A vehicle lateral dynamics steady state identification method and related device

Abstract

The present application belongs to the field of vehicle active safety and intelligent driving technology, and discloses a vehicle lateral dynamics stable state identification method and related device. A Stacking integrated learning model is constructed with CART and SVM as base models and MLP as a meta model. A dataset based on sinusoidal sweep steering conditions and using a progressive maximum mutual information coefficient method for feature screening and state division is used for training. The final model has high precision identification and good generalization ability, and can realize quantitative grading discrimination of the vehicle stable state, thereby significantly improving the state identification reliability under complex dynamic conditions and providing a solid foundation for precise intervention of the control strategy.

Description

Technical Field

[0001] This invention belongs to the field of vehicle active safety and intelligent driving technology, and particularly relates to a method and related device for identifying the lateral dynamic stability state of a vehicle. Background Technology

[0002] With the rapid iteration and development of intelligent, electric, and high-speed automotive technologies, vehicle driving conditions are becoming increasingly complex, and the risk of lateral instability is continuously rising, highlighting the growing importance of vehicle lateral stability control technology. Lateral stability control, by monitoring and actively intervening in the vehicle's motion in real time, effectively avoids dangerous instability conditions such as skidding and fishtailing, and is a core technology for improving vehicle active safety performance and ensuring the stable and reliable operation of autonomous driving systems. Among these technologies, vehicle lateral dynamic stability state identification, as a core component of the stability state discrimination system, can identify instability risks during vehicle operation in real time, providing accurate data support and basis for vehicle control decisions. Accurate vehicle lateral dynamic stability state identification is a necessary prerequisite for achieving efficient and precise control of vehicle lateral stability, and is crucial for ensuring vehicle driving safety.

[0003] Current vehicle lateral stability determination technologies are mainly divided into two categories: those based on dynamic models and those based on data-driven approaches. However, existing state determination criteria also have many technical shortcomings, and the overall identification effect is insufficient to meet the application requirements of high precision and high adaptability. Traditional model-based identification methods, which initially relied on linear two-degree-of-freedom models combined with fixed thresholds to complete stability judgments, cannot adapt to the highly nonlinear and strongly coupled dynamic driving characteristics of vehicle dynamics, resulting in a significant decrease in identification accuracy under extreme conditions. Subsequent improvements such as the Lyapunov exponent and the self-adjusting phase plane method still heavily rely on accurate vehicle model parameters and can only perform binary state division of stable and unstable states, failing to quantify stability levels and hindering the precise intervention of control strategies. While data-driven identification methods do not rely on precise dynamic models and can adapt to the nonlinear characteristics of vehicles, existing technologies have significant drawbacks. First, the experimental conditions are limited and the data coverage is narrow, failing to fully recreate the complex dynamic behavior of vehicles after instability, resulting in insufficient reliability in state discrimination. Second, the selection of feature parameters is complex and highly subjective, with numerous redundant correlations between parameters, increasing the computational burden on equipment and easily leading to model bias, further reducing identification accuracy and reliability. At the state discrimination criterion level, distance-based discrimination methods are greatly affected by data quality and distribution characteristics, exhibiting poor adaptability and generalization ability in dynamic environments. Single neural network discrimination models are easily affected by their own structure, hyperparameters, and data quality, exhibiting inherent limitations in their ability to discriminate highly nonlinear dynamic systems. While existing stacked generalization ensemble learning models can integrate the advantages of multiple models and improve identification performance, their core weakness lies in the lack of a scientific base model combination selection mechanism. The randomness of base model configuration directly affects the meta-model fusion effect and the final identification performance, making it difficult to achieve the optimal performance of ensemble learning.

