets with a recognition latency controlled within 150 milliseconds and provides over 86% accuracy in target motion prediction, providing a foundation for dynamic obstacle avoidance and energy-saving startup strategies.

[0039] This method integrates traffic state information features with dynamic target distribution features to perform traffic situation prediction analysis and construct a traffic situation prediction map. It presents the dynamic evolution paths, potential conflict points, and state change trends of all key targets in the entire traffic scenario ahead within a short period (15 seconds). A graph neural network (GNN) structure is used, with each target (vehicle, pedestrian, traffic light) as a graph node, and the relative motion relationships between different targets, predicted trajectory overlap, and spatial proximity as graph edge weights. After training, the scenario's situation tensor is output. This map effectively reflects potential risk hotspots, acceleration / deceleration nodes induced by signals, and path accessibility levels. In a set of complex intersection scenarios, the traffic situation prediction map identified 17 potential vehicle conflict trends and provided intervention prompts 0.8-1.5 seconds in advance, providing a safety redundancy window for the vehicle control module.

[0040] Traffic situation prediction maps and environmental parameter features are input into the AI ​​perception and decision-making module to construct a "vehicle scene prediction and perception model," which is the final high-order fusion model before implementing intelligent vehicle start-up control strategies. This model uses a Transformer architecture at its core, incorporating a time attention mechanism to continuously predict the state of the scene in the near future, while embedding the vehicle's current state (such as remaining battery power, motor response capability, and dynamic state) for personalized adjustments. In the model output, each scene configuration (e.g., green light + flat road + low-density traffic ahead + no pedestrians) is mapped to an optimal start-up response strategy, including parameters such as motor output delay, maximum starting power, and battery release curve. Experimental data shows that before and after the control strategy optimization, the average vehicle start-up time was reduced by 0.9 seconds, power consumption decreased by 6.3%, start-up comfort score improved by 15%, and model inference latency was controlled within 300 milliseconds, fully meeting the requirements for real-time deployment.

[0041] In this embodiment, step S4 includes the following steps: The vehicle's comprehensive performance index is calculated using the vehicle's synchronous digital simulation model to generate a dynamic performance baseline. Environmental constraints are evaluated for the vehicle scene prediction and perception model, and the maximum available acceleration for the scene is calculated to generate the maximum available acceleration parameters for the scene. Based on the maximum available acceleration parameters and dynamic performance baseline of the scenario, multi-objective optimization and adjustment are performed to generate the optimal starting torque curve; Based on the optimal starting torque curve, an environmental adaptive power distribution strategy is constructed. The system predicts power demand based on personalized driving profiles and redistributes battery power according to power distribution strategies to build an intelligent matching acceleration control engine.

[0042] In this embodiment, the calculation of the vehicle's comprehensive performance index for the synchronous digital simulation model relies on a previously constructed multi-source data-driven model that includes battery, motor, chassis, and dynamic characteristic maps. This model integrates factors such as maximum power output, motor response characteristics, vehicle mass, and aerodynamic parameters to simulate the vehicle's dynamic response under multiple typical driving conditions. Through structured analysis of the simulation results under different scenarios—for example, quantifying multiple parameters such as acceleration curves, vehicle start-up time, torque output curves, yaw rate of change, and energy consumption—and then using a normalization method to unify these heterogeneous data into a dimensionless comparison benchmark, the weights of each performance index are determined using the analytic hierarchy process (AHP), ultimately forming the vehicle's comprehensive performance index under different operating states. This comprehensive index not only reflects the vehicle's instantaneous performance capabilities but also, through time series analysis, depicts the vehicle's performance evolution curve during the complete acceleration process. By applying curve envelope processing to the performance indices under all conditions, a dynamic performance benchmark covering different driving scenarios can be generated. This baseline represents the theoretical maximum performance boundary that the vehicle can achieve under different states under the current hardware capabilities and control strategies. It serves as a key reference baseline for subsequent multi-objective optimization of the control strategy. Environmental constraint assessment is performed on the vehicle scene prediction perception model, and based on this, the maximum usable acceleration capability of the vehicle in a specific scenario is calculated. This step uses the previously constructed multimodal perception system to provide environmental state input, including traffic signal information identified by cameras, terrain and obstacle distribution acquired by radar, vehicle attitude information provided by the IMU, and wind speed, direction, and slope estimated by sensors. The environmental perception module comprehensively models information such as road surface conditions, adhesion coefficient, and slope angle, and combines this with dynamic target features such as real-time traffic flow density, preceding vehicle behavior, and pedestrian activity areas to constrain and calculate the safe acceleration that the vehicle can achieve in the current environment. For example, on slippery urban curved roads, based on an empirical regression model combined with real-time road surface reflection information, it is determined that the current road adhesion coefficient is insufficient to support standard start-up acceleration, thus dynamically adjusting the maximum usable acceleration. During this process, the vehicle's own power capability remains in the model as a hardware upper limit, but an "environmentally tolerable maximum acceleration value" is generated based on the environmental modeling results. This value is continuously updated over time and serves as a dynamic lower bound encapsulation of the performance baseline. Based on the integration of traffic flow behavior, it can also predict interactive risks during acceleration, such as potential braking by the vehicle ahead or the possibility of a vehicle entering the lane from the side. This allows for further safety redundancy adjustments to the acceleration, ensuring stable, controllable, and abrupt vehicle behavior, and providing precise dynamic boundaries for intelligent control strategies.

