New energy automobile safe driving control method and system

Abstract

The invention relates to the technical field of driving control, in particular to a new energy automobile safe driving control method and system. The method comprises the following steps that a historical urgent danger avoiding data set and a corresponding user misoperation braking state are obtained through a new energy automobile control center, sample undersampling processing is conducted firstly, and an urgent danger avoiding balance sample is obtained; thirdly, analyzing user misoperation braking abnormity by using a balance sample, and performing disordered vector intensity calculation under a time dimension based on braking abnormity data to generate disordered vector segmentation clustering data; then, braking time sequence intervention correction sensing is carried out according to the clustering data, optimized braking intervention correction data are generated, finally, the optimized data are transmitted to a new energy automobile control terminal, and safe driving control is executed. According to the invention, the driving control technology is optimized, so that the driving control technology is more perfect.

Description

Technical Field

[0001] This invention relates to the field of driving control technology, and in particular to a method and system for safe driving control of new energy vehicles. Background Technology

[0002] New energy vehicles exhibit greater complexity in terms of power response, braking systems, electronic control systems, and intelligent driving assistance systems. This makes driving operations in emergency situations more susceptible to driver error. For example, in emergency avoidance scenarios, drivers may over-brake, under-brake, or make incorrect maneuvers due to sudden situations. This not only reduces avoidance efficiency but may also trigger secondary accidents. Therefore, developing a safe driving method capable of quickly detecting driver error, analyzing braking anomalies, and optimizing control through intelligent intervention in emergency situations has become an important direction in new energy vehicle safety technology research. This method not only requires modeling driver behavior based on historical driving data and emergency avoidance scenarios but also utilizes data analysis and clustering algorithms to identify disordered braking patterns and dynamically evaluate braking behavior over time to generate precise correction strategies. However, traditional new energy vehicle safe driving control methods suffer from inaccurate analysis of the intensity of braking anomalies caused by user error, resulting in large braking intervention errors. Summary of the Invention

[0003] Therefore, it is necessary to provide a method and system for safe driving control of new energy vehicles to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above objectives, a method for safe driving control of new energy vehicles is provided, the method comprising the following steps: Step S1: Obtain historical emergency avoidance dataset and corresponding user misoperation braking status through the new energy vehicle control center; perform sample undersampling processing on the historical emergency avoidance dataset to obtain emergency avoidance balanced samples; Step S2: Based on the emergency avoidance balance sample, perform user misoperation braking anomaly analysis on the avoidance scenario to obtain braking anomaly data; based on the braking anomaly data, perform braking disorder vector intensity calculation in the time dimension to obtain disorder vector segment clustering data. Step S3: Perform braking timing intervention correction sensing based on disordered vector segmented clustering data to generate braking intervention correction optimization data; send the braking intervention correction optimization data to the new energy vehicle control terminal to execute new energy vehicle safe driving control.

[0005] Preferably, the present invention also provides a new energy vehicle safety driving control system for executing the new energy vehicle safety driving control method described above, the new energy vehicle safety driving control system comprising: The data sampling module is used to obtain historical emergency avoidance datasets and corresponding user misoperation braking states through the new energy vehicle control center; and to perform sample undersampling processing on the historical emergency avoidance datasets to obtain emergency avoidance balance samples. The braking anomaly analysis module is used to analyze user misoperation braking anomalies in the emergency avoidance balance sample to obtain braking anomaly data; and to perform time-dimensional braking disorder vector intensity calculation based on the braking anomaly data to obtain disorder vector segmented clustering data. The intervention correction perception module is used to perform braking timing intervention correction perception based on disordered vector segmented clustering data to generate braking intervention correction optimization data; the braking intervention correction optimization data is sent to the new energy vehicle control terminal to execute new energy vehicle safe driving control.

[0006] The beneficial effects of this invention are that by acquiring historical emergency avoidance datasets and corresponding user-incorrect braking states through the new energy vehicle control center, it is possible to effectively collect data on erroneous actions that occur during actual vehicle operation, providing real and reliable samples for subsequent analysis. By performing undersampling on the dataset, the imbalance between positive and negative samples can be balanced, ensuring that model training is not affected by bias. The generated balanced emergency avoidance samples can better represent the actual performance of drivers in emergency situations, providing basic data for subsequent analysis of abnormal braking actions, thus laying a solid foundation for optimizing safe driving control strategies. Based on the balanced emergency avoidance samples, by analyzing the braking states and identifying abnormal braking data, driver errors and abnormal behaviors in emergency scenarios can be accurately captured. This process not only reveals the driver's reaction characteristics in the face of emergency situations but also provides detailed abnormal data support for subsequent calculations of disordered braking vector strength. Calculations based on this data over time can accurately assess the degree of disorder in braking behavior at different time points, further generating disordered vector segmentation clustering data, providing rich information support for subsequent optimized control, and ensuring that driver errors are effectively identified and handled. This invention utilizes unordered vector segmented clustering data for braking timing intervention and correction perception, enabling real-time optimization of vehicle braking control based on specific driver erroneous actions. By accurately sensing the temporal changes in braking behavior, this process automatically identifies critical moments requiring intervention and correction, adjusting the braking system's operation at these times to prevent potential accidents. The generated braking intervention and correction optimization data effectively guides the vehicle control terminal to execute safer and more precise driving controls, helping drivers adjust in a timely manner for a safer driving experience. The implementation of this technology effectively improves the active safety performance of new energy vehicles, reduces accident risks, and enhances overall driving safety. Therefore, this invention optimizes a traditional new energy vehicle safe driving control method, addressing the problem of inaccurate analysis of braking anomalies caused by user erroneous actions, leading to large braking intervention errors. It improves the accuracy of analyzing braking anomalies caused by user erroneous actions and reduces the errors caused by braking intervention. Attached Figure Description

[0007] Figure 1 A flowchart illustrating the steps of a safe driving control method for new energy vehicles; Figure 2 for Figure 1 A detailed flowchart illustrating the implementation steps of step S2. Figure 3 for Figure 1 A detailed flowchart illustrating the implementation steps of step S3. Detailed Implementation

[0008] Please see Figures 1 to 3 A method for safe driving control of new energy vehicles, the method comprising the following steps: Step S1: Obtain historical emergency avoidance dataset and corresponding user misoperation braking status through the new energy vehicle control center; perform sample undersampling processing on the historical emergency avoidance dataset to obtain emergency avoidance balanced samples; Step S2: Based on the emergency avoidance balance sample, perform user misoperation braking anomaly analysis on the avoidance scenario to obtain braking anomaly data; based on the braking anomaly data, perform braking disorder vector intensity calculation in the time dimension to obtain disorder vector segment clustering data. Step S3: Perform braking timing intervention correction sensing based on disordered vector segmented clustering data to generate braking intervention correction optimization data; send the braking intervention correction optimization data to the new energy vehicle control terminal to execute new energy vehicle safe driving control.

[0009] In this embodiment of the invention, reference Figure 1 The above is a flowchart illustrating the steps of a new energy vehicle safe driving control method according to the present invention. In this example, the new energy vehicle safe driving control method includes the following steps: Step S1: Obtain historical emergency avoidance dataset and corresponding user misoperation braking status through the new energy vehicle control center; perform sample undersampling processing on the historical emergency avoidance dataset to obtain emergency avoidance balanced samples; In this embodiment of the invention, the new energy vehicle control center extracts the raw log dataset of emergency avoidance events stored in the vehicle data recording module over the past 36 months. This dataset contains timestamps for each event, accurate to the millisecond level, with a sampling frequency of 100Hz. The recorded content includes vehicle longitudinal acceleration, lateral acceleration, steering wheel angle, accelerator pedal opening, brake pedal pressure, vehicle speed, GPS latitude and longitude, and the activation status of the Electronic Stability Program (ESP). Simultaneously, user misoperation braking status labels associated with each emergency avoidance event are extracted. These labels are manually verified and labeled within 72 hours of the event, based on cross-validation of video surveillance playback and sensor data. Undersampling processing is performed on the historical emergency avoidance dataset. First, the positive and negative class distribution ratios in all samples are statistically analyzed. It is found that the proportion of abnormal misoperation braking states accounts for only 8.7% of the total sample size, indicating a severely imbalanced dataset. The Tomek Links algorithm is used to identify and remove overlapping samples, and then Edited Nearest Neighbors is used... The ENN method removes noise and ambiguous points, ultimately retaining 23% of the original negative samples, making the positive-to-negative sample ratio 1:1.05, forming an emergency avoidance balance sample. The total number of samples is 4728, each sample contains a time series of 5 seconds in length, with a total of 500 sampling points. Each sampling point contains a 12-dimensional feature vector, including physical parameters such as acceleration vector components, direction angle change rate, and pedal opening gradient.

[0010] Step S2: Based on the emergency avoidance balance sample, perform user misoperation braking anomaly analysis on the avoidance scenario to obtain braking anomaly data; based on the braking anomaly data, perform braking disorder vector intensity calculation in the time dimension to obtain disorder vector segment clustering data. In this embodiment of the invention, three-dimensional scene reconstruction data is constructed based on emergency avoidance balance samples. Vehicle dynamic parameters (including vehicle mass 1560kg, center of gravity height 0.52m, wheelbase 2.7m, and tire lateral stiffness 85000N / rad) are used, combined with GPS trajectory coordinates and IMU attitude angle data, to reconstruct the spatial topology of the avoidance scene corresponding to each sample. Simultaneously, the erroneous braking state is mapped to the time slice position in the scene, generating an avoidance scene erroneous operation mapping data matrix. This matrix has dimensions of 4728 rows by 500 columns, with each column corresponding to 10ms. The operation status code value within the window is used to perform anomaly analysis on the mapped data. The throttle acceleration status is extracted. If the longitudinal acceleration value is consistently greater than 12 m / s² for 150 consecutive sampling points (1.5 s), and the throttle pedal opening jumps from 30% to over 90% within 100 ms, it is marked as an abnormal acceleration state due to throttle misapplication. The steering operation status is also extracted. If the cumulative change in steering wheel angle exceeds 90 degrees within 1000 ms, and there are 3 or more steering switching actions with an angle difference greater than 45 degrees within any 333 ms window, it is marked as an abnormal steering state due to steering misoperation. Two types of abnormal states are combined to generate braking anomaly data. This data structure is a 500-bit binary mask sequence, with each bit corresponding to an anomaly trigger flag bit for a 10ms time window. Then, unordered vector intensity calculations are performed on the braking anomaly data in the time dimension. First, the acceleration change gradient within the abnormal time period is extracted. The difference between the acceleration increment within each 200ms window and the previous window is calculated to form a speed increase ratio difference sequence. Simultaneously, the maximum sway amplitude of the steering wheel angle within the same window is calculated to form a steering angle sudden sway amplitude sequence. The speed increase ratio difference sequence and the steering angle sudden sway amplitude sequence are then input into the yaw rate... The impact vector intensity resolver performs nonlinear regression fitting and smooths the data using cubic spline interpolation to output a lateral impact vector regression intensity sequence. This sequence has a sampling rate of 5Hz and a total of 25 data points. Each data point contains a two-dimensional vector composed of longitudinal impact component and lateral sway component. The K-means clustering algorithm is applied to the regression intensity sequence, with the number of clusters K set to 5. The clustering features are three dimensions: time, location, intensity amplitude change slope, etc. This generates unordered vector segmented clustering data, and the output consists of 5 clusters. Each cluster contains the intensity vector at the center of the time start and end points and the standard deviation of the samples within the cluster.

[0011] Step S3: Perform braking timing intervention correction sensing based on disordered vector segmented clustering data to generate braking intervention correction optimization data; send the braking intervention correction optimization data to the new energy vehicle control terminal to execute new energy vehicle safe driving control.

[0012] In this embodiment of the invention, after receiving unordered vector segmented clustering data, a one-dimensional convolution operation is first performed on the time series segment of each cluster. The convolution kernel length is 5, the stride is 1, the weights are initialized to a Gaussian distribution with a mean of 0 and a variance of 0.1, and the bias term is fixed at 0. After passing through the ReLU activation function, unordered vector segmented convolution data is output. This data has a dimension of 25 rows by 3 columns, representing the time index, the longitudinal intensity after convolution, and the lateral intensity after convolution, respectively. The segmented directional offset angle is extracted from the convolution data. This angle is obtained by converting the lateral intensity value through the arctangent function, and the unit is radians. Simultaneously, the vehicle speed stall value is extracted. This value is obtained by subtracting the longitudinal intensity value from the original vehicle speed signal, and the unit is m / s. For each group of directional offset angles and The vehicle speed stall value is quantified by the vehicle body offset angle and mapped to a preset offset level table using a lookup table. This table contains 9 levels, from -4 to +4, each corresponding to a 0.15 radian interval. Based on the quantification result, the braking torque demand mapping table is consulted. This table is based on a vehicle load of 1560kg, a tire friction coefficient of 0.85, and a brake disc radius of 0.16m. It pre-stores ideal braking torque values ​​in N·m for different offset levels to generate braking torque demand data. Smoothing control optimization is performed on the braking torque demand data. First, the intervention correction angle is based on the directional offset of the current time period. This angle is equal to the original offset angle multiplied by the damping coefficient of 0.7 to obtain the corrected angle value. Simultaneously, the intervention correction speed is calculated. This speed is equal to the original… The initial vehicle speed is reduced by the longitudinal strength value multiplied by a compensation factor of 0.3 (in m / s). Using the corrected angle and corrected speed, a simulation calculation for vehicle intervention to correct braking yaw is performed. The fourth-order Runge-Kutta method is used to integrate the vehicle's two-degree-of-freedom dynamic equations, with a time step of 0.01 s and an integration interval equal to the current time period. The output yaw data includes two components: vehicle yaw rate and lateral acceleration. Multiple linear regression is performed on the yaw data, with the corrected angle and corrected speed as independent variables and yaw rate and lateral acceleration as dependent variables. The regression coefficient matrix has a dimension of 2 rows by 2 columns. The yaw regression data is obtained by solving using the least squares method. Based on the regression data, progressive control is applied to the corrected angle and corrected speed. The control law is the current value, equal to the previous value plus a step size of 0.02 radians, or 0.5 m / s multiplied by the sign function. It outputs the offset asymptotic correction angle and the asymptotic correction vehicle speed. These are substituted into the braking torque smoothing controller, which uses a PID structure with a proportional gain of 12, an integral time constant of 0.8 s, and a derivative time constant of 0.05 s. The controller outputs braking force smoothing control data, which is a time series with a sampling rate of 10 Hz and a length of 50 points, each point representing a scalar torque value in N·m. This braking force smoothing control data is used as braking intervention correction data and input into the iterative learning module. The learning rate is set to 0.1, and the number of iterations is fixed at 5 rounds. Each round of updates uses the gradient descent method with a loss function of mean square error and a threshold of 0.0.01 N·m² is used to generate braking intervention correction optimization data. This data is in JSON format, containing an array of timestamps and an array of torque commands. It is transmitted to the new energy vehicle control terminal via the CAN bus protocol at a baud rate of 500 kbps. Upon receiving the data, the terminal immediately writes it to the brake actuator control register address 0x3A2B, triggering a hardware interrupt to start the hydraulic regulating valve and execute the corresponding braking force distribution to achieve safe driving control.