[0004] It is evident that existing vehicle lateral dynamic stability identification technologies generally suffer from low identification accuracy, poor generalization ability, and inability to achieve quantitative classification of stability levels, making them unsuitable for the precise identification requirements of vehicle lateral stability under complex working conditions. Summary of the Invention

[0005] This invention provides a method and related device for identifying the lateral dynamic stability state of a vehicle. This method can effectively solve the defects of existing vehicle lateral dynamic stability state identification technologies, such as low identification accuracy, poor generalization ability, and inability to achieve quantitative classification of stability degree. It can meet the needs of accurate identification of the lateral stability state of a vehicle under complex working conditions.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for identifying the lateral dynamic stability state of a vehicle includes: Collect the lateral dynamic characteristics of the vehicle to be identified; The lateral dynamic features of the vehicle to be identified are input into the pre-trained vehicle lateral dynamic stability state discrimination model, and the vehicle lateral dynamic stability state identification result is output. The vehicle lateral dynamic stability state discrimination model is based on a Stacking ensemble learning framework with CART and SVM as base models and MLP as meta-model. The vehicle lateral dynamic stability state discrimination model is trained on a training dataset based on the Stacking ensemble learning framework. The steps for constructing the training dataset are as follows: Based on the sinusoidal sweep frequency steering condition, the vehicle's lateral dynamic characteristic parameters are collected, and sample preprocessing is performed to remove duplicate samples to obtain the initial feature dataset. The progressive maximum mutual information coefficient method is used to sequentially perform feature redundancy removal, K-Means clustering state partitioning and labeling, and feature-label correlation filtering operations on the initial feature dataset to construct the training dataset.

[0007] Furthermore, the process of acquiring vehicle lateral dynamics characteristic parameters based on sinusoidal sweep frequency steering conditions, performing sample preprocessing operations, and removing duplicate samples to obtain an initial feature dataset includes: A vehicle model was built based on a vehicle dynamics simulation platform, and data was collected using a sinusoidal sweep frequency steering condition. The steering wheel angle was executed sequentially with decreasing amplitude sinusoidal steering in preset time periods, and the steering wheel angle was kept at zero in subsequent time periods. The simulation conditions covered a reasonable range of longitudinal vehicle speed and road surface adhesion coefficient and were divided according to preset step sizes to form several simulation conditions. Multiple characteristic parameters related to vehicle lateral dynamics are collected at a preset sampling frequency; after removing duplicate samples, a valid sample set is obtained, which constitutes the initial feature dataset; wherein, the multiple characteristic parameters include yaw rate, center of gravity sideslip angle, center of gravity sideslip rate, vehicle roll angle, vehicle roll rate, longitudinal vehicle speed, lateral vehicle speed, lateral acceleration, front axle lateral load transfer rate, rear axle lateral load transfer rate, front wheel steering angle, and road adhesion coefficient.

[0008] Furthermore, the execution feature redundancy removal includes: The maximum mutual information coefficient algorithm is used to calculate the correlation between features in the initial feature dataset. The process of gridding, normalization and selection of maximum mutual information is executed in sequence to construct the feature redundancy matrix. Based on the feature redundancy matrix, completely redundant features are removed to obtain the feature parameter set after redundancy removal.

[0009] Furthermore, the K-Means clustering state partitioning and labeling includes: The feature parameter set after redundancy removal is split into two feature parameter subsets corresponding to the forward and reverse steering angles of the steering wheel. Perform elbow rule analysis on the two subsets separately to determine the optimal number of clusters. The lateral dynamic stability state of vehicles is classified into four categories: stable, tending towards stability, tending towards instability, and unstable. K-Means clustering is performed based on preset clustering parameters to assign state classification labels to all samples; Merge the subsets with the same label within the positive and negative corner subsets to obtain the feature dataset with state labels.

[0010] Furthermore, the feature-label relevance screening includes: The maximum mutual information coefficient algorithm is used to calculate the correlation between each feature and the state label in the feature dataset with state labels, and the correlation score of each feature is obtained. Features with a relevance score lower than the first preset score and no substantial correlation with the label, as well as features with a relevance score lower than the second preset score and difficult to measure in real vehicles, are removed. After removing features, the final effective feature set is determined, and the training dataset is constructed.