[0043] Based on the aforementioned dynamic performance baseline and the maximum available acceleration parameters for the scenario, the core control optimization stage will commence, namely, multi-objective optimization adjustment to generate the optimal starting torque curve. This process aims to improve power performance, ensure safety, and balance comfort and energy efficiency, using the motor output torque variation sequence as the optimization variable for a global search. First, an optimization time window (e.g., the 0-5 second start-up phase) is set, and this phase is discretized into multiple time slices, each corresponding to a motor output torque value, forming a time-series torque control curve. To achieve the optimization search, a multi-objective genetic algorithm based on an evolutionary mechanism is used, trained with a large number of simulation samples to obtain a torque output scheme with high adaptability. Each candidate curve undergoes vehicle acceleration simulation in the simulation platform, calculating its performance score across five dimensions: maximum acceleration proximity, acceleration smoothness, overall energy consumption, tire slippage probability, and vehicle stability. Curves that do not meet the scenario's maximum available acceleration limit are directly eliminated. After extensive iterations and selection, an optimal starting torque curve balancing various objective constraints is finally generated. This curve may provide a high response torque in the initial stage of start-up to achieve rapid departure from a standstill, then gradually increase to a stable plateau range to ensure traction continuity, while automatically converging near high loads to avoid slippage risks. This curve not only serves as a reference input for the motor controller but can also be adjusted in real time according to changes in environmental constraints through parameterization, forming a dynamic curve template with adaptive characteristics. The various levels of vehicle power controllers are coordinated to implement an environmentally adaptive power distribution strategy. Specifically, this involves real-time analysis of various factors such as the drive motor efficiency range, current battery temperature (e.g., suppressing high torque requests when below 0°C), transmission efficiency between the motor and wheel ends, and the activation status of energy recovery mode. In single-motor or dual-motor configurations, the front-to-rear drive ratio must also be adjusted in real time based on adhesion conditions and tire load to avoid power slippage and energy waste. Furthermore, in complex environments such as slopes or icy roads, the starting torque upper limit will be adaptively reduced, acceleration and climbing response time will be extended, and the anti-slip control level will be automatically increased to improve stability. This strategy not only ensures the intelligent starting capability of electric vehicles in different climates, terrains and traffic environments, but also has the ability to adaptively adjust to long-term operating conditions such as battery aging and drive performance degradation, achieving closed-loop optimization of intelligent start-up and acceleration control under all operating conditions.

[0044] In this embodiment, the process of predicting power demand based on personalized driving profiles and reallocating battery power according to the power distribution strategy to build an intelligent matching acceleration control engine includes the following steps: Power demand prediction is performed on personalized driving profiles to generate target power demand prediction values ​​for driver's licenses; Dynamic acceleration trajectory planning is performed based on the predicted target power demand value of the driver's license to generate an acceleration trajectory that matches the power demand. Based on the acceleration trajectory, the power distribution strategy is used to redistribute battery pack power and construct an intelligent matching acceleration control engine. Real-time acceleration control is performed based on an intelligent matching acceleration control engine, and acceleration intervals are calculated. During acceleration intervals, braking energy recovery and intelligent battery balancing management are performed, and intelligent vehicle start-up and acceleration control operations are executed based on the intelligent start-up mode selection strategy.