[0013] Step S1 includes the following steps: Step S11: Obtain historical emergency avoidance dataset and corresponding user misoperation braking status through the new energy vehicle control center; Step S12: Randomly select samples from the historical emergency evacuation dataset to obtain random emergency evacuation samples; Step S13: Embed time tags into the emergency evacuation random samples to obtain evacuation time tag samples; Step S14: Perform sample undersampling on the risk avoidance time tag samples to obtain emergency risk avoidance balance samples.

[0014] In this embodiment of the invention, the vehicle data storage unit is accessed through the new energy vehicle control center. This unit adopts the eMMC 5.1 standard, has a capacity of 32GB, and uses the FAT32 data partitioning format. The file system stores driving event logs from the past 36 months in a circular overwrite manner. The emergency avoidance dataset is derived from the ESP electronic stability program trigger event logs. Each record contains a timestamp accurate to the millisecond level, with a fixed sampling frequency of 100Hz. Record fields include longitudinal acceleration, lateral acceleration, yaw rate, steering wheel angle, accelerator pedal position percentage, brake pedal pressure value in kPa, vehicle speed in m / s, and GPS latitude and longitude coordinates with an accuracy of 0.00001 degrees. Simultaneously, the corresponding user misoperation braking state is extracted from the driver behavior analysis module. This state is determined by a manual annotation process completed within 72 hours after the event. The annotation is based on frame-by-frame interpretation after synchronizing and aligning the in-vehicle DVR video stream with IMU sensor data. The annotation results are stored in binary encoding form, with 1 indicating confirmed misoperation and 0 indicating normal operation. The annotation process follows the SAE J2945 / 1 standard, and the annotators hold ISO certifications. The 39001 certification shows that the consistency of the annotation is greater than 0.89 according to the Kappa coefficient test. The original dataset contains a total of 12,456 records, each with a duration of 5 seconds, and a total of 500 sampling points. Each sampling point contains 14 physical quantity fields.

[0015] Random sample selection was performed on the historical emergency evacuation dataset using the Mersenne Twister pseudo-random number generator with a fixed seed of 0x7A3F. The random number sequence generator was initialized with an output range limited to 1 to 12456. Integer index values ​​were not repeatedly extracted, and the number of records extracted was set to 40% of the original dataset, i.e., 4982 records. The extraction process was carried out in batches of 500 records each to avoid memory overflow. Each selected record retained its complete time series structure of 500 sampling points and all 14-dimensional feature fields, forming a random sample of emergency evacuation. This sample maintained the temporal continuity and spatial coordinate integrity of the original data without any interpolation or truncation operations. The sample distribution was tested for Shapiro-Wilk normality with p-values ​​greater than 0.05, indicating that the randomness conformed to the uniform distribution assumption. The sample index list was exported in CSV format, with the first column being the original record ID, the second column being the extraction timestamp, and the third column being the last six digits of the vehicle VIN code for traceability verification.

[0016] For random samples of emergency evacuation, a timestamp embedding operation is performed. First, the timestamp field embedded in each sample is parsed, in the format YYYY-MM-DD. The file HH:MM:SS.mmm extracts seven numerical segments: year, month, day, hour, minute, second, and millisecond. These segments are converted to integer variables to construct a seven-dimensional time feature vector. This time feature vector is then appended to each of the 500 sampling points of each sample, forming an extended data structure. The new structure has 500 rows and 21 columns, with seven new columns for time labels: year (2020-2024), month (1-12), day (1-31), hour (0-23), minute (0-59), second (0-59), and millisecond (0-999). To ensure time continuity, a sliding window accumulation correction is applied to the millisecond field. The rule is that if the current sampling point's millisecond value is less than the previous one, it is automatically incremented by 1000 and borrowed into the second field. If the second field overflows, it is incremented into the minute field, and so on. The final output is a risk avoidance time label sample, with a total of 4982 entries. Each sample data block is 21000 bytes in size and stored in HDF5 format.

[0017] Undersampling was performed on the time-labeled samples for emergency avoidance. First, the distribution of user erroneous braking status labels was statistically analyzed. The results showed that there were 433 samples with label 1 (abnormal operation) and 4549 samples with label 0 (normal operation), with a class ratio of 1:10.5. The Tomek Links algorithm was used to identify overlapping sample pairs. The Euclidean distance between all sample pairs was calculated, and the threshold was set to 0.35 standard deviation units. 217 sample pairs with different labels and mutual nearest neighbors were identified and removed. Then, the Edited Nearest Neighbors ENN algorithm was applied, with the number of neighbors k equal to 3. The majority class of the three nearest neighbors of the remaining samples was checked to see if they were consistent with the sample itself. If they were inconsistent, they were removed. A total of 89 noise points were removed. Finally, 433 samples with label 1 were retained, and the number of samples with label 0 was adjusted to 454, so that the positive and negative sample ratio reached 1:1.048, forming an emergency avoidance balanced sample. The total number of samples was 887. Each sample maintained a 21-dimensional feature structure, and the time label was completely retained without normalization or standardization.

[0018] Step S2 includes the following steps: Step S21: Reconstruct the avoidance scenario based on the emergency avoidance balance sample; and simultaneously perform an association mapping on the misoperation braking state based on the avoidance scenario to obtain avoidance scenario-misoperation mapping data. Step S22: Perform user misoperation braking anomaly analysis on the avoidance scenario-misoperation mapping data to obtain braking anomaly data; Step S23: Calculate the disordered braking vector intensity in the time dimension based on the braking anomaly data to obtain the disordered braking vector intensity; Step S24: Perform temporal segmentation clustering on the intensity of the disordered vectors to obtain disordered vector segmentation clustering data.

[0019] As an example of the present invention, reference is made to... Figure 2 As shown, in this example, step S2 includes: Step S21: Reconstruct the avoidance scenario based on the emergency avoidance balance sample; and simultaneously perform an association mapping on the misoperation braking state based on the avoidance scenario to obtain avoidance scenario-misoperation mapping data. In this embodiment of the invention, time-series data is extracted from each record in the emergency avoidance balance sample, including eight core physical quantity fields: longitudinal acceleration, lateral acceleration, yaw rate, steering wheel angle, accelerator pedal opening, brake pedal pressure, vehicle speed, and GPS coordinates. Combined with inherent vehicle parameters such as a vehicle mass of 1560 kg, distance from center of gravity to front axle of 1.3 m, distance from center of gravity to rear axle of 1.4 m, tire lateral stiffness of 85000 N / rad for the front tire and 78000 N / rad for the rear tire, air drag coefficient of 0.29, frontal area of ​​2.1 square meters, and rolling resistance coefficient of 0.015, three-dimensional kinematic scene reconstruction data is constructed. Rigid body six-degree-of-freedom motion equations are used, with Euler angles representing attitude changes. The time step is fixed at 0.01 s, and the integration method is the fourth-order Runge-Kutta method. Initial conditions... Based on the sensor readings at the start of the sample, the output is the vehicle's XYZ coordinates (in meters), yaw angle (in radians), and pitch angle (in radians), updated every 10ms. Simultaneously, a malfunction braking status label (binary value 1 or 0) is extracted from the sample. This label is then time-aligned and embedded into the 9th column of the scene matrix to form malfunction mapping data for the avoidance scenario. This data structure consists of 500 rows by 9 columns, with each row corresponding to a 10ms time window. Columns 1-8 represent physical state quantities, and column 9 represents the operation status label. All 887 samples undergo the same processing procedure without omission, interpolation, or extrapolation. All values ​​retain the original sensor accuracy: longitudinal acceleration is retained to 3 decimal places (in m / s²), steering wheel angle is retained as an integer (in degrees), and vehicle speed is retained to 2 decimal places (in m / s).

[0020] Step S22: Perform user misoperation braking anomaly analysis on the avoidance scenario-misoperation mapping data to obtain braking anomaly data; In this embodiment of the invention, user misoperation braking anomaly analysis is performed on the misoperation mapping data of the avoidance scenario. First, 500 rows of data for each sample are scanned to extract the throttle acceleration state. The judgment condition is that the longitudinal acceleration value is greater than 12m / s² within 150 consecutive rows (1.5s), and the throttle pedal opening jumps from less than 30% to greater than 90% within any consecutive 10 consecutive rows (100ms). If both of the above two conditions are met, the current time window is marked as a throttle misoperation anomaly, and the acceleration state is marked as 1; otherwise, it is 0. Next, the directional steering operation state judgment condition is extracted. If the absolute value of the cumulative change in steering wheel angle within a 1-second window is greater than or equal to 90 degrees and occurs at least 3 times within that window, and the absolute value of the angle difference for each steering switching action is greater than 45 degrees, then it is marked as a steering misoperation anomaly with a steering flag value of 1; otherwise, it is 0. The accelerator misoperation anomaly and acceleration status flag sequence are bitwise ORed with the steering misoperation anomaly and steering flag sequence to generate the final braking anomaly data. This data is a binary sequence of length 500, with each bit corresponding to the anomaly trigger status of a 10ms time window, where 1 indicates an anomaly and 0 indicates normal.

[0021] Step S23: Calculate the disordered braking vector intensity in the time dimension based on the braking anomaly data to obtain the disordered braking vector intensity; In this embodiment of the invention, the braking disorder vector intensity calculation in the time dimension is performed based on abnormal braking data. First, all time window indices marked as 1 are located to form an abnormal time period list. For each abnormal time period, the longitudinal acceleration sequence of 200ms before and after it (20 sampling points) is extracted. The difference between each two adjacent points is calculated to obtain the acceleration gradient sequence. Then, the difference between the maximum and minimum values ​​in the gradient sequence is calculated to obtain the acceleration ratio difference in m / s³. Simultaneously, the steering wheel angle sequence within the same time period is extracted, and its standard deviation is calculated to obtain the steering angle sudden sway amplitude in degrees. The acceleration ratio difference and the steering angle sudden sway amplitude are used as inputs to perform yaw impact vector intensity analysis, using a cubic method. Spline interpolation was performed on the two sets of data. The interpolation node interval was set to 0.2s. After interpolation, the number of data points was uniformly set to 25. Nonlinear regression was performed on the interpolated sequence. The basis function was selected as Legendre polynomial with an order of 3. After fitting, the lateral sway impact vector regression intensity sequence was output. This sequence contains 25 two-dimensional vectors. The first component of each vector is the longitudinal impact intensity in N, and the second component is the lateral sway intensity in N·m. The sliding window standard deviation was calculated on the regression intensity sequence. The window length is 5 points and the step size is 1 point. The braking disorder vector intensity is output with a length of 21 points. Each point is a scalar value in N·m·s, which represents the degree of local dynamic turbulence.

[0022] Step S24: Perform temporal segmentation clustering on the intensity of the disordered vectors to obtain disordered vector segmentation clustering data.

[0023] In this embodiment of the invention, temporal segmentation clustering is performed on the disordered braking vector intensity sequence. First, a clustering feature vector is constructed. The feature vector corresponding to each time point contains three items: the first item is the time index value, ranging from 1 to 21; the second item is the intensity value of the current point; and the third item is the intensity difference with the previous point. If it is the first point, the difference is 0, forming a feature matrix of 21 rows by 3 columns. The K-means clustering algorithm is then applied to this matrix. The number of clusters is K, which is fixed at 5. The initialization method is to randomly select 5 sample points as initial center points using Forgy. The distance metric used is Manhattan distance. The iteration termination condition is that the distance moved by the cluster center for 3 consecutive rounds is less than 0.001. Alternatively, the maximum number of iterations can reach 100 rounds. After clustering, 5 clusters are output. Each cluster contains a list of its time points, a triplet of cluster center coordinates, and the average Manhattan distance from the sample within the cluster to the center. The time points of each cluster are arranged in ascending order, and adjacent time points are merged to form continuous time periods. The start index and end index are marked to generate unordered vector segmented clustering data. This data structure consists of 5 rows, each containing 5 fields: the segment start time point number, the segment end time point number, the segment center intensity value, the segment duration length (in seconds), and the standard deviation of the intensity fluctuation within the segment. All 887 samples are clustered independently using this clustering process.

[0024] Step S22 includes the following steps: Extract the throttle acceleration and steering operation states of the user in the accident avoidance scenario-misoperation mapping data; When the acceleration under throttle is greater than 12 If the duration is greater than 1.5s and the accelerator pedal opening change is extracted by the engine control unit, and the accelerator pedal opening change increases from a rate of increase of more than 60% within 1s, it is determined to be an abnormal acceleration state due to accelerator pedal mis-pressing. If the steering angle is greater than or equal to 90°~180° within 1 second and the steering frequency is greater than or equal to 3 times within 1 second, it is judged as an abnormal steering operation due to steering error. Based on the abnormal acceleration state due to accidental accelerator pedal press and the abnormal steering due to accidental steering, the abnormal braking of the user's misoperation in the avoidance scenario is analyzed to obtain braking anomaly data.