[0011] Furthermore, the Stacking ensemble learning framework includes: The framework is set up with a two-layer structure: the first layer is the base model layer, and the second layer is the meta-model layer. The base model layer takes effective features from the training dataset as input and outputs predicted probabilities for four types of stable states. The meta-model layer takes as input the predicted probabilities from the base model layer and outputs as the classification results of the vehicle's lateral dynamic stability state.

[0012] Furthermore, the training process of the vehicle lateral dynamic stability state discrimination model is as follows: The training dataset is divided into a training subset and a test subset according to a preset ratio; A five-fold cross-validation method was used to pre-train multiple candidate base models, and candidate base models with satisfactory accuracy were selected based on the validation accuracy. Through ablation experiments, CART and SVM were identified as the optimal ensemble framework for Stacking ensemble learning, with MLP as the meta-model. The model training and validation are completed based on the training subset and test subset and the optimal ensemble framework, resulting in a well-trained vehicle lateral dynamic stability state discrimination model. The candidate base models include Classification and Regression Tree (CART), Multilayer Perceptron (MLP), Gaussian Naive Bayes, Support Vector Machine (SVM), k-Nearest Neighbors (k-NN), and Linear Logistic Regression (LRRP).

[0013] A vehicle lateral dynamic stability state identification system includes: The data acquisition module is used to collect the lateral dynamic characteristics of the vehicle to be identified; The state identification module is used to input the lateral dynamic features of the vehicle to be identified into the pre-trained vehicle lateral dynamic stability state identification model and output the vehicle lateral dynamic stability state identification result. The vehicle lateral dynamic stability state discrimination model is based on a Stacking ensemble learning framework with CART and SVM as base models and MLP as meta-model. The vehicle lateral dynamic stability state discrimination model is trained on a training dataset based on the Stacking ensemble learning framework. The specific steps for constructing the training dataset are as follows: Based on the sinusoidal sweep frequency steering condition, the vehicle's lateral dynamic characteristic parameters are collected, and sample preprocessing is performed to remove duplicate samples to obtain the initial feature dataset. The progressive maximum mutual information coefficient method is used to sequentially perform feature redundancy removal, K-Means clustering state partitioning and labeling, and feature-label correlation filtering operations on the initial feature dataset to construct the training dataset.

[0014] A vehicle lateral dynamic stability state identification device includes: Memory, used to store computer programs; A processor is used to implement the above-described vehicle lateral dynamic stability state identification method when executing the computer program.

[0015] A computer-readable storage medium storing a computer program, which, when executed by a processor, is used to implement the above-described vehicle lateral dynamic stability state identification method.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a method for identifying the lateral dynamic stability state of a vehicle. It constructs a Stacking ensemble learning model with CART and SVM as base models and MLP as the meta-model. The model is trained on a dataset based on sinusoidal sweep steering conditions and using a progressive maximum mutual information coefficient method for feature selection and state classification. The Stacking framework, by fusing the outputs of decision trees and support vector machines—two base models with vastly different principles—and deeply integrating them through a neural network meta-model, effectively captures the highly nonlinear and coupled characteristics of vehicle lateral dynamics. The progressive feature selection method scientifically eliminates redundant parameters and strengthens the correlation between features and state labels, ensuring high-quality and representative training data. This results in a final model that combines high-precision identification with good generalization ability, enabling quantitative and hierarchical discrimination of vehicle stability states. This significantly improves the reliability of state identification under complex dynamic conditions, providing a solid foundation for precise intervention of control strategies. Attached Figure Description