[0045] In this embodiment, a personalized driving profile of the driver is modeled to accurately predict their power demand in different scenarios. This driving profile is based on the collection and labeling of long-term driving behavior data, including accelerator pedal depressor depth and duration, throttle response frequency, start-up timing, acceleration / deceleration style (rapid / smooth), driving time distribution, and road type preference. During the training phase, supervised learning models (such as XGBoost and Random Forest) are used to extract the correlation mapping between different driver feature dimensions and power demand. For example, after collecting data on a driver's driving behavior over approximately 800km in 3 months, it was found that in urban roads, the driver maintained an acceleration of 0.8-1.3 m / s² during the start-up phase for 90% of the time, and preferred 2.2 m / s² on elevated roads. Combining environmental parameters, real-time driving context, and personal style factors (such as fuel efficiency preference and high responsiveness preference), the predicted value of the driver's target power demand in the current scenario is output and dynamically adjusted over time to form an "individualized power expectation curve," which directly guides subsequent power trajectory planning.

[0046] Next, based on the predicted target power demand, the dynamic acceleration trajectory planning stage begins. This process aims to transform the power demand into a dynamic trajectory of time-velocity-acceleration, guiding motor output control and torque response distribution. First, the vehicle dynamics simulation module is invoked to determine the feasible acceleration range under current physical conditions based on parameters such as vehicle weight, gradient, wind resistance, and tire adhesion. Then, considering the driver's desired power level, a continuous and smooth acceleration trajectory is generated within a 3-5 second time window using deep sequence modeling (such as Bi-LSTM). The planning process also incorporates a "driving comfort" constraint to control the rate of change of acceleration (jerk) within an acceptable range, ensuring that power output does not cause discomfort to occupants. The traffic environment ahead is also considered; for example, if the vehicle ahead is predicted to brake or turn within 2 seconds, the planned trajectory will automatically suppress power output and lengthen the acceleration curve period. The final generated acceleration trajectory not only matches the driver's expectations but also possesses environmental adaptability and energy efficiency optimization characteristics, serving as the most direct basis for engine control.

[0047] The generated acceleration trajectory will serve as the core for reconfiguring the vehicle's power distribution strategy, specifically in the real-time allocation and coordinated management of battery pack power. The key at this stage is translating the target acceleration into a practically feasible battery power output strategy, avoiding over-discharge, thermal runaway, or individual cell voltage drift. The acceleration trajectory will be subdivided into several power output points. Combining state parameters such as the battery pack's current SOC (State of Charge), individual cell temperature distribution, and internal resistance trends, the optimal battery output power at each time point will be determined using constrained optimization algorithms (such as a QP-based power scheduling model). Simultaneously, in a multi-mode drive architecture, the control system must also determine the output distribution ratio between the front and rear axle motors. For example, in low-traction environments, the power ratio will be tilted towards the front axle to improve traction stability. The entire power output scheduling is completed by an intelligent acceleration control engine, which integrates power prediction, load balancing, thermal safety regulation, and voltage protection mechanisms, forming a highly responsive and adaptable power distribution execution platform.

[0048] The intelligent acceleration control engine enters the real-time acceleration control phase, responsible for rapidly converting planned power output and torque commands into motor control commands to ensure the vehicle executes stably along the acceleration trajectory. It continuously monitors actual vehicle speed, torque response, wheel slippage, and drive efficiency through high-frequency sampling (typically at the 10ms level) and compares these values ​​with target values ​​in real time. If a deviation exceeds a set threshold (e.g., torque deviation exceeding ±10Nm or acceleration deviation exceeding ±0.15m / s²), the control commands are immediately corrected to achieve precise closed-loop power control. Simultaneously, it records the transition intervals between each power output, i.e., "acceleration intervals." These intervals are often critical periods with discontinuous power output but significant energy management potential. These periods are particularly noticeable in urban congestion, slow-moving traffic, and frequent stop-and-go scenarios, occurring an average of 2-3 times per minute. Effectively utilizing these intervals helps improve the overall vehicle energy efficiency.

[0049] During acceleration intervals, the system automatically switches to energy recovery and battery balancing management modes. For energy recovery, the vehicle converts kinetic energy into electrical energy via the brake motor, which flows back to the battery. This energy is precisely distributed based on the current State of Charge (SOC) difference of each cell to achieve active battery balancing. The control system dynamically adjusts the charging rate of each cell, prioritizing cells with lower SOC or lower temperature, thereby reducing the rate of degradation across the entire battery pack. This process relies on a high-precision individual cell state monitoring module, achieving an average voltage difference compensation effect of 1-2% every 10 minutes. Simultaneously, based on an intelligent start-up mode selection strategy, considering driver operating habits, environmental scenario labels, and traffic condition predictions, the system determines the power release rhythm for the next start (e.g., ease mode, energy-saving mode, response-priority mode), and adjusts control engine parameters in real time. This closed-loop process fully realizes the entire process from understanding driver behavior, predicting target needs, generating acceleration trajectories, scheduling power resources, to energy recovery and reuse. Through this personalized, environmentally adaptable, and fast-responding control, electric vehicles can achieve higher energy efficiency, better driving comfort, and stronger safety during the start-up and acceleration phases, providing solid technical support for intelligent decision-making on individual vehicle behaviors in intelligent transportation.