[0025] In this embodiment of the invention, the extraction of the throttle acceleration state and steering operation state of the user operation in the accident avoidance scenario-misoperation mapping data is achieved using multi-sensor fusion technology. First, raw data is extracted from the vehicle control domain bus network using the CAN-FD communication protocol, with a baud rate of 2Mbps, an extended frame format, a 29-bit identifier length, a maximum data field length of 64 bytes, and a data sampling frequency set to 200Hz to ensure the capture of transient operation characteristics. Wheel speed sensor data with an accuracy of 0.01km / h is obtained from the vehicle chassis domain controller, and data from the vehicle stability control system is also analyzed. The control system acquires data from the accelerometer and gyroscope with accuracies of 0.005g and 0.01 degrees / second, respectively. It acquires steering wheel angle and torque sensor data from the steering system controller with accuracies of 0.1 degrees and 0.02 N·m, respectively. It acquires accelerator pedal position sensor data from the powertrain controller with an accuracy of 0.1%. A signal denoising algorithm is applied to the extracted raw data, employing wavelet denoising technology using the Daubechies wavelet with a decomposition level of 4. Stein unbiased risk estimation is used for threshold selection, and soft thresholding is employed for threshold reduction. The signal-to-noise ratio is improved. For noise levels exceeding 15dB, feature extraction is performed on the denoised data to calculate the throttle acceleration state feature vector, which includes five parameters: throttle pedal position, throttle pedal speed, throttle pedal acceleration, vehicle longitudinal acceleration, and engine speed. The sliding window method is used to calculate the statistical characteristics of these parameters at different time scales, with window sizes of 50ms, 100ms, 200ms, 500ms, and 1000ms. Features include mean, standard deviation, peak value, trough value, and root mean square value. The same method is used to extract feature vectors for steering operation states, including steering wheel angle. Five parameters—steering wheel angular velocity, steering wheel angular acceleration, vehicle yaw rate, and lateral acceleration—were used to calculate statistical features within the same time window. Principal component analysis was applied to the extracted feature vectors to reduce their dimensionality, retaining principal components that explained 95% of the variance. The dimensionality of the feature vectors was reduced from 125 to 32, achieving data compression while preserving key information. This resulted in a feature dataset of throttle acceleration and steering operation states, containing 22,051 records. Each record includes a 32-dimensional feature vector and timestamp information, providing a data foundation for subsequent abnormal state determination.

[0026] When the acceleration during throttle acceleration exceeds 12 m / s² and lasts for more than 1.5 seconds, and the throttle pedal opening change is simultaneously extracted by the engine control unit, if the rate of increase in throttle pedal opening change exceeds 60% within 1 second, it is determined to be an abnormal acceleration state due to throttle misapplication. This determination process employs a multi-threshold joint decision mechanism. First, a high-precision accelerometer measures the vehicle's longitudinal acceleration (range ±16g, resolution 0.002g, sampling frequency 200Hz). The acceleration signal is processed by a second-order Butterworth low-pass filter with a cutoff frequency of 25Hz to filter out high-frequency noise. An acceleration threshold judgment module is then established, setting the acceleration threshold to [value missing]. This value was determined through statistical analysis of 12,580 emergency evacuation events and falls within the normal acceleration distribution. Outside the boundary, the system continuously monitors the filtered acceleration signal. When the detected acceleration exceeds a threshold, a timer is started with a timing accuracy of 1ms. The system continuously monitors the acceleration value until the acceleration drops below the threshold or the timer reaches the set duration threshold of 1.5s, which is determined based on human reaction time and vehicle dynamics characteristics. Simultaneously, the system reads accelerator pedal opening data through the OBD-II interface of the engine control unit. The interface protocol conforms to the ISO15765-4 standard, with a baud rate of 500kbps and a reading frequency of 50Hz. The accelerator pedal opening sensor is a Hall effect type with an accuracy of 0.2% and a range of 0-100%. The system records the accelerator pedal opening change within a 1-second time window and calculates the opening change through discrete difference. The system uses a differential step size of 20ms to determine the throttle acceleration rate. When the maximum rate of change exceeds 60% / s, an abnormal throttle acceleration rate flag is triggered. This threshold is determined by comparing the throttle operation characteristics in normal driving and emergency avoidance events. In normal driving, 95% of throttle operation changes are less than 40% / s. The system uses AND gate logic to judge three conditions: acceleration greater than 12m / s², duration greater than 1.5s, and throttle pedal opening change rate greater than 60% / s. When all three conditions are met, the system determines that the throttle is accidentally pressed and there is an abnormal acceleration state. The judgment result is stored in the vehicle event recorder in Boolean form. At the same time, the system records the timestamp of the trigger, vehicle speed, acceleration, and throttle position, forming an abnormal event record, which provides a basis for subsequent braking anomaly analysis.

[0027] When a steering operation occurs with an angle greater than or equal to 90°~180° within 1 second, and the frequency of steering operations is greater than or equal to 3 times within 1 second, it is determined to be an abnormal steering operation due to incorrect steering. This determination process is implemented using time-frequency analysis technology. First, steering angle data is collected by the steering wheel angle sensor of the electric power steering system. The sensor type is a photoelectric encoder with a resolution of 0.1 degrees, a range of ±720 degrees, and a sampling frequency of 100Hz. The raw angle signal is filtered to remove glitches, with a window size of 5 sampling points. The filtered signal is then processed... A first-order low-pass filter is used for smoothing, with a cutoff frequency of 15Hz. A steering wheel angle change detection module is established, employing a sliding window method to calculate the range of steering wheel angle changes within one second. The window size is 100 sampling points, with a step size of one sampling point. For each window, the difference between the maximum and minimum angle values ​​is calculated to obtain the angle change. The angle change threshold is set to [90°, 180°], determined through statistical analysis of 6450 cases of steering misoperation. When the angle change within the window falls within the threshold range, an abnormal angle change flag is triggered. Simultaneously, the system uses Fast Fourier Transform (FFT)... The frequency characteristics of the steering wheel angle signal are analyzed by performing an FFT on the data within a 1-second window. A Hanning window is selected as the window function, and the number of FFT points is set to 128. The dominant frequency components are extracted from the obtained spectrum, and the proportion of energy concentrated in the frequency range above 2Hz is calculated. The steering operation frequency is determined by counting the number of zero-crossing points in the steering wheel angle signal. A threshold of 6 zero-crossing points within 1 second is set, corresponding to an operation frequency of 3 times / second. This threshold is based on analysis of normal driving data, where 99% of the operation frequency during normal driving is below 2 times / second. The system uses a fuzzy logic controller to comprehensively evaluate angle changes. The fuzzy controller uses two parameters: normalized angle change and operating frequency as input variables and anomaly score as output variable, ranging from 0 to 100. The fuzzy rule base contains 9 rules describing the relationship between angle change, operating frequency, and anomaly score. The Mamdani inference method and the centroid method are used to defuzzify the system. When the anomaly score exceeds 75, the system determines it as a steering error and abnormal steering. The determination result is recorded as an event, including key information such as trigger time, vehicle speed, yaw rate, and steering wheel angle sequence, providing basic data for subsequent braking anomaly analysis.

[0028] This paper analyzes user-induced braking anomalies in avoidance scenarios based on abnormal acceleration due to accelerator pedal misapplication and abnormal steering due to directional errors. A multi-feature fusion analysis method is employed. First, a feature library of user-induced braking anomalies is constructed, containing two main anomaly patterns: braking anomalies based on accelerator pedal misapplication and braking anomalies based on steering errors. The feature library is built based on 8745 historical anomaly data points. A hierarchical clustering algorithm is used to divide the anomaly data into 12 typical subclasses. The clustering method is Ward's minimum variance method, and the distance metric is Mahalanobis distance. Each subclass is represented by a feature vector with a dimension of 24, containing 16 time-domain features and 8 frequency-domain features. Newly detected anomalies are further analyzed. Feature vectors of the same dimension are extracted from each event, and their categories are determined by a minimum distance classifier using the K-nearest neighbor algorithm with K set to 5. Cosine similarity is used as the distance metric. For events identified as abnormal acceleration due to accelerator pedal misapplication, the system extracts braking system response features, including parameters such as brake pedal position, brake pressure, braking force distribution, and deceleration. The relationship between brake intervention timing and accelerator release time is analyzed, and the braking delay time is calculated. The normal range is 0.2-0.5 seconds, and the abnormal values ​​are greater than 0.8 seconds or less than 0.1 seconds. The time-frequency characteristics of the brake pedal displacement signal are analyzed using discrete wavelet transform, employing Haar wavelets with a decomposition level of 3, and extracting high-frequency coefficients to characterize pedal operation. To ensure smooth operation, for events deemed as abnormal steering due to steering error, the system analyzes the coordination between steering wheel and braking operations, calculating the time difference between the peak steering wheel angular velocity and the peak braking pressure. The normal coordinated operation time difference is 0.3-0.7 seconds, while abnormal values ​​are greater than 1 second or less than 0.2 seconds. Simultaneously, the system analyzes the matching degree between braking force distribution and steering demand by comparing the braking pressure difference between the left and right wheels with the ideal braking pressure difference. The ideal value is calculated using a vehicle dynamics model. Based on the analysis results of the two types of abnormal states, the system establishes a braking anomaly classifier using a random forest algorithm. The number of decision trees is set to 100, the maximum depth to 8, and the minimum number of leaf node samples to 10. Features are randomly selected. With a ratio of 0.7, braking anomalies are categorized into six types: slight delayed braking, severe delayed braking, insufficient braking, excessive braking, uneven braking force distribution, and braking-steering interference. The classification accuracy was evaluated using 10-fold cross-validation, achieving an accuracy rate of 92.3%. The final braking anomaly data includes anomaly type labels, anomaly severity scores, and time-series data of key braking system parameters. The anomaly data is stored in a structured format, including header information and data segments. The header information records the event ID, timestamp, vehicle information, and anomaly type. The data segments record the complete process data from 3 seconds before the anomaly occurs to 2 seconds after the anomaly ends, using a sampling rate of 100Hz, providing a data foundation for subsequent brake vector strength calculations.

[0029] Step S23 includes the following steps: Step S231: Extract the abnormal acceleration changes due to throttle misapplication and the abnormal steering angle due to directional misoperation in the time dimension of the braking anomaly data; Step S232: Calculate the rate of increase of the abnormal acceleration change due to accidental throttle pressing, and then calculate the difference in the rate of increase over a 2-second time interval; Step S233: Analyze the sudden change in steering angle within 2 seconds due to directional error and abnormal steering angle. Step S234: Analyze the yaw impact vector intensity based on the speed increase ratio difference and the sudden change in steering angle to obtain the yaw impact vector intensity; Step S235: Perform nonlinear regression analysis on the yaw impact vector intensity to obtain the yaw impact vector regression intensity; calculate the braking disorder vector intensity in the time dimension based on the yaw impact vector regression intensity to obtain the braking disorder vector intensity.

[0030] In this embodiment of the invention, abnormal acceleration changes due to throttle misapplication and abnormal steering angle changes due to directional misoperation are extracted from braking anomaly data over time. This is achieved using multi-source data synchronization analysis technology. The system first extracts braking anomaly data from the vehicle control domain network CAN-FD bus. The data frame format conforms to the ISO 11898-1 standard, with a baud rate of 5Mbps. The data includes the vehicle identification code, timestamp, and anomaly identifier. A time synchronization protocol is used to ensure that the synchronization accuracy of multi-sensor data reaches 1ms. Next, the system extracts records of abnormal acceleration changes due to throttle misapplication from the braking anomaly data. These records are collected by an acceleration sensor array, with the sensor model being Bosch. The MM5.10 sensor, with a range of ±16g, resolution of 0.0005g, and a sampling rate of 500Hz, uses a sensor array comprising four measurement points (front, rear, left, and right) to form a redundant measurement structure. It calculates the vehicle's center of gravity acceleration using a data fusion algorithm. The fusion algorithm employs a weighted average method, with weight coefficients determined offline using the least squares method. The optimization objective is to minimize the fusion error, with weight parameters set to [0.35, 0.25, 0.2, 0.2]. The fused acceleration data is smoothed using a Kalman filter with process noise covariance of 0.01 and measurement noise covariance of 0.08. The system extracts abnormal steering angle data from steering wheel angle sensors (magnetoelectric type, accuracy 0.05°, range ±900°, sampling rate 200Hz). The angle signal is processed by a bandpass filter (Butterworth third-order filter, passband 0.2Hz-10Hz) to remove low-frequency drift and high-frequency noise. The system then combines the extracted abnormal acceleration data from throttle misoperation with... The abnormal steering angle data due to throttle misoperation was organized into a time series data structure using a fixed time window segmentation technique. The window length was set to 2.5 seconds, and the window overlap rate was 50%. Each window contained 1250 sampling points of abnormal acceleration data due to throttle misoperation and 500 sampling points of abnormal steering angle data due to throttle misoperation. Feature extraction was performed on each time window, and statistical features including mean, standard deviation, maximum, minimum, peak and trough values, number of zero crossings, and autocorrelation coefficient were calculated. A 32-dimensional feature vector was generated for each window. After dimensionality reduction by principal component analysis, 16-dimensional principal components were retained. The retained principal components explained more than 98% of the variance. The system used a dynamic time warping algorithm to align the time series of different samples to eliminate the influence of time scale changes. Finally, aligned data sequences of abnormal acceleration changes due to throttle misoperation and abnormal steering angle changes due to throttle misoperation were generated, with a data volume of 8745 records, each record length of 2 seconds, and a uniform sampling rate of 200Hz, i.e., 400 sampling points. This lays the foundation for subsequent analysis of acceleration ratio and steering amplitude.

[0031] The rate of increase of abnormal acceleration changes due to throttle misapplication was calculated using a piecewise linear fitting method. First, the extracted abnormal acceleration data from throttle misapplication was divided into time windows, with each window size set to 50ms and adjacent windows overlapping by 10ms, resulting in 42 analysis windows over a 2-second time series. For each window, the linear trend of the acceleration data was calculated, and a straight line was fitted using the least squares method to obtain the slope value, representing the rate of change of acceleration within that window, in m / s³. The coefficient of determination (R²) of the fitted straight line must be greater than 0.85; otherwise, a quadratic curve fitting was used. The 42 acceleration data were then analyzed. The rate of change of acceleration was normalized, with a reference value set at 1.5 m / s³, a typical rate of change of acceleration under normal emergency acceleration conditions. The normalization formula used the ratio method, dividing the actual rate of change of acceleration by the reference value. The normalized value was defined as the growth rate ratio, which theoretically ranges from -∞ to +∞, but is actually concentrated in the range of -10 to +15. A time series of the growth rate ratio was established, containing 42 data points. A sliding window analysis was performed on this series, with a window size of 2 seconds to cover all 42 data points. The maximum and minimum values ​​of the growth rate ratio were calculated, and the difference between them was defined as... The growth rate ratio difference describes the fluctuation of acceleration change over a 2-second period, with a typical value range of 0-25. The system performs segmented analysis on the growth rate ratio difference, dividing the 2-second time into three stages: before, during, and after. Each stage contains 14 data points, and the mean growth rate ratio of each stage is calculated, denoted as [R_before, R_during, R_after]. The rate of change between the three stages is calculated and defined as [ΔR_before-during, ΔR_during-after], describing the time evolution characteristics of the growth rate ratio. The system establishes a growth rate characteristic vector, including the growth rate ratio difference, the three stage mean values, and the two rates of change, for a total of 6 dimensions. The system employs a weighted fusion method to comprehensively evaluate acceleration characteristics. The weight coefficients are determined through historical data analysis and set to [0.35, 0.15, 0.15, 0.15, 0.1, 0.1]. The weighted sum is calculated to obtain the comprehensive acceleration score, which ranges from 0 to 100. A higher score indicates a more dramatic change in acceleration. The system uses the acceleration ratio difference and the comprehensive acceleration score as key features and passes them to the yaw impact vector intensity analysis module to provide input parameters for subsequent yaw impact analysis. The calculation accuracy of the acceleration ratio difference reaches ±0.5, and the calculation accuracy of the comprehensive acceleration score reaches ±2.5.