[0017] Figure 1 A flowchart illustrating the implementation of a vehicle lateral dynamic stability state identification method provided in this embodiment of the invention; Figure 2 The steering wheel angle curve for sinusoidal sweep frequency steering provided in the embodiments of the present invention; Figure 3 The MIC feature redundancy matrix is ​​the initial feature parameter provided in the embodiments of the present invention; Figure 4 The elbow rule analysis results provided for the embodiments of the present invention; wherein, (a) is a rotation angle greater than 0; (b) is a rotation angle less than 0; Figure 5 The three-dimensional scatter plot and feature maps of the K-means clustering results provided in this embodiment of the invention are shown; wherein, (a) is a three-dimensional clustering scatter plot; and (b) is a two-dimensional projection scatter plot. Figure 6 The MIC importance score statistics of the classification labels provided in the embodiments of the present invention; wherein, (a) is a horizontal bar chart of feature importance ranking based on MIC; and (b) is a vertical bar chart of MIC scores of all features. Figure 7 The optimized CART-SVM-MLP Stacking integrated model structure provided in this embodiment of the invention; Figure 8 This is a schematic diagram illustrating the performance of the Stacking model provided in an embodiment of the present invention; wherein, (a) is a visualization of the comparison of prediction results on the training set; (b) is a visualization of the comparison of prediction results on the test set; (c) is the confusion matrix of the training set; and (d) is the confusion matrix of the test set. Figure 9This is a core flowchart of a vehicle lateral dynamic stability state identification method provided in an embodiment of the present invention; Figure 10 This is a schematic diagram of a vehicle lateral dynamic stability state identification system provided in an embodiment of the present invention. Detailed Implementation

[0018] To further understand the content of this invention, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments are merely illustrative and not limiting of the invention.

[0019] As described in the background section, data-driven methods do not rely on precise dynamic models and are better suited to the nonlinear characteristics of vehicles. However, existing research has significant drawbacks: First, the experimental conditions are relatively simple, resulting in insufficient data coverage and making it difficult to fully reflect the complex dynamic behavior of vehicles after instability, thus limiting the reliability of state identification. Second, too many feature parameters are selected, ignoring the redundancy between parameters. Furthermore, parameter selection is highly subjective, and highly correlated redundant features not only increase the computational burden but also cause model bias, reducing the reliability of identification. Therefore, implementing correlation analysis is necessary. The MIC algorithm can effectively capture the complex nonlinear relationships between vehicle dynamic features and is suitable for redundancy identification and dependency mining in high-dimensional parameters.

[0020] In constructing state discrimination criteria, existing research can be mainly divided into two categories: distance-based and data-driven. Distance-based methods heavily rely on data quality and distribution characteristics for classification performance, exhibiting poor adaptability to dynamic environments and limited generalization ability. While neural network methods perform well in stability discrimination, their performance is significantly affected by model structure, hyperparameters, and data quality; a single model still has limitations in discriminating highly nonlinear dynamic systems. Stacking generalization methods, by integrating the prediction information of multiple base models to construct a meta-model, effectively combine the advantages of different models, enhancing generalization ability and prediction accuracy. However, the performance of stacking models is highly dependent on the selection of base model combinations—different base models constitute different knowledge representations, directly affecting the fusion effect and final performance of the meta-model. Therefore, scientifically selecting and determining the optimal base model combination becomes the core challenge in constructing high-performance stacking recognition models.

[0021] To address the aforementioned issues, this embodiment provides a method for identifying the lateral dynamic stability state of a vehicle. This method solves the problems of low identification accuracy, poor generalization, and inability to quantify and classify in the prior art. It constructs a high-quality training dataset through a progressive MIC method and combines it with the Stacking ensemble learning framework to achieve high-precision and high-generalization identification of the lateral stability state of a vehicle.

[0022] For example, such asFigure 9 As shown, this embodiment provides a method for identifying the lateral dynamic stability state of a vehicle, including: Collect the lateral dynamic characteristics of the vehicle to be identified; The lateral dynamic features of the vehicle to be identified are input into the pre-trained vehicle lateral dynamic stability state discrimination model, and the vehicle lateral dynamic stability state identification result is output. The vehicle lateral dynamic stability state discrimination model is based on a Stacking ensemble learning framework with CART and SVM as base models and MLP as meta-model. The vehicle lateral dynamic stability state discrimination model is trained on a training dataset based on the Stacking ensemble learning framework. The steps for constructing the training dataset are as follows: Based on the sinusoidal sweep frequency steering condition, the vehicle's lateral dynamic characteristic parameters are collected, and sample preprocessing is performed to remove duplicate samples to obtain the initial feature dataset. The progressive maximum mutual information coefficient method is used to sequentially perform feature redundancy removal, K-Means clustering state partitioning and labeling, and feature-label correlation filtering operations on the initial feature dataset to construct the training dataset.