[0050] In this embodiment, an AI-based intelligent start-up and acceleration control system for electric vehicles is provided, used to execute the AI-based intelligent start-up and acceleration control method for electric vehicles as described above, including: The intelligent start module is used to obtain the driver's driving log, optimize the intelligent vehicle after startup, and build an intelligent start mode selection strategy. The driving profile module is used to acquire driver physiological state data, analyze state changes and evolution, and build a personalized driving profile. The synchronous simulation module is used to collect battery pack monitoring parameters, perform holographic state perception and vehicle virtual simulation, and build a synchronous digital simulation model of the vehicle. The adaptive power distribution module is used to predict power demand based on the vehicle's synchronous digital simulation model and personalized driving profile, and to redistribute battery pack power based on the intelligent start-up mode selection strategy.

[0051] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0052] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement it. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein are implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. An AI-based intelligent start-up and acceleration control method for electric vehicles, characterized in that, Includes the following steps: Step S1: Obtain the driver's driving log, optimize the intelligent vehicle after startup, and build an intelligent startup mode selection strategy; Step S2: Obtain driver's physiological state data, analyze state changes and evolution, and construct a personalized driving profile; Step S3: Collect battery pack monitoring parameters, perform holographic status perception and vehicle virtual simulation, and construct a synchronous digital simulation model of the vehicle; Step S4: Based on the vehicle's synchronous digital simulation model and personalized driving profile, predict power demand and redistribute battery pack power based on the intelligent start-up mode selection strategy. Specifically, step S3 involves the following steps: Collect multi-dimensional monitoring parameters of the battery pack; Temperature distribution analysis is performed on the multi-dimensional monitoring parameters, voltage difference calculation and internal resistance change identification of individual cells are performed, and holographic state perception is performed to construct a holographic state perception map of the battery pack. Data on the motor rotor position, stator temperature, and magnetic field strength are collected, and the motor system performance is analyzed to obtain a real-time profile of the motor system performance. The vehicle attitude, tire pressure, and suspension load are obtained, and the chassis system state information is fitted. Perform vehicle dynamics evolution analysis on chassis system state information to construct a vehicle dynamics feature map; Virtual vehicle simulation is performed on the vehicle dynamics feature map, real-time performance profile of the motor system, and holographic state perception map of the battery pack to construct a synchronous digital simulation model of the vehicle. Collect radar feedback sensor information and real-time monitoring images in front of vehicles to perform traffic situation prediction and vehicle scene prediction perception, and build a vehicle scene prediction perception model. The specific steps of step S4 are as follows: The vehicle's comprehensive performance index is calculated using the vehicle's synchronous digital simulation model to generate a dynamic performance baseline. Environmental constraints are evaluated on the vehicle scene prediction and perception model, and the maximum available acceleration for the scene is calculated to generate the maximum available acceleration parameters for the scene. Based on the maximum available acceleration parameters and dynamic performance baseline of the scenario, multi-objective optimization and adjustment are performed to generate the optimal starting torque curve; Based on the optimal starting torque curve, environmental adaptive power distribution is performed to construct a power distribution strategy; The system predicts power demand based on personalized driving profiles and redistributes battery power according to power distribution strategies to build an intelligent matching acceleration control engine. The specific steps for predicting power demand based on personalized driving profiles and redistributing battery power according to the power distribution strategy to build an intelligent matching acceleration control engine are as follows: Power demand prediction is performed on personalized driving profiles to generate target power demand prediction values ​​for driver's licenses; Dynamic acceleration trajectory planning is performed based on the predicted target power demand value of the driver's license to generate an acceleration trajectory that matches the power demand. Based on the acceleration trajectory, the power distribution strategy is used to redistribute battery pack power and construct an intelligent matching acceleration control engine. Real-time acceleration control is performed based on an intelligent matching acceleration control engine, and acceleration intervals are calculated. During acceleration intervals, braking energy recovery and intelligent battery balancing management are performed, and intelligent vehicle start-up and acceleration control operations are executed based on the intelligent start-up mode selection strategy.