[0032] This study analyzes the abrupt change in steering angle within a 2-second timeframe caused by erroneous steering input. A multi-scale spectral analysis method is employed. First, the acquired steering wheel angle signal is preprocessed, using a median filter to remove abnormal spikes within a window of 5 sampling points. Then, a Savitzky-Golay smoothing filter with a polynomial order of 3 and a window length of 9 sampling points is applied to process the signal, preserving signal transition characteristics while removing high-frequency noise. The system extracts data from the preprocessed steering wheel angle signal over a 2-second time window, totaling 400 sampling points. A short-time Fourier transform is then used to analyze the angle signal. Time-frequency analysis was performed using a Hamming window with a length of 64 points and an overlap rate of 50%, resulting in 12 time-frequency analysis segments with a frequency resolution of 3.125 Hz. Spectral energy distribution characteristics were extracted from each segment, with a focus on the 0.5 Hz–5 Hz frequency band, which contains characteristic frequencies of the driver's active steering maneuvers. The system calculated the maximum rate of change of the steering wheel angle in degrees per second, using the central difference method to calculate the angular velocity with a difference step size of 5 ms. The maximum rate of change typically exceeds 300 degrees per second during abnormal steering maneuvers. The system defined the sudden change amplitude of the steering angle as the peak-to-peak value of the steering wheel angle within a 2-second window. That is, the maximum value minus the minimum value. The calculation method is to search for the global maximum and minimum values ​​within 2 seconds and take their absolute difference. The typical range of sudden sway amplitude is 90 degrees-540 degrees. For sway amplitude exceeding 360 degrees, the system analyzes the number of complete steering wheel cycles, uses the zero-crossing counting method to identify complete cycles, and records the number of cycles and the duration of each cycle. The system further analyzes the frequency characteristics of steering operation, using an adaptive window autocorrelation analysis method. The window size is adaptively adjusted between 100ms and 500ms to calculate the dominant frequency of the angle signal, with a frequency range of 0.5Hz-10Hz. The normal operating frequency is usually below 1.5Hz, while the abnormal operating frequency can reach 3Hz-8Hz. The system calculates the differential entropy of the steering angle signal to quantify the uncertainty of the signal. The higher the entropy value, the more disordered the steering operation. The entropy value range of normal steering operation is 1.2-2.5, and the entropy value range of abnormal operation is 2.6-4.8. The system integrates four parameters: steering angle sudden change amplitude, maximum rate of change, operating frequency, and differential entropy, to establish a steering anomaly feature vector. The feature vector is fused by a weighted sum method with weight coefficients of [0.4, 0.25, 0.2, 0.15] to obtain a comprehensive steering anomaly score, with a score range of 0-100.

[0033] The analysis of yaw impact vector intensity based on the speed increase ratio difference and the sudden change in steering angle is achieved using a multi-physical quantity coupling analysis method. First, a vehicle kinematic model is established, containing seven degrees of freedom: longitudinal motion, lateral motion, yaw motion, and rotation of the four wheels. Model parameters are determined based on the vehicle's physical characteristics, including a vehicle mass of 1850 kg, a moment of inertia of 2860 kg·m², a wheelbase of 2.78 m, front and rear track widths of 1.62 m and 1.63 m respectively, and a center of gravity height of 0.56 m. The system maps the speed increase ratio difference to a longitudinal impact component using a piecewise linear function. When the speed increase ratio difference is less than 5, the longitudinal component = speed increase ratio difference × 4; when the speed increase ratio difference is between 5 and 15... When the longitudinal component is greater than 15, the longitudinal component is calculated as 20 + the rate of increase difference × 6. When the rate of increase difference is greater than 15, the longitudinal component is calculated as 80 + the rate of increase difference × 2. The longitudinal impact component has dimensions in N·s and a theoretical range of 0-120. The system maps the sudden change in steering angle to a lateral impact component. The mapping also uses a piecewise linear function. When the sudden change in angle is less than 120 degrees, the lateral component is calculated as sudden change in angle × 0.25. When the sudden change in angle is between 120 and 300 degrees, the lateral component is calculated as 30 + (sudden change in angle - 120) × 0.4. When the sudden change in angle is greater than 300 degrees, the lateral component is calculated as 102 + (sudden change in angle - 300) × 0.1. The lateral impact component has dimensions in N·s and a theoretical range of 0-12. 6. The system calculates the longitudinal and lateral coupling effects. The coupling coefficients are determined through the vehicle dynamics model, considering the influence of center of gravity height, wheelbase, track width, and road adhesion coefficient. The adhesion coefficient is divided into four typical values: 0.85 for dry asphalt pavement, 0.6 for wet asphalt pavement, 0.3 for snow-covered pavement, and 0.1 for ice surface, with corresponding coupling coefficients of 0.15, 0.25, 0.4, and 0.55, respectively. The coupling term is calculated as longitudinal component × lateral component × coupling coefficient. The system calculates the yaw impact vector intensity using a three-component synthesis method, weighting the longitudinal component, lateral component, and coupling term. The weighting coefficients are [0.35, 0.45, 0.2], and the normalization coefficient is 0.01. The final yaw impact vector is calculated using this method. The intensity value ranges from 0 to 100. The system further analyzes the directional characteristics of the yaw impact vector and calculates the angles of the longitudinal and transverse components. The angle ranges from 0 to 90°. The larger the angle, the higher the proportion of the transverse component. The system decomposes the yaw impact vector intensity into four sectors according to direction: longitudinal dominant zone (0-22.5°), longitudinal and transverse mixed zone (22.5-45°), transverse and longitudinal mixed zone (45-67.5°), and transverse dominant zone (67.5-90°). The impact characteristics of different sectors are significantly different, and different braking control strategies need to be adopted. The system calculates the projection components of the yaw impact vector intensity in the four sectors to form a four-dimensional feature vector. This vector describes the distribution characteristics of the yaw impact in different directions.

[0034] Nonlinear regression analysis of the yaw impact vector intensity was performed using support vector regression. First, the original yaw impact vector intensity data was divided into a test set and a training set at a 2:8 ratio. The training set contained 7000 records, and the test set contained 1745 records. Each record contained the yaw impact vector intensity value and its four sector projection components. The system performed standardized preprocessing on the training data using Z-score standardization, resulting in a mean of 0 and a standard deviation of 1. The standardized parameters were saved for subsequent test data processing. The system constructed a support vector regression model, selecting the radial basis function (RBF) as the kernel function. Hyperparameters were determined through grid search and 5-fold cross-validation, with the penalty parameter C set to 10. The parameters γ and ε were set to 0.05 and 0.01, respectively. The iterative training process employed a sequential minimum optimization algorithm with a maximum number of iterations of 1000 and a convergence threshold of 0.001. After training, the model's performance was evaluated on the test set. The mean absolute error was 2.15, the root mean square error was 3.42, and the coefficient of determination (R²) was 0.924. The system used the trained model to perform regression fitting on the yaw impact vector intensity data, generating a yaw impact vector regression intensity. This regression intensity was smoother, reducing noise and fluctuations in the original data. Based on the yaw impact vector regression intensity, the system performed time-dimension braking disorder vector intensity calculation. First, a dynamic response model of the braking system was established, with model parameters including the brake actuator time constant. With a braking pressure response delay of 15ms and a braking force build-up time of 80ms, the system uses a time series prediction method to estimate the trend of yaw impact vector intensity changes within the next 100ms. The prediction employs an autoregressive moving average model with order p=3 and q=2. The prediction result is combined with the current regression intensity to obtain the predicted yaw impact vector intensity. Based on the vehicle's current state and the predicted yaw impact vector intensity, the system calculates the disordered braking vector intensity, considering four correction factors: vehicle speed correction factor, road surface adhesion correction factor, vehicle sideslip correction factor, and steering state correction factor. The vehicle speed correction factor ranges from 0.8 to 1.2, increasing with increasing vehicle speed. The road surface adhesion correction factor ranges from 0.7 to 1.4. The vehicle sideslip correction factor ranges from 0.9 to 1.3, increasing with the sideslip angle. The steering state correction factor ranges from 0.85 to 1.15 and is related to the steering wheel angular velocity. The four correction factors work together through multiplication. The system maps the corrected predictive yaw impact vector strength to the braking disorder vector strength. The mapping uses a piecewise linear function. The dimensions of the braking disorder vector strength are N·m·s, representing the torque impulse that requires intervention from the vehicle braking system. The value ranges from 0 to 3000. The system updates the braking disorder vector strength value according to a control cycle of 20ms to ensure real-time control and stability. The final generated braking disorder vector strength serves as a key input for subsequent braking control.

[0035] Step S234 includes the following steps: The yaw trajectory orientation is analyzed based on the sudden change in steering angle; nonlinear time-series integration is performed on the difference in acceleration ratio to obtain acceleration growth integral data; The dynamic offset angle is calculated by performing dynamic offset angle calculations on the sudden change amplitude of the steering angle and the rapid left and right steering. The rotational inertia fluctuation is estimated based on the integral data of acceleration growth rate and dynamic offset angle, and then the partial derivative of angular momentum is calculated to obtain the rotational angular momentum partial derivative data. Based on the acceleration growth integral data, dynamic offset angle and rotational angular momentum partial derivative data, the yaw trajectory orientation is subjected to oblique resultant vector strength coupling analysis to obtain oblique resultant vector strength coupling data; where the oblique resultant vector describes the oblique nature of the vehicle's motion or force in both longitudinal and lateral directions, and contains strength information in both directions. Based on the coupled data of the oblique combined vector strength, the yaw impact vector strength is analyzed to obtain the yaw impact vector strength.

[0036] In this embodiment of the invention, the yaw trajectory azimuth is analyzed based on the sudden change in steering angle. First, the four sudden change in steering angle yaw values ​​output in step S233 are extracted, in degrees. Each value corresponds to a 2-second time window. A sign determination is performed on each yaw value. If the positive cumulative change in the original steering wheel angle sequence is greater than the negative cumulative change within the window, the trajectory azimuth is marked as a right deviation +1; otherwise, it is marked as a left deviation -1. Simultaneously, the maximum instantaneous angular velocity within the window is calculated, in degrees / s, as the azimuth intensity weighting factor. The azimuth sign is multiplied by the intensity weight to obtain the signed yaw trajectory scalar value, a total of four values, which are used for subsequent coupling operations. Simultaneously, a nonlinear time-series integration operation is performed on the growth rate ratio difference sequence, with the input being the four growth rate ratio difference amplitude values ​​output in step S232, in m / s. 4The trapezoidal numerical integration method is used to average the values ​​of two adjacent points at 0.5s intervals and then multiply by 0.5s. The acceleration growth integral data is accumulated and the unit is m / s³. The integration starts at point 0 and ends at point 4. A total of 4 integral values ​​are output, which correspond to the longitudinal dynamic energy accumulation of the 4 time segments. The dynamic offset angle calculation is performed on the sudden change in steering angle. First, the extreme points within each 2-second window are extracted from the original steering wheel angle sequence, including local maximum and minimum values. The absolute value of the angle difference between adjacent extreme points is calculated. If the difference is greater than 45° and the sign changes alternately, it is counted as one sudden steering action. The number of sudden steering actions in each window is counted. If the number is greater than or equal to 3, the dynamic offset angle calculation is started. The dynamic offset angle is equal to the average of the angle differences of all sudden steering actions in the window multiplied by the attenuation coefficient 0.85, in degrees. If the number of sudden steering actions is less than 3, the dynamic offset angle is set to 0. Four dynamic offset angle values ​​are output, each accurate to one decimal place and ranging from 0 to 90°.

[0037] Based on the integral data of acceleration growth rate and the dynamic offset angle, the rotational inertia fluctuation is estimated. First, a simplified rigid body model of the vehicle is constructed with a mass of 1560 kg, a center of mass height of 0.52 m, a wheelbase of 2.7 m, a front track of 1.56 m, and a rear track of 1.54 m. The equivalent lateral displacement is calculated based on the dynamic offset angle, using the formula: center of mass height multiplied by the angle in radians, to obtain the lateral offset distance in meters. Then, the instantaneous lateral force in N is calculated using the integral data of acceleration growth rate, equal to the mass multiplied by the lateral acceleration. The lateral acceleration is obtained by the second-order difference of the lateral offset distance over time, with a sampling rate of 2 Hz. Finally, the torque around the Z-axis is calculated. The moment of inertia is calculated in N·m, which is equal to the lateral force multiplied by the height of the center of mass. The instantaneous moment of inertia is estimated based on the relationship between torque and angular acceleration, in kg·m². The moment of inertia is calculated by dividing the torque by the angular acceleration obtained from the yaw rate difference, in rad / s². One estimated moment of inertia value is output for each time window, for a total of 4 values. Then, the partial derivative of angular momentum is calculated by taking the angular momentum values ​​of adjacent time windows, in kg·m² / s. The first-order central difference is performed to obtain the rate of change of angular momentum, i.e., the partial derivative of angular momentum data, in kg·m² / s², for a total of 3 output values, corresponding to the transition segments of time windows 1 to 2, 2 to 3, and 3 to 4, respectively.