[0023] The identification method provided in this embodiment will be further explained below with reference to the accompanying drawings: For example, this embodiment provides a method for identifying the lateral dynamic stability state of a vehicle, including two main processes: constructing and training a vehicle lateral dynamic stability state identification model and using the vehicle lateral dynamic stability state identification model for state identification. The specific steps for constructing and training the vehicle lateral dynamic stability state identification model are as follows: S1: Collect vehicle lateral dynamic characteristic parameters based on sinusoidal sweep frequency steering conditions, complete sample preprocessing, and obtain the initial feature dataset.

[0024] A vehicle model was built based on the Carsim simulation platform. Data acquisition was completed using a sinusoidal sweep frequency steering condition. The steering wheel angle input rules were as follows: the vehicle completed sinusoidal steering with amplitudes of 180°, 90°, and 45° in 0s-8s, 8s-12s, and 12s-14s respectively, and kept the steering wheel angle at 0 in 14s-20s. The simulation conditions cover longitudinal vehicle speeds of 30-120 km / h with a step size of 10 km / h; road surface adhesion coefficients of 0.4-1 with a step size of 0.1; a total of 70 simulation conditions are included, with a sampling frequency of 10 Hz. Collect yaw rate , centroid side slip angle angular velocity of center of mass deflection Body roll angle Body roll rate Longitudinal speed Lateral speed lateral acceleration , , Front wheel steering angle Road surface adhesion coefficient There are 12 parameters, among which, The output matrix can be represented as:

[0025] After removing duplicate samples, 11,989 valid samples were obtained, which constituted the initial feature dataset.

[0026] S2: Based on the initial feature dataset, a high-quality model training dataset is constructed using the progressive maximum mutual information coefficient (MIC) method.

[0027] Phase 1: Feature Redundancy Analysis. The MIC algorithm is used to calculate the correlation between features in the initial feature set, construct a feature redundancy matrix, and eliminate completely redundant features to obtain a preliminary feature set. Specifically, this includes: The correlation among the 12 initial features is calculated using the Maximum Mutual Information Coefficient (MIC) algorithm. The MIC calculation steps are as follows: (1) Gridding: Randomly select two features from the initial feature set. X and Y The two-dimensional scatter plot composed of the two features is divided along the coordinate axis direction into... i List j The grid of rows, calculate the values ​​under that grid. X and Y The mutual information matrix is ​​expressed as:

[0028] In the formula, is for and The joint probability density, and They are respectively and The marginal probability density.

[0029] (2) Normalization: The maximum mutual information value in the mutual information matrix is ​​normalized. The expression is:

[0030] In the formula, Indicates size is The set of grid partitions.

[0031] (3) Select the maximum mutual information value under different grid sizes as the final MIC value, and the expression is:

[0032] In the formula, It is the maximum possible number of mesh divisions, usually set to , This represents the sample size.

[0033] Based on the above steps, a MIC feature redundancy matrix is ​​constructed, and lateral vehicle speeds and lateral accelerations that are completely redundant with the centroid sideslip angles are eliminated. The final feature parameter set is obtained. .

[0034] The second stage involves using the K-Means clustering algorithm to partition the feature parameter set into stable states, specifically including: The K-Means clustering algorithm is used to analyze the feature parameter set. Perform steady state division Finally, the state label of each sample is obtained. Training dataset for constructing a vehicle lateral dynamic stability state identification model The initial feature set is split into feature parameter subsets where the steering wheel angle is greater than 0 and less than 0. , Elbow rule analysis is used to classify the lateral dynamic stability state of the vehicle into four categories: stable state, tending towards stable state, tending towards unstable state, and unstable state.