2. The AI-based intelligent start-up and acceleration control method for electric vehicles according to claim 1, characterized in that, The specific steps of step S1 are as follows: Obtain the vehicle owner's driving log; Define a time-series behavior period and extract the vehicle startup behavior from the owner's driving log to obtain the vehicle startup behavior data stream within the period; The driving behavior is analyzed sequentially after each vehicle startup from the vehicle startup behavior data stream to generate driving behavior features; Extract the vehicle startup timestamp from the vehicle startup behavior data stream; Perform periodic time series fitting on the timestamps to construct a periodic timestamp marker map; Intelligent vehicle startup optimization is performed based on driving behavior characteristics and periodic timestamp markers to construct an intelligent startup mode selection strategy.

3. The AI-based intelligent start-up and acceleration control method for electric vehicles according to claim 2, characterized in that, The specific steps for optimizing intelligent vehicle startup based on driving behavior characteristics and periodic timestamp markers, and constructing an intelligent startup mode selection strategy, are as follows: Identify the real-time entry time of the vehicle owner's vehicle; The optimal similarity matching calculation is performed on the real-time vehicle entry time of the car owner based on the periodic timestamp marker map to obtain the most similar start timestamp in history. Based on the most similar historical startup timestamp, analyze the car owner's usage scenario to obtain the current car owner's usage scenario; The travel time delay is calculated based on the most similar start timestamp in history to obtain the travel delay parameters for this trip; The vehicle driving mode is intelligently selected based on driving behavior characteristics and the delay parameters of this trip to obtain a pre-loaded driving mode; Intelligent vehicle startup optimization is performed based on pre-loaded driving modes, and an intelligent startup mode selection strategy is constructed.

4. The AI-based intelligent start-up and acceleration control method for electric vehicles according to claim 1, characterized in that, The specific steps of step S2 are as follows: Based on the real-time acquisition of driver's heart rate variability, eye movement trajectory, blink frequency and closure degree by vehicle-mounted multimodal sensor array, driver physiological state data is constructed. Multi-time point state change evolution analysis was performed on the multi-dimensional feature vector of the driver's physiological state data to obtain multi-time point state evolution features; The potential driving intentions and the urgency of driving scenarios are mined from the multi-time point state evolution characteristics to generate potential driving intentions and the urgency of driving scenarios. Personalized real-time profiles are built based on potential driving intentions and the urgency of driving scenarios to create personalized driving profiles.

5. The AI-based intelligent start-up and acceleration control method for electric vehicles according to claim 1, characterized in that, The specific steps for collecting radar feedback sensor information and real-time monitoring images of the vehicle's front, performing traffic situation prediction and vehicle scene prediction perception, and constructing a vehicle scene prediction perception model are as follows: The vehicle-mounted LiDAR performs an all-around environmental scan to obtain radar feedback sensing information. Based on radar feedback sensor information, road slope, road surface friction coefficient and wind speed and direction are calculated to obtain environmental parameter characteristics; Acquire real-time monitoring images of the area in front of the vehicle; Traffic flow status identification and traffic light change cycle calculation are performed on real-time monitoring images of the vehicle ahead to obtain traffic status information features; Based on real-time monitoring images of the front of the vehicle, deep visual recognition is performed to analyze the dynamic distribution of pedestrians and changes in vehicle displacement, thereby obtaining the dynamic target distribution characteristics of the scene. Traffic situation prediction is performed based on traffic condition information features and dynamic target distribution features in the scene, and a traffic situation prediction map is constructed. Vehicle scene prediction and perception are performed based on traffic situation prediction maps and environmental parameter characteristics, and a vehicle scene prediction and perception model is constructed.

6. An AI-based intelligent start-up and acceleration control system for electric vehicles, characterized in that, The method for executing the AI-based intelligent start-up and acceleration control method for electric vehicles as described in claim 1 includes: The intelligent start module is used to obtain the driver's driving log, optimize the intelligent vehicle after startup, and build an intelligent start mode selection strategy. The driving profile module is used to acquire driver physiological state data, analyze state changes and evolution, and build a personalized driving profile. The synchronous simulation module is used to collect battery pack monitoring parameters, perform holographic state perception and vehicle virtual simulation, and build a synchronous digital simulation model of the vehicle. The adaptive power distribution module is used to predict power demand based on the vehicle's synchronous digital simulation model and personalized driving profile, and to redistribute battery pack power based on the intelligent start-up mode selection strategy.

Images