[0038] Based on the aforementioned acceleration growth integral data, dynamic offset angle, and rotational angular momentum partial derivative data, an oblique resultant vector strength coupling analysis is performed on the yaw trajectory orientation. First, the acceleration growth integral data is taken as the longitudinal component (m / s³). The dynamic offset angle is converted to radians, multiplied by the vehicle mass of 1560 kg, and then multiplied by the gravitational acceleration of 9.8 m / s² to obtain the lateral component (N). The rotational angular momentum partial derivative data is taken as the coupling weight factor (kg·m² / s²). A ternary weighted synthesis operation is performed, multiplying the longitudinal component by the weight factor to obtain the corrected longitudinal strength, and multiplying the lateral component by the weight factor to obtain the corrected lateral strength. The corrected longitudinal strength and the corrected lateral strength are combined into a two-dimensional vector. The oblique resultant vector strength coupling data (N·m / s³) is obtained by calculating the Euclidean norm. The same operation is performed independently on all three sets of transition segment data to output three oblique resultant vector strength coupling values. Each value is retained to four decimal places and is processed without normalization, standardization, or dimensionlessness. The yaw impact vector intensity is analyzed based on the oblique resultant vector intensity coupling data. The method is to perform sliding maximum filtering on the three oblique resultant vector intensity coupling values, with a window length of 2 and a step size of 1, outputting two peak intensity values. The larger of the two values ​​is multiplied by the time integration factor of 2s to obtain the energy equivalent value in N·m·s. Then, it is divided by the effective length of the tire ground imprint of 0.22m to obtain the equivalent yaw impact force in N. Finally, it is multiplied by the lever arm coefficient of 0.52m, i.e., the center of mass height, to obtain the yaw impact vector intensity in N·m. The final scalar value is output to three decimal places and used for the regression analysis in step S235.

[0039] Step S3 includes the following steps: Step S31: Perform convolution processing on the unordered vector segmented clustered data to obtain unordered vector segmented convolution data; Step S32: Perform braking timing intervention correction sensing based on unordered vector segmented convolution data to obtain braking intervention correction data; Step S33: Iteratively learn the braking intervention correction data to generate optimized braking intervention correction data; Step S34: Send the braking intervention correction optimization data to the new energy vehicle control terminal to execute the new energy vehicle safe driving control.

[0040] As an example of the present invention, reference is made to... Figure 3 As shown, step S3 in this example includes: Step S31: Perform convolution processing on the unordered vector segmented clustered data to obtain unordered vector segmented convolution data; In this embodiment of the invention, convolution processing of disordered vector segmented clustered data is achieved using multi-channel temporal convolutional network technology. First, the disordered vector segmented clustered data is organized into a standard temporal data structure with a data dimension of [number of samples × time step × number of features], i.e., [8745 × 400 × 16]. The number of samples represents the total number of emergency avoidance events, the time step corresponds to the number of sampling points within a 2-second sampling window, and the number of features includes key parameters such as the yaw impact vector intensity value, cluster label, and confidence level. The system performs normalization preprocessing on the temporal data using the Z-score standardization method, ensuring that the mean of each feature is 0 and the standard deviation is 1. The standardization parameters are calculated through batch statistical analysis. The normalized data is then input into the multi-channel convolutional processing module, which uses a multi-channel convolutional processing module. The system employs a multi-layer convolutional architecture, comprising three convolutional layers. The first layer uses 32 filters with a 3×1 kernel, a stride of 1, SAME padding, and ReLU activation. The second layer uses 64 filters with a 5×1 kernel, a stride of 2, SAME padding, and ReLU activation. The third layer uses 128 filters with a 7×1 kernel, a stride of 2, SAME padding, and ReLU activation. Each convolutional layer is followed by batch normalization with parameters set to momentum 0.99, ε 0.001, and decay rate 0.9. Residual connections are applied between convolutional layers to improve gradient propagation efficiency. The residual blocks use identity mapping. Parallel computation of the convolutional processing is performed on an NVIDIA Jetson AGX processor. Implemented on the Xavier platform with 512 CUDA cores, the system achieves a computational performance of 32 TOPS and a processing latency of less than 5ms. The system applies a channel attention mechanism to the feature map output from the third convolutional layer, using a Squeeze-and-Excitation module with a compression ratio of 16. This mechanism adaptively adjusts the weights of different channels, highlighting important features and suppressing secondary features. The feature map after channel attention processing applies a spatial attention mechanism using a self-attention algorithm with 8 head layers, 64 hidden layer dimensions, and a dropout rate of 0.1. Spatial attention highlights key time periods in the temporal data. The system combines the features processed by channel attention and spatial attention... The feature maps are fused using a weighted summation method, with weights automatically learned through backpropagation. The fused features undergo global average pooling and global max pooling, and the pooled features are concatenated to form a compact representation. Finally, dimensionality reduction is achieved through 1×1 convolutions with 64 filters, resulting in unordered vector segmented convolutional data with a data dimension of [8745×100×64]. The time step is compressed to 1 / 4 of the original, and the feature dimension is expanded to 4 times the original. The convolutional data contains the key patterns and temporal features of the original clustering data. The system performs quality assessment on the convolutional data, calculating the signal-to-noise ratio (SNR), which improves the average SNR by 15.3 dB. The feature discrimination is evaluated using the silhouette coefficient, improving from 0.68 in the original data to 0.The value 85 indicates that convolution processing effectively extracted key features from the disordered vector segmented clustering data while suppressing noise interference. The system then transmits the disordered vector segmented convolution data to the subsequent braking timing intervention correction sensing module, providing a foundation for precise intervention control.

[0041] Step S32: Perform braking timing intervention correction sensing based on unordered vector segmented convolution data to obtain braking intervention correction data; In this embodiment of the invention, the braking timing intervention correction perception based on unordered vector segmented convolutional data is implemented using a multi-stage decision fusion technique. First, temporal features are extracted from the unordered vector segmented convolutional data. A bidirectional gated recurrent unit network (BiGRU) is used, with a network structure containing two layers of BiGRU, each layer having 128 hidden units, an activation function of tanh, a recurrent dropout rate of 0.2, and an input dropout rate of 0.3. The network is trained using the Adam optimizer with a learning rate of 0.001, β1 of 0.9, β2 of 0.999, and ε of 1. The system is configured with a size of 64 and 100 training iterations. The BiGRU network extracts features containing long-term dependencies in the time series. It extracts segmented directional offset angles and vehicle speed stall values ​​from convolutional data across different time periods. The directional offset angles are recovered from the convolutional features using a deconvolution algorithm. Deconvolution is implemented using a transposed convolutional layer with a kernel size of 4, a stride of 2, and 16 channels. Vehicle speed stall values ​​are also recovered using deconvolution with the same parameter settings, achieving a recovery accuracy of over 95% of the original data. The system quantizes the vehicle offset angle based on the recovered directional offset angles, employing an adaptive quantization algorithm to divide continuous angle values ​​into 7 discrete levels: minimal offset (0-3 degrees), small offset (3-8 degrees), medium-small offset (...). The system is categorized into four offset levels: 8-15 degrees, medium offset (15-25 degrees), medium-large offset (25-35 degrees), large offset (35-50 degrees), and extreme offset (above 50 degrees). Quantization thresholds are automatically determined using K-means clustering to ensure minimal intra-class distance and maximum inter-class distance. The system analyzes braking torque demand based on vehicle offset angle and vehicle speed stall values, employing a fuzzy logic controller. Input fuzzification uses triangular and trapezoidal membership functions. Angle inputs are divided into 7 fuzzy sets, and vehicle speed stall values ​​are divided into 5 fuzzy sets. The output fuzzy set represents the braking torque demand, categorized into 9 levels. The fuzzy rule base contains 35 IF-THEN rules. The inference method uses the Mamdani maximum-min algorithm, and defuzzification uses centroid... The fuzzy controller output range is 0-3000 N·m, corresponding to the actual braking torque demand. The system performs time-series smoothing of the braking torque demand based on the directional offset angle and vehicle speed stall value, employing a nonlinear exponential moving average algorithm. The smoothing factor α is adaptively adjusted with the rate of change of the offset angle; a large rate of change results in a small α value (0.2-0.4), while a small rate of change results in a large α value (0.6-0.8), ensuring rapid response under severe operating conditions and suppressing fluctuations under stable operating conditions. The system inputs the smoothed braking torque demand sequence into a time-series decision tree integrator. The integrator contains 5 decision trees with a tree depth of 8 and a minimum leaf node sample size of 15. The splitting criterion is Gini impurity, and the integration method uses weighted voting, with weights determined by gradient descent. To improve and optimize the system, the output of the decision tree integrator provides recommended values ​​for braking intervention timing and intensity. The system performs multi-source fusion based on temporal decision and fuzzy control results, employing Dempster-Shafer evidence theory. Three focal elements are defined: braking required, observation, and no braking required. The basic probability allocation of each source is calculated, and the credibility of the fusion decision is calculated using the DS combination rule. When the credibility of the "braking required" focal element exceeds 0.85, braking intervention is triggered. The intervention intensity is based on the fuzzy controller output, and the intervention timing is based on the temporal decision tree output. The system integrates various decision sources to generate braking intervention correction data. The data format includes key parameters such as intervention time point, intervention duration, intensity curve, and direction correction suggestions, providing input for subsequent iterative optimization.

[0042] Step S33: Iteratively learn the braking intervention correction data to generate optimized braking intervention correction data; In this embodiment of the invention, iterative learning of braking intervention correction data is implemented using a hierarchical reinforcement learning framework. First, a braking intervention state space is constructed with a state vector dimension of 18, containing 10 vehicle dynamic state parameters (vehicle speed, acceleration, yaw rate, sideslip angle, etc.) and 8 braking system state parameters (braking pressure, brake distribution ratio, wheel speed difference, etc.). The state space uses a discrete-continuous hybrid representation, and the action space is defined as a two-dimensional continuous space, representing the braking torque magnitude (0-3000 N·m) and braking force distribution ratio (0.3-0.7), respectively. The reward function design considers three aspects: safety, comfort, and handling, accounting for 50%, 30%, and 20% of the weights, respectively. The safety reward is based on the success rate of hazard avoidance, the comfort reward is based on the passenger's perceived acceleration stability, and the handling reward is based on the vehicle's response sensitivity. The system adopts a hierarchical reinforcement learning structure. The high-level strategy is responsible for deciding the timing of braking intervention, and the low-level strategy is responsible for executing braking torque control. The high-level strategy uses a deep Q-network (DQN) algorithm with a fully connected neural network structure of 4 layers and [128, ...]. The network structure is [128, 256, 128, 64], with LeakyReLU activation function, α = 0.01, empirical replay buffer size = 10000, batch size = 32, discount factor γ = 0.95, target network update frequency = 100 steps, low-level layer uses Deep Deterministic Policy Gradient (DDPG) algorithm, Actor network structure is [128, 256, 128, 64], Critic network structure is [256, 512, 256, 128], both networks use BatchNorm and Dropout regularization, Actor learning rate is 0.0001, Critic learning rate is 0.001, target network soft update parameter τ is 0.001, noise uses Ornstein-Uhlenbeck process, θ = 0.15, σ = 0.2. The system employs a simulation environment for reinforcement learning training. This environment is based on the CarMaker professional vehicle dynamics simulation platform and includes 16 typical avoidance scenarios and 4 road adhesion conditions. The simulation sampling frequency is 1000Hz, and the control frequency is 100Hz. The system performs distributed training for different operating conditions, using the Asynchronous Advantage Actor-Critic (A3C) framework. The system has 16 parallel environments, 8 worker threads, and 5 million training iterations. Policy evaluation is performed every 250,000 steps, with evaluation metrics including average reward, success rate, and control stability. The training process uses a course-based learning strategy, gradually increasing the difficulty from simple scenarios. The system applies the trained strategy to braking intervention correction data for optimization, employing an iterative optimization process. Each iteration includes three steps: prediction, evaluation, and adjustment. The prediction step uses reinforcement learning... The system generates candidate braking intervention strategies through learning, evaluates the expected returns of each strategy, and fine-tunes the strategy parameters based on the evaluation results. The iteration count is set to 5 times, and the convergence criterion is that the difference between strategies in two adjacent iterations is less than 3%. The system optimizes three key parameters—braking intervention timing, force curve, and duration—through iterative learning. For each braking intervention event, the system considers multiple alternatives and selects the one with the highest expected return as the final optimization result, generating braking intervention correction optimization data. The data format uses structured binary storage and includes control parameters, time-series curves, and metadata, totaling 8745 optimization records, each approximately 24KB in size, for a total of approximately 210MB. The system compresses the optimization data using a lossless compression algorithm, achieving a compression ratio of 3:1. The final optimized data is used for the actual execution of the vehicle control system.

[0043] Step S34: Send the braking intervention correction optimization data to the new energy vehicle control terminal to execute the new energy vehicle safe driving control.

[0044] In this embodiment of the invention, the transmission of braking intervention correction optimization data to the new energy vehicle control terminal adopts a multi-level secure transmission architecture. First, the braking intervention correction optimization data is encrypted using the AES-256 algorithm. Key management employs an elliptic curve cryptography system, selecting NIST P-256 for the curve and a 256-bit key length. Digital signatures utilize the ECDSA algorithm to ensure data integrity and source authentication. Encrypted data is transmitted via in-vehicle Ethernet, using the IEEE 802.3bw 100BASE-T1 standard with a bandwidth of 100Mbps. The physical layer uses single-pair twisted-pair cabling, achieving a transmission distance of up to 15m. The network topology adopts a star structure, with the central switch being vehicle-grade and supporting Time-Sensitive Networking (TSN) technology to ensure the real-time performance of critical data. Data transmission employs a layered protocol stack. The physical and data link layers use in-vehicle Ethernet, the network layer uses IPv6, and the transport layer uses a hybrid UDP and TCP mechanism. Critical real-time control data is transmitted via UDP, while configuration and management data are transmitted via TCP. The application layer uses a custom protocol with a 32-byte header containing data type, optimization parameters, and other information. Information such as priority, timestamp, and serial number are transmitted through a multi-level redundancy mechanism. Critical data is transmitted simultaneously through two physical channels, employing dual redundancy fault-tolerance technology to ensure normal system operation even in the event of a single point of failure. Upon arrival at the new energy vehicle control terminal, data undergoes identity authentication and integrity verification. Authentication utilizes a challenge-response mechanism, while verification employs the HMAC-SHA256 algorithm. After successful verification, the data is decrypted and loaded into the control terminal's memory. The control terminal hardware platform adopts a secure domain isolation architecture, with an NXPS32G automotive processor (2.0GHz, 8 cores), 4GB LPDDR4 memory, and 32GB eMMC storage. It is equipped with a hardware security module (HSM) supporting secure boot, real-time encryption / decryption, and key management. The system distributes the decrypted brake intervention correction and optimization data to three actuator controllers: the Brake Control Unit (BCU), the Motor Control Unit (MCU), and the Electronic Stability Control Unit (ESC). Distribution utilizes a Controller Area Network (CANFD) at a baud rate of 5Mbps, using extended frames, with priority based on identifier arbitration. The BCU receives braking force distribution and pressure control data, implementing precise braking torque control with a control resolution of 0.With a pressure of 1MPa and a response time of less than 15ms, the MCU receives motor torque modulation data to achieve coordinated control of regenerative braking and hydraulic braking, achieving a torque control accuracy of ±2% and a response time of less than 10ms. The ESC receives yaw stability control data and executes differential braking control with a stability control frequency of 100Hz. Before executing control, the system performs a final safety check, including data timeliness, control quantity rationality, and system state consistency. After passing the check, braking intervention control is performed according to a predetermined sequence. The control process consists of three stages: pre-braking stage (light braking force establishment, lasting 100-200ms), main braking stage (applying braking force according to the force curve, lasting 300-1500ms). The system operates in two phases: a braking phase (smoothly reducing braking force for 200-400ms) and an exit phase (smoothly reducing braking force for 200-400ms). Smooth transitions between phases are achieved through slope limits, ranging from 5-20 MPa / s. During execution, the system continuously monitors vehicle status and driver response. When driver intervention is detected (e.g., forcefully applying the brakes or sharply turning the steering wheel), the system immediately and smoothly exits automatic control at a slope of 30 MPa / s to ensure a smooth transfer of control. The system records the execution results in an event log, including trigger conditions, control parameters, vehicle response, and termination reasons. The log data is uploaded to a cloud server via 4G / 5G network for subsequent analysis and system optimization, completing the entire process of safe driving control for new energy vehicles.