[0035] The K-means clustering algorithm parameters are set as follows: number of clusters 7, maximum number of iterations 100. After clustering, the corresponding classification labels are assigned to each of the 11989 samples, and the subsets with the same labels within the positive and negative corner subsets are merged.

[0036] Phase 3: Feature-label correlation analysis, specifically including: The MIC algorithm is used to calculate the correlation between each feature in the initial feature set and the classification label, and the MIC importance score of each feature is obtained. The road adhesion coefficient with a MIC score of 0.002 and no correlation with the label is removed, as well as the centroid sideslip angular velocity and vehicle body roll angular velocity with a MIC score of less than 0.5, which are difficult to measure accurately in real vehicles.

[0037] Final choice The feature parameters, the final dataset can be represented as:

[0038] S3: Based on the training dataset, construct a Stacking ensemble learning identification model, determine the optimal ensemble framework with CART and SVM as base models and MLP as meta-model through ablation experiments, and complete model training and validation.

[0039] The Stacking ensemble learning identification model adopts a two-layer structure: the first layer is the base model layer, and the second layer is the meta-model layer. The base model layer takes into account the effective features of the training dataset and outputs the predicted probabilities of four stable state levels. The meta-model layer takes into account the output of the base model and outputs the final stable state classification result.

[0040] Five-fold cross-validation was used to pre-train the candidate base models. A model dataset was constructed based on effective features and divided into training and test sets in an 8:2 ratio. The candidate base models included CART (Classification and Regression Tree), Gaussian Naive Bayes, Support Vector Machine (SVM), k-Nearest Neighbors (k-NN), and Linear Logistic Regression. Model scores were used for evaluation. Verification accuracy can be expressed as:

[0041] In the formula, This represents the number of models trained based on different hyperparameters.

[0042] CART, Gaussian Naive Bayes, SVM, and k-NN, which have a validation accuracy of over 85%, were initially selected as candidate base models.

[0043] Explained, a Multilayer Perceptron (MLP) is a classic feedforward artificial neural network that can capture complex nonlinear relationships in input data through hidden layers and activation functions, automatically learning high-order features of the input data. It has the advantages of simple structure and high computational efficiency. Furthermore, MLPs make no structural assumptions about the input data, making them more suitable for the dataset established above. Therefore, the meta-model adopts a Multilayer Perceptron (MLP).

[0044] Ablation experiments were conducted using a stepwise removal method, with accuracy as the primary evaluation metric and training time and single-step computation time as the evaluation standards. The base model outputs the probabilities of four classes as input to the meta-model. The hidden layer of the MLP meta-model consists of two fully connected layers with 64 and 32 neurons respectively, employing ReLU activation and batch normalization. The output layer uses the softmax activation function to transform the original output of the neural network into a probability distribution, with a dimension of 4. The maximum number of training epochs was 300, with an initial learning rate of 0.002 and a learning rate decay factor of 0.5. Early stopping and regularization techniques were implemented to prevent overfitting, with a regularization coefficient of 0.001. Considering the variability in model performance across training iterations, each model was trained five times, and the optimal result was displayed, with the average single-step computation time taken as the benchmark. Based on the ablation experiment results, the final base model combination was determined to be CART and SVM. The trained model achieved a validation accuracy of 99.4%, with recall and precision exceeding 99% for all four classes.

[0045] Therefore, the design idea of ​​this embodiment is to use the progressive MIC feature selection method to complete the construction of a high-quality dataset, to achieve the four-level quantization and classification of vehicle stability through K-means clustering, to build a Stacking ensemble learning framework to complete the construction of a high-precision identification model, and finally to verify the classification performance and generalization ability of the model through performance verification.

[0046] For example, the identification method provided in this embodiment has been specifically applied and implemented, and the implementation process is as follows: In this embodiment, considering that the unstable state of the vehicle during the data acquisition process may endanger the driver's safety, the characteristic parameters are collected by the vehicle simulation software Carsim. The vehicle parameters are shown in Table 1.

[0047] Table 1 shows the vehicle parameters