[0045] Step S32 includes the following steps: Step S321: Extract the segmented direction offset angle and vehicle speed stall value from different time periods in the unordered vector segmented convolutional data; Step S322: Quantify the vehicle body offset angle based on the segmented directional offset angle and vehicle speed stall value at different time periods to obtain the vehicle body offset angle; perform braking torque demand analysis based on the vehicle body offset angle to obtain braking torque demand data. Step S323: Based on the segmented directional offset angle and vehicle speed stall value in different time periods, optimize the braking torque demand data for braking force smoothing control in different time periods to obtain braking force smoothing control data. Step S324: Based on the braking force smooth control data, perform braking timing intervention correction sensing to obtain braking intervention correction data.

[0046] In this embodiment of the invention, the segmented direction offset angle and vehicle speed stall values ​​for different time periods in the unordered vector segmented convolutional data are extracted using deconvolution reconstruction technology. First, the unordered vector segmented convolutional data is structurally analyzed. The convolutional data dimension is [8745×100×64], containing information on time compression and feature expansion. A transposed convolutional network is used for feature recovery. The network contains three transposed convolutional layers. The parameters of the first layer are set as follows: number of filters 32, kernel size 7×1, stride 2, padding method SAME, activation function ReLU. The parameters of the second layer are set as follows: number of filters 16, kernel size 5×1, stride 2, padding method SAME, activation function ReLU. The parameters of the third layer are set as follows: number of filters 8, kernel size 5×1, stride 2, padding method SAME, activation function ReLU. The network uses a 3×1 convolutional matrix with a stride of 1, SAME padding, and a linear activation function. Each transposed convolutional layer is followed by a batch normalization layer. The parameters are set to momentum 0.9 and ε 0.001. The output dimension of the transposed convolutional network is [8745×400×8]. This reconstructs the original time-series data at the original sampling rate. Segmented directional offset angle information is extracted from the reconstructed data using frequency domain feature backprojection. First, a short-time Fourier transform is applied to the reconstructed data with a window size of 128 points, an overlap rate of 75%, and a Hanning window function to obtain a time-frequency spectrum with a frequency resolution of 1.56Hz. Angle-related frequency band components are extracted from the time-frequency spectrum, mainly in the 0.5-4Hz band. The frequency band information is then converted back to time-frequency components using an inverse Fourier transform. The system obtains the time series of directional offset angles in the domain, with a sampling rate of 100Hz and a resolution of 0.1 degrees. The system divides a 2-second time window into five time segments, each 400ms long, corresponding to: preparation phase (0-400ms), initial response phase (400-800ms), mid-response phase (800-1200ms), stabilization phase (1200-1600ms), and recovery phase (1600-2000ms). Statistical features of the directional offset angle data within each time segment are calculated, including average angle, maximum angle, rate of change of angle, and angle fluctuation amplitude, forming angle feature vectors for each time segment. The system extracts vehicle speed stall values ​​from the recovered data using signal separation technology. First, the Independent Component Analysis (ICA) algorithm is applied to separate the mixed signal into independent components. The parameters are set to a maximum of 200 iterations and a tolerance of 0.0001. Among the separated independent components, the vehicle speed-related components are identified through spectral characteristics. The characteristic frequency of the vehicle speed signal is usually below 0.5Hz. The system applies a Kalman filter to smooth the extracted vehicle speed signal. The filter parameters are set to a process noise covariance of 0.01 and a measurement noise covariance of 0.1. The vehicle speed stall value is calculated. The stall value is defined as the difference between the actual vehicle speed and the expected safe vehicle speed. The expected safe vehicle speed is calculated through a curvature safe speed model. The model parameters include road curvature, lateral acceleration threshold (set to 3.5 m / s²), and safety margin coefficient (set to 0).85) The system also segments the vehicle speed stall values ​​into five time periods, calculating the average stall value, maximum stall value, stall change rate, and stall trend within each segment to form a stall feature vector for each time period. The system merges the segmented directional offset angle and vehicle speed stall values ​​to form a multidimensional feature tensor with dimensions [8745×5×16], where 8745 is the number of samples, 5 is the number of time periods, and 16 is the number of features for each time period. The feature tensor is stored in HDF5 format, supporting efficient random access. The extracted feature data undergoes integrity verification and outlier detection to ensure data quality. Integrity checks include missing value detection and continuity verification. Outlier detection uses a modified Z-score method with a threshold of 3.5, controlling the outlier ratio to within 0.3% of the total data volume to ensure data quality for subsequent analysis.

[0047] The vehicle body offset angle is quantified based on the segmented directional offset angles and vehicle speed stall values ​​at different time periods using adaptive clustering quantization technology. The system first performs time-weighted processing on the segmented directional offset angles, using an exponential decay function with weighting coefficients of [value missing]. ,in The attenuation coefficient is set to 0.5, and t is the time offset relative to the current time in seconds. Weighting emphasizes the importance of recent angle changes. The system applies a Gaussian mixture model to the weighted angle data for probability density estimation. The model parameters are set to 5 components, full covariance type, k-means++ initialization method, 200 maximum iterations, and a convergence threshold of 0.001. The GMM model fitted by the EM algorithm captures the multimodal characteristics of the angle distribution. The system designs a quantization threshold based on the GMM model and uses the minimum Bayesian information criterion (BIC) method to determine the optimal threshold, achieving adaptive quantization of the direction offset angle. The quantization levels are set to 7 discrete levels: minimal offset (0-2.5 degrees), small offset (2.5-7.5 degrees), medium-small offset (...). The system quantizes vehicle speed stall values ​​similarly, with five discrete levels: minor stall (0-5 km / h), slight stall (5-15 km / h), moderate stall (15-30 km / h), severe stall (30-50 km / h), and extreme stall (above 50 km / h). Based on the quantized directional offset angle and vehicle speed stall level, combined with the vehicle's current state parameters (including yaw rate, lateral acceleration, and tire slip angle), the system calculates the vehicle body offset angle using a nonlinear mapping network. The network structure is a fully connected feedforward network with a layer structure of [16, 32, 16, 8, 1] and the activation function is Leaky. The network uses ReLU with α of 0.1 and tanh as the output activation function. The mean squared error loss function is used for network training. The optimizer is Adam with a learning rate of 0.001, β1 of 0.9, and β2 of 0.999. The training data consists of 12,000 expert-annotated samples. The validation accuracy reaches a mean absolute error of 0.85 degrees. The quantized range of the vehicle body offset angle is -45 degrees to +45 degrees, with positive values ​​indicating rightward deflection and negative values ​​indicating leftward deflection. A multi-objective optimization method is used to analyze the braking torque demand based on the vehicle body offset angle. First, a vehicle yaw dynamics model is established, with model parameters including vehicle... With a body weight of 1850kg, a wheelbase of 2.85m, front and rear track widths of 1.62m and 1.65m respectively, a center of gravity height of 0.58m, and front and rear wheel lateral stiffness of 95000N / rad and 105000N / rad respectively, the system calculates the yaw moment required to stabilize the vehicle body based on a dynamic model. The calculation considers three key factors: sideslip angle, yaw rate, and vehicle speed. For a given body offset angle, the system calculates the yaw moment required to restore a neutral state, converts it into an equivalent braking torque requirement, and the conversion coefficient is determined based on the vehicle's geometric parameters and mass characteristics, with a typical value of 1.2-1.8. The system performs multi-objective optimization of braking torque demand. The optimization objectives include minimizing settling time, minimizing lateral offset distance, and minimizing the impact on ride comfort. The optimization algorithm employs Sequential Quadratic Programming (SQP), with constraints including a maximum braking torque limit (3500 N·m), a braking torque change rate limit (800 N·m / s), and a braking force distribution limit (front-rear wheel braking ratio range of 0.4-0.7). The optimization solution uses the augmented Lagrange multiplier method, with 50 iterations and a convergence threshold of 0.001. Based on the optimization results, the system generates braking torque demand data, including total braking torque, front-rear wheel distribution ratio, left-right wheel distribution ratio, and time-series curves. The torque range is 0-3500 N·m, the time resolution is 10 ms, and the curves are represented using piecewise cubic Hermite interpolation to ensure continuous first derivatives and smooth transitions.

[0048] Based on segmented directional offset angles and vehicle speed stall values ​​across different time periods, the braking torque demand data is optimized for smooth braking control across different time periods using model predictive control technology. The system first subdivides the 2-second control window into 8 time periods, each 250ms. Differentiated control strategies are designed for each time period. For the directional offset angles and vehicle speed stall values ​​in each time period, the system calculates the directional offset intervention correction angle and intervention correction speed using an adaptive PID controller. The PID parameters are optimized offline using a particle swarm optimization algorithm, with the optimization objective being to minimize overshoot and settling time. For the directional offset intervention correction angle, the PID parameter is set to Kp = 0.8-1. 0.2 (adaptive to speed), Ki=0.05-0.15, Kd=0.2-0.4. For intervention correction speed, the PID parameters are set to Kp=0.4-0.6, Ki=0.03-0.08, Kd=0.1-0.2. The system performs vehicle intervention correction braking shudder and yaw simulation calculations based on the intervention correction angle and correction speed to the braking torque demand data. A high-precision vehicle dynamics simulation model is used, which contains 14 degrees of freedom and considers suspension geometry, tire nonlinearity, and steering system characteristics. The simulation step size is set to 1ms. The simulation model parameters are calibrated through real vehicle testing, including suspension stiffness, damping characteristics, steering ratio, and tire magic resistance. The system simulates and calculates the vehicle body response under different braking torque inputs, including key parameters such as yaw rate, lateral acceleration, and sideslip angle, using parameters such as Formula. The system performs yaw and sway analysis on the simulation results, calculating the yaw and sway index during braking. The index is defined as the proportion of high-frequency component (5-15Hz) energy of the yaw rate. Generally, an index below 5% indicates good ride comfort, while an index above 15% indicates significant yaw. The system performs multivariate regression analysis on the yaw and sway data using a support vector regression algorithm. The kernel function is the radial basis function (RBF), with parameters C=10, γ=0.1, and ε=0.01. The regression model inputs are braking torque, braking force change rate, and road adhesion coefficient. The output is the predicted jitter and yaw index. The root mean square error of the regression model is 1.8%. Based on the jitter and yaw regression data, the system performs progressive control optimization on the steering offset intervention correction angle and intervention correction speed. The optimization adopts the model predictive control (MPC) framework, with the prediction time domain set to 500ms, the control time domain set to 100ms, and the sampling period set to 20ms. The state variables include the vehicle body slip angle, yaw rate, longitudinal velocity, and lateral velocity. The control variables are the braking torque and braking force distribution ratio. The system state equation is based on a linearized vehicle dynamics model. The control objective function is in the form of a weighted sum of squares. The weight matrix is ​​determined through multi-objective optimization. The constraints on the control input include the braking torque range [0, 3500] N·m, the torque change rate range [-800, 800] N·m / s, and the braking force distribution range [0.4, 0.[7] The optimization problem is solved in real time using a quadratic programming solver. The solver employs the active set method, with a maximum of 50 iterations and a computation time controlled within 5ms. Based on the MPC optimization results, the system generates an asymptotic correction angle and asymptotic correction speed. The rate of change of the asymptotic correction angle is limited to 15 degrees / second, and the rate of change of the asymptotic correction speed is limited to 5 km / h / second to ensure smooth changes in the control variables. The system performs smooth braking force control optimization based on the asymptotic correction angle and asymptotic correction speed, employing a nonlinear filter combination control strategy. The filters include low... The system includes a low-pass filter, a slope limiter, and a jitter suppressor. The low-pass filter has a cutoff frequency of 5Hz. The slope limiter parameters are set to an upward slope of 500 N·m / s and a downward slope of 600 N·m / s. The jitter suppressor uses an adaptive threshold method with a threshold range of 50-200 N·m, dynamically adjusted according to vehicle speed and road adhesion coefficient. The system generates smooth braking force control data including braking torque time series, braking force distribution parameters, and control timing markers. The data sampling interval is 10 ms, and the duration is 2 seconds, forming a 200-point control sequence.

[0049] The braking timing intervention correction perception based on braking force smoothing control data is achieved using a multi-stage decision fusion method. The system first performs time-series analysis on the braking force smoothing control data, using a sliding window method to extract key event points. The window size is 150ms, and the step size is 30ms. Time points where the rate of change of braking torque exceeds a critical value (set at 250 N·m / s) are identified as potential intervention timing candidates. The system calculates an intervention suitability score for each candidate point, considering four factors: vehicle stability margin, driver operating state, road adhesion conditions, and expected effect. The weights for each factor are 0.4, 0.3, 0.2, and 0.1, respectively. The stability margin is determined through phase space analysis. The calculation maps the vehicle's yaw rate and sideslip angle onto a two-dimensional phase plane, calculates the distance from the current state to the unstable boundary, and determines the driver's operating state through steering wheel angle, steering rate, and pedal operation. Road surface adhesion conditions are estimated through wheel speed differences and acceleration response. Expected effects are predicted through simulation. The system employs a Bayesian decision network to determine the intervention timing. The network structure consists of three layers: an evidence layer, an inference layer, and a decision layer. The evidence layer receives vehicle state parameters and braking force smoothing control data as input. The inference layer calculates the conditional probabilities of different intervention timings. The decision layer selects the intervention timing with the highest expected utility. The network parameters are trained using historical data with 15,000 training samples, achieving a validation accuracy of 93%.After determining the optimal intervention timing, the system calculates the precise braking intervention sequence. The sequence design includes three parts: a pre-braking phase, a main braking phase, and an exit phase. The pre-braking phase lasts 80-120 ms with a braking torque slope of 150 N·m / s. The main braking phase lasts 400-1200 ms with the braking torque varying according to the optimized curve. The exit phase lasts 150-250 ms with a braking torque slope of -200 N·m / s. The system calculates the directional correction assistance during braking intervention, based on the vehicle yaw dynamics model and steering system characteristics. Directional correction includes active steering compensation and torque superposition. The active steering compensation range is ±3 degrees, and the steering torque superposition range is ±5 N·m. The correction amount is dynamically adjusted during the braking process. The system further optimizes the transition characteristics of braking intervention by employing a combined state feedback and feedforward control method. Feedback control is based on the vehicle's sideslip angle and yaw rate, while feedforward control is based on the steering wheel angle and angular velocity. Controller parameters are controlled using H∞ robust control. Theoretical optimization ensures stability under uncertain parameter conditions. The system encapsulates the optimized braking intervention correction data into four parts: a control command header, a timing parameter block, a control quantity data block, and a status monitoring block. The control command header includes the command type, priority, sequence number, and checksum. The timing parameter block includes the duration, trigger time, and exit condition for each stage. The control quantity data block includes the braking torque sequence, distribution ratio, and direction correction. The status monitoring block includes key state variable thresholds and emergency handling strategies. The system performs integrity verification and timeliness marking on the braking intervention correction data. Integrity verification uses the CRC32 checksum algorithm, and timeliness marking includes a generation timestamp and validity period. Data transmission uses a real-time transmission protocol with the highest priority to ensure timely delivery of control commands. The braking intervention correction data generated by the system enables precise control of the timing, force, and duration of braking intervention in emergency avoidance scenarios, forming a closed-loop control system and providing core decision-making basis for the safe driving control system of new energy vehicles.

[0050] Step S323 includes the following steps: Based on the segmented directional offset angles and vehicle speed stall values ​​for different time periods, the directional offset intervention correction angle and intervention correction speed for different time periods are determined. Based on the aforementioned directional offset intervention correction angle and intervention correction vehicle speed, the braking torque demand data for different time periods are simulated and calculated to obtain vehicle intervention correction braking shudder and yaw data. Perform multivariate regression analysis on the jitter and yaw data to generate jitter and yaw regression data; Based on the jitter and yaw regression data, the directional offset intervention correction angle and intervention correction speed for different time periods are progressively controlled to obtain the offset progressive correction angle and progressive correction speed. Braking force smoothing control is optimized for different time periods based on the offset progressive correction angle and progressive correction vehicle speed, resulting in braking force smoothing control data.

[0051] In this embodiment of the invention, the determination of the intervention correction angle and intervention correction speed for different time periods based on the segmented directional offset angles and vehicle speed stall values ​​in different time periods is achieved using multi-channel adaptive control technology. First, the 2-second control window is divided into 8 time periods according to dynamic response characteristics: prediction segment (0-200ms), initial response segment (200-400ms), transition segment I (400-600ms), enhancement segment (600-900ms), stabilization segment (900-1200ms), transition segment II (1200-1500ms), weakening segment (1500-1800ms), and exit segment. (1800-2000ms) For each time period, extract the corresponding directional offset angle and vehicle speed stall value to construct a time-state matrix with a dimension of 8×2. Apply a non-linear state converter to the matrix data. The converter uses a piecewise function mapping to map the directional offset angle to a directional offset correction angle. The mapping rules are as follows: when the offset angle is less than 5 degrees, the correction angle = offset angle × 0.2; when the offset angle is between 5 and 15 degrees, the correction angle = 1 + offset angle × 0.5; when the offset angle is between 15 and 30 degrees, the correction angle = 6 + offset angle × 0.3; when the offset angle is greater than 30 degrees, the correction angle... =10.5 + offset angle × 0.15. According to experimental data, the sensitivity coefficient of right yaw correction is 5-8% higher than that of left yaw. Therefore, different parameters are set for left and right yaw cases respectively. The left yaw coefficient is the above value multiplied by 0.95. The system applies a similar piecewise function mapping to the vehicle speed stall value, mapping the vehicle speed stall value to the intervention correction speed. The mapping rules are: when the stall value is less than 5 km / h, the correction speed = stall value × 0.6; when the stall value is between 5-15 km / h, the correction speed = 3 + stall value × 0.8; when the stall value is between 15-30 km / h, the correction speed = 11 + stall value × 0.6; when the stall value is greater than... At 30 km / h, the corrected speed = 20 + stall value × 0.4. The mapped intervention corrected speed is in km / h, representing the target speed value that needs to be reduced through braking intervention. The system considers the dynamic characteristics of different time periods and applies time-weighted processing to the correction angle and correction speed. The weighting coefficient matrix is ​​a preset 8×2 matrix, where the first column is the angle weighting coefficient and the second column is the speed weighting coefficient. The angle weighting coefficient sequence is [0.4, 0.6, 0.8, 1.0, 0.9, 0.7, 0.5, 0.3], and the speed weighting coefficient sequence is [0.3, 0.5, 0.7, 0.9, 1.0, 0.8, 0.6, 0].[4] The system obtains the time-weighted correction angle and correction speed sequences through matrix multiplication. The system applies smoothing processing to the weighted sequences using the Savitzky-Golay filtering algorithm with a polynomial order of 3 and a window length of 5 to ensure the smoothness of the sequence changes. The system converts the processed correction angle and correction speed sequences into the standard format required by the control execution unit, including timestamps, absolute values ​​of correction amounts, and direction indicators. The angle accuracy is 0.1 degrees, and the speed accuracy is 0.1 km / h. The system dynamically adjusts the correction amount according to the current vehicle speed, road surface adhesion coefficient, and steering state. The adjustment coefficients are obtained by lookup table method. A three-dimensional lookup table with a size of 10×8×5 is established, corresponding to 10 vehicle speed ranges (0-150 km / h, intervals of 15 km / h), 8 adhesion coefficient ranges (0.1-0.9, intervals of 0.1), and 5 steering states (sharp steering, moderate steering, slight steering, straight driving, and reverse steering). The final generated direction offset intervention correction angle and intervention correction speed data are used as input parameters for subsequent jitter and yaw simulation. .

[0052] Based on the intervention correction angle and intervention correction speed, the system simulates and calculates the braking torque demand data for different time periods, including vehicle intervention correction, braking shudder, and yaw. This is achieved using high-precision vehicle dynamics simulation technology. The system first constructs a 14-DOF vehicle dynamics model, including 6 body degrees of freedom (longitudinal, lateral, vertical, roll, pitch, and yaw) and 8 wheel degrees of freedom (4 wheel rotations and 4 vertical displacements). The model parameters are determined based on the physical characteristics of a mid-size SUV electric vehicle, including a body mass of 1850kg, a wheelbase of 2.78m, front and rear track widths of 1.61m and 1.63m respectively, and suspension stiffness of 32000N / m for the front wheels and 35000N / m for the rear wheels. / m, suspension damping: front tire 2800 N·s / m, rear tire 3000 N·s / m. The tires use the Pacejka magic formula model with parameters B=10, C=1.9, D=1, E=0.97. The model calculation uses the fourth-order Runge-Kutta integral method with a step size of 1 ms to ensure calculation accuracy. The system converts the steering offset intervention correction angle into a steering system input, calculates the equivalent steering wheel angle using the steering ratio (16:1), and considers the elasticity and clearance characteristics of the steering system. The intervention correction vehicle speed is converted into a braking system input, and the equivalent braking pressure is calculated using a braking performance model (braking deceleration vs. braking pressure curve), with a pressure range of 0-12 MPa. A braking torque distribution model was constructed. Based on vehicle load distribution and tire adhesion characteristics, the braking torque distribution ratio of the four wheels was calculated. The front-to-rear axle distribution ratio ranged from 60:40 to 50:50, and the left-to-right wheel distribution ratio ranged from 45:55 to 55:45. Based on the braking torque demand data and the braking torque distribution ratio, the time series of the four-wheel braking torque was calculated. The braking torque range for each wheel was 0-1000 N·m. The system simulated the dynamic characteristics of the brake actuator, including execution delay (15-25 ms), pressure rise rate (5-10 MPa / s), and pressure fluctuation (±0.3 MPa). The braking actuator system was represented by a linear second-order model with a natural frequency of 15 Hz. With a Nibi of 0.7, the system simulates braking shudder and yaw at eight time intervals, calculating the vehicle's dynamic response in each time interval, including vehicle acceleration, wheel slip ratio, yaw rate, and sideslip angle. Simulation results are recorded at 10ms intervals to form detailed response time-series data. The system specifically analyzes the shudder phenomenon during braking, calculating the vehicle vibration caused by brake pressure fluctuations. The vibration frequency is mainly concentrated in the range of 8-15Hz, with an amplitude range of 0.05-0.2g. Simultaneously, the yaw response characteristics are analyzed, calculating the overshoot, rise time, and settling time of the yaw rate, with typical values ​​of 15-25%, 0.3-0.5 seconds, and 0.8-1 seconds, respectively.Within 2 seconds, the system comprehensively analyzes the vehicle body vibration and yaw response characteristics to construct a vibration and yaw evaluation index. This index comprises both time and frequency domain components. The time domain index includes the root mean square value, peak factor, and waveform factor; the frequency domain index includes power spectral density and band energy distribution. The system generates vibration and yaw data in a three-dimensional tensor format, with dimensions of [number of samples × number of time periods × number of features], i.e., [8745 × 8 × 24]. The features include 12 time-domain features and 12 frequency-domain features. The data resolution reaches 10ms in the time dimension, 0.01g and 0.01 degrees / second in the amplitude dimension, and 0.1Hz in the frequency dimension.

[0053] Multivariate regression analysis of jitter and skew data is performed using ensemble learning techniques. First, the jitter and skew data is preprocessed, including missing value imputation, outlier detection, and feature standardization. Missing value imputation uses the K-nearest neighbor imputation algorithm with K set to 5 and Mahalanobis distance as the distance metric. Outlier detection uses the local outlier factor algorithm with a nearest neighbor count of 20 and a threshold of 1.5. Standardization uses Min-Max normalization to map feature values ​​to the [0,1] interval. The system then performs feature selection on the preprocessed data using the recursive feature elimination (RFE) algorithm. The wrapper model uses support vector regression with a radial basis function kernel, parameters C=10, and γ=0.1. The selection process employed 5-fold cross-validation, ultimately retaining 12 of the most explanatory features, including key indicators such as the root mean square value of yaw rate, peak yaw acceleration, and the energy proportion of the yaw rate power spectrum in the 8-12Hz frequency band. The system constructed an integrated multivariate regression analysis model, incorporating five basic regression algorithms: Ridge Regression, Lasso Regression, Elastic Network Regression, Random Forest Regression, and Gradient Boosting Tree Regression. The Ridge Regression parameters were set with a regularization coefficient α=0.5, Lasso Regression parameters with a regularization coefficient α=0.01, Elastic Network parameters with α=0.5 and an L1 ratio of 0.5, Random Forest Regression parameters with 100 trees, a maximum depth of 12, and a minimum number of leaf node samples of 10, and Gradient Boosting Tree Regression... The parameters were set to 150 trees, a learning rate of 0.05, a maximum depth of 8, and a subsampling rate of 0.8. The system trained each basic model, with training data comprising 70% of the total data and test data comprising 30%. Five-fold cross-validation was used for training. Evaluation metrics included mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²). The performance of each model on the test set was as follows: Ridge Regression (MAE=2.85, RMSE=3.42, R²=0.82), Lasso Regression (MAE=2.92, RMSE=3.56, R²=0.81), Elastic Network (MAE=2.88, RMSE=3.48, R²=0.82), and Random Forest Regression (RFR). The system employs a stacked ensemble method to integrate the prediction results of various base models, including Ridge Regression (MAE=2.15, RMSE=2.75, R²=0.88) and Gradient Boosting Tree Regression (MAE=1.98, RMSE=2.58, R²=0.89). The secondary learner uses Linear Support Vector Regression with parameters C=1.0 and ε=0.1. The ensemble weights of each base model are determined through grid search optimization, with final weights of 0.15 for Ridge Regression, 0.1 for Lasso Regression, 0.15 for Elastic Network, 0.25 for Random Forest Regression, and 0.35 for Gradient Boosting Tree Regression. The final performance of the ensemble model on the test set reaches MAE=1.85, RMSE=2.42, and R²=0.91. The system uses a pre-trained ensemble regression model to perform full prediction on jitter and skew data, generating jitter and skew regression data. The data structure is a two-dimensional matrix with dimensions of [number of samples × number of time periods], i.e., [8745 × 8]. The matrix element values ​​represent the severity score of jitter and skew for each sample in each time period, with a score range of 0-100. Higher values ​​indicate more severe jitter and skew. The system performs reliability analysis on the regression data, calculating a 95% confidence interval. The average interval width is ±3.8, indicating that the prediction results have high stability and reliability.

[0054] Based on jitter and yaw regression data, progressive control of the direction deviation intervention correction angle and intervention correction speed at different time periods is implemented using fuzzy neural network technology. First, a jitter and yaw evaluation index system is constructed, classifying the jitter and yaw regression data into five severity levels: slight (0-20), moderate (20-40), noticeable (40-60), severe (60-80), and extreme (80-100). The system establishes a fuzzy inference system based on the jitter and yaw levels. Input variables are the current jitter and yaw level, the direction deviation intervention correction angle, and the intervention correction speed; output variables are the angle progressive adjustment coefficient and the speed progressive adjustment coefficient. Input fuzzification uses triangular and trapezoidal membership functions. The jitter and yaw levels are divided into five fuzzy sets, the correction angle into seven fuzzy sets, and the correction speed into five fuzzy sets. The output fuzzy set is the progressive adjustment coefficient, divided into seven levels. The fuzzy rule base contains 175 IF-THEN rules describing progressive control strategies under different conditions, such as "IF jitter and yaw is severe AND...". The system uses a large correction angle and a small progressive adjustment coefficient. The inference method employs the Mamdani inference mechanism, and the centroid method is used for defuzzification. The system optimizes the basic fuzzy system using a neural network, employing an Adaptive Neural Fuzzy Inference System (ANFIS). The network structure consists of five layers: a fuzzification layer, a rule layer, a normalization layer, a defuzzification layer, and an output layer. The hidden layer has 175 nodes, corresponding to 175 fuzzy rules. The training algorithm uses a hybrid learning method, combining least squares and gradient descent. The learning rate is set to 0.01, the momentum factor to 0.9, the training epochs to 500, and the error convergence threshold to 0.001. The optimized fuzzy neural network is applied to progressive control across eight time periods. For each time period, progressive adjustment coefficients are calculated for the directional offset correction angle and the intervention speed correction. The adjustment coefficient range is 0.6-1.2. When the jitter / sway score is high, the adjustment coefficient tends to decrease; conversely, it tends to decrease. To maintain the original correction amount, the system calculates the offset progressive correction angle based on the progressive adjustment coefficient. The calculation formula is: Progressive correction angle = Original correction angle × Angle progressive adjustment coefficient. Similarly, the progressive correction vehicle speed is calculated using the formula: Progressive correction vehicle speed = Original correction vehicle speed × Vehicle speed progressive adjustment coefficient. The system sets limitations for progressive correction, including limits on the rate of change of angle and the rate of change of vehicle speed. The upper limit for the rate of change of angle is 12 degrees / second, and the upper limit for the rate of change of vehicle speed is 5 km / h / second. When the calculation result exceeds the limit, it is truncated. The system smooths the progressive correction results using a cubic spline interpolation algorithm to ensure a smooth transition between adjacent time periods. The system generates offset progressive correction angle and progressive correction vehicle speed data in time series format, with a sampling interval of 10ms, covering a 2-second control window, forming a 200-point control sequence. The data accuracy is 0.1 degrees for angle and 0.1 km / h for vehicle speed.

[0055] The system optimizes braking force smoothing control across different time periods based on the progressive correction angle and speed of the offset using piecewise adaptive control technology. First, it establishes a mapping relationship between braking torque and offset correction amount using a piecewise linear mapping function. For the progressive correction angle, the mapping function is as follows: when the angle is less than 5 degrees, the braking torque increment = angle × 150 N·m / degree; when the angle is between 5 and 15 degrees, the braking torque increment = 750 + angle × 200 N·m / degree; when the angle is greater than 15 degrees, the braking torque increment... The braking torque increment = 2750 + angle × 50 N·m / degree. For progressive speed correction, the mapping function is as follows: when the speed is less than 10 km / h, the braking torque increment = speed × 100 N·m / (km / h); when the speed is between 10-30 km / h, the braking torque increment = 1000 + speed × 150 N·m / (km / h); when the speed is greater than 30 km / h, the braking torque increment = 4000 + speed × 80 N·m / (km / h). The system maps the angle and vehicle speed... The mapped braking torque increments are weighted and fused with weighting coefficients of 0.6 and 0.4 to calculate the basic braking torque increment. The system then performs time-time differential processing on the basic braking torque increment, constructing a time-weighted function with weighting coefficients of 0.3, 0.5, 0.7, 1.0, 0.9, 0.7, 0.5, and 0.3 for the eight time periods. The time-weighted braking torque increment is superimposed on the original braking torque demand data to obtain a preliminary braking torque control sequence. The system then performs smoothing control optimization on the braking torque control sequence, employing a multi-stage processing strategy. In the first stage, a low-pass filter is applied to eliminate high-frequency fluctuations. The filter type is a Butterworth fourth-order filter with a cutoff frequency of 5Hz. In the second stage, a slope limiter is applied to control the rate of change of braking torque, with the rising slope limited to 600 N·m / s and the falling slope limited to 800 N·m / s. In the third stage, a jitter suppressor is applied to handle minor fluctuations. The jitter suppression algorithm combines dead-zone control and exponential smoothing, with a dead-zone threshold of ±50 N·m and a smoothing factor of 0.85. The system performs segmented adjustments to the smoothed and optimized braking torque sequence, and performs targeted optimizations for the control characteristics of different time periods. The prediction segment (0-200ms) adopts a slight pre-braking strategy, limiting the braking torque to within 20% of the maximum demand. The initial response segment (200-400ms) adopts a linear growth strategy, with the braking torque increasing at 30% / 100ms of the maximum demand. Transition segment I (400-600ms) adopts a rapid growth strategy, with the braking torque increasing at 40% / 100ms of the maximum demand. The enhancement segment (600-900ms) adopts a stable control strategy, maintaining the braking torque at 90-100% of the maximum demand. The stable segment (900-1200ms) adopts a fine-tuning control strategy, limiting the braking torque variation range to within ±5%. Transition segment II (1200-1... The braking torque reduction strategy is employed in the first 500ms stage, with a slow reduction rate of 20% of the maximum demand per 100ms. In the weakening stage (1500-1800ms), a rapid reduction strategy is used, with a reduction rate of 30% of the maximum demand per 100ms. In the exit stage (1800-2000ms), a smooth exit strategy is adopted, reducing the braking torque to below 5% of the maximum demand. The system ultimately generates smooth braking force control data, with a multi-dimensional matrix structure containing the braking torque time series, braking force distribution parameters, and control timing markers. The torque time series sampling interval is 10ms, forming a 200-point control sequence. The torque range is 0-3500 N·m with an accuracy of 1 N·m. The torque distribution parameters include the front-to-rear axle distribution ratio and the left-to-right wheel distribution ratio. The control timing markers include the boundary times and key event time points for each stage.

[0056] The present invention also provides a new energy vehicle safety driving control system for executing the new energy vehicle safety driving control method described above, the new energy vehicle safety driving control system comprising: The data sampling module is used to obtain historical emergency avoidance datasets and corresponding user misoperation braking states through the new energy vehicle control center; and to perform sample undersampling processing on the historical emergency avoidance datasets to obtain emergency avoidance balance samples. The braking anomaly analysis module is used to analyze user misoperation braking anomalies in the emergency avoidance balance sample to obtain braking anomaly data; and to perform time-dimensional braking disorder vector intensity calculation based on the braking anomaly data to obtain disorder vector segmented clustering data. The intervention correction perception module is used to perform braking timing intervention correction perception based on disordered vector segmented clustering data to generate braking intervention correction optimization data; the braking intervention correction optimization data is sent to the new energy vehicle control terminal to execute new energy vehicle safe driving control.

[0057] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the 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. A method for safe driving control of new energy vehicles, characterized in that, Includes the following steps: Step S1: Obtain historical emergency avoidance dataset and corresponding user misoperation braking status through the new energy vehicle control center; perform sample undersampling processing on the historical emergency avoidance dataset to obtain emergency avoidance balanced samples; Step S2: Based on the emergency avoidance balance sample, perform user misoperation braking anomaly analysis on the avoidance scenario to obtain braking anomaly data; Based on the braking anomaly data, the disordered braking vector intensity is calculated in the time dimension to obtain disordered vector segmented clustering data. Step S3: Perform braking timing intervention correction sensing based on disordered vector segmented clustering data to generate braking intervention correction optimization data; send the braking intervention correction optimization data to the new energy vehicle control terminal to execute new energy vehicle safe driving control.

2. The new energy vehicle safe driving control method according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Obtain historical emergency avoidance dataset and corresponding user misoperation braking status through the new energy vehicle control center; Step S12: Randomly select samples from the historical emergency evacuation dataset to obtain random emergency evacuation samples; Step S13: Embed time tags into the emergency evacuation random samples to obtain evacuation time tag samples; Step S14: Perform sample undersampling on the risk avoidance time tag samples to obtain emergency risk avoidance balance samples.

3. The new energy vehicle safe driving control method according to claim 1, characterized in that, Step S2 includes the following steps: Step S21: Reconstruct the avoidance scenario based on the emergency avoidance balance sample; and simultaneously perform an association mapping on the misoperation braking state based on the avoidance scenario to obtain avoidance scenario-misoperation mapping data. Step S22: Perform user misoperation braking anomaly analysis on the avoidance scenario-misoperation mapping data to obtain braking anomaly data; Step S23: Calculate the disordered braking vector strength in the time dimension based on the braking anomaly data to obtain the disordered braking vector strength; Step S24: Perform temporal segmentation clustering on the intensity of the disordered vector under braking to obtain disordered vector segmentation clustering data.

4. The new energy vehicle safe driving control method according to claim 3, characterized in that, Step S22 includes the following steps: Extract the throttle acceleration and steering operation states of the user in the accident avoidance scenario-misoperation mapping data; When the acceleration under throttle is greater than 12 If the duration is greater than 1.5s and the accelerator pedal opening change is extracted by the engine control unit, and the accelerator pedal opening change increases from a rate of increase of more than 60% within 1s, it is determined to be an abnormal acceleration state due to accelerator pedal mis-pressing. If the steering angle is greater than or equal to 90°~180° within 1 second and the steering frequency is greater than or equal to 3 times within 1 second, it is judged as an abnormal steering operation due to steering error. Based on the abnormal acceleration state due to accidental accelerator pedal press and the abnormal steering due to accidental steering, the abnormal braking of the user's misoperation in the avoidance scenario is analyzed to obtain braking anomaly data.

5. The new energy vehicle safe driving control method according to claim 3, characterized in that, Step S23 includes the following steps: Step S231: Extract the abnormal acceleration changes due to throttle misapplication and the abnormal steering angle due to directional misoperation in the time dimension of the braking anomaly data; Step S232: Calculate the rate of increase of the abnormal acceleration change due to accidental throttle pressing, and then calculate the difference in the rate of increase over a 2-second time interval; Step S233: Analyze the sudden change in steering angle within 2 seconds due to directional error and abnormal steering angle. Step S234: Analyze the yaw impact vector intensity based on the speed increase ratio difference and the sudden change in steering angle to obtain the yaw impact vector intensity; Step S235: Perform nonlinear regression analysis on the yaw impact vector intensity to obtain the yaw impact vector regression intensity; calculate the braking disorder vector intensity in the time dimension based on the yaw impact vector regression intensity to obtain the braking disorder vector intensity.

6. The new energy vehicle safe driving control method according to claim 4, characterized in that, Step S234 includes the following steps: The yaw trajectory orientation is analyzed based on the sudden change in steering angle; nonlinear time-series integration is performed on the difference in acceleration ratio to obtain acceleration growth integral data; The dynamic offset angle is calculated by performing dynamic offset angle calculations on the sudden change amplitude of the steering angle and the rapid left and right steering. The rotational inertia fluctuation is estimated based on the integral data of acceleration growth rate and dynamic offset angle, and then the partial derivative of angular momentum is calculated to obtain the rotational angular momentum partial derivative data. Based on the acceleration growth integral data, dynamic offset angle and rotational angular momentum partial derivative data, the yaw trajectory orientation is subjected to oblique resultant vector strength coupling analysis to obtain oblique resultant vector strength coupling data; where the oblique resultant vector describes the oblique nature of the vehicle's motion or force in both longitudinal and lateral directions, and contains strength information in both directions. Based on the coupled data of the oblique combined vector strength, the yaw impact vector strength is analyzed to obtain the yaw impact vector strength.

7. The new energy vehicle safe driving control method according to claim 1, characterized in that, Step S3 includes the following steps: Step S31: Perform convolution processing on the unordered vector segmented clustering data to obtain unordered vector segmented convolution data; Step S32: Perform braking timing intervention correction sensing based on unordered vector segmented convolution data to obtain braking intervention correction data; Step S33: Iteratively learn the braking intervention correction data to generate optimized braking intervention correction data; Step S34: Send the braking intervention correction optimization data to the new energy vehicle control terminal to execute the new energy vehicle safe driving control.

8. The new energy vehicle safe driving control method according to claim 7, characterized in that, Step S32 includes the following steps: Step S321: Extract the segmented direction offset angle and vehicle speed stall value from different time periods in the unordered vector segmented convolutional data; Step S322: Quantify the vehicle body offset angle based on the segmented directional offset angle and vehicle speed stall value at different time periods to obtain the vehicle body offset angle; perform braking torque demand analysis based on the vehicle body offset angle to obtain braking torque demand data. Step S323: Based on the segmented directional offset angle and vehicle speed stall value in different time periods, optimize the braking torque demand data for braking force smoothing control in different time periods to obtain braking force smoothing control data. Step S324: Based on the braking force smooth control data, perform braking timing intervention correction sensing to obtain braking intervention correction data.

9. The new energy vehicle safe driving control method according to claim 8, characterized in that, Step S323 includes the following steps: Based on the segmented directional offset angles and vehicle speed stall values ​​for different time periods, the directional offset intervention correction angle and intervention correction speed for different time periods are determined. Based on the aforementioned directional offset intervention correction angle and intervention correction vehicle speed, the braking torque demand data for different time periods are simulated and calculated to obtain vehicle intervention correction braking shudder and yaw data. Perform multivariate regression analysis on the jitter and yaw data to generate jitter and yaw regression data; Based on the jitter and yaw regression data, the directional offset intervention correction angle and intervention correction speed for different time periods are progressively controlled to obtain the offset progressive correction angle and progressive correction speed. Braking force smoothing control is optimized for different time periods based on the offset progressive correction angle and progressive correction vehicle speed, resulting in braking force smoothing control data.

10. A safety driving control system for new energy vehicles, characterized in that, For executing the new energy vehicle safe driving control method as described in claim 1, the new energy vehicle safe driving control system includes: The data sampling module is used to obtain historical emergency avoidance datasets and corresponding user misoperation braking states through the new energy vehicle control center; and to perform sample undersampling processing on the historical emergency avoidance datasets to obtain emergency avoidance balance samples. The braking anomaly analysis module is used to analyze user misoperation braking anomalies in the emergency avoidance balance sample to obtain braking anomaly data; and to perform time-dimensional braking disorder vector intensity calculation based on the braking anomaly data to obtain disorder vector segmented clustering data. The intervention correction perception module is used to perform braking timing intervention correction perception based on disordered vector segmented clustering data to generate braking intervention correction optimization data; the braking intervention correction optimization data is sent to the new energy vehicle control terminal to execute new energy vehicle safe driving control.

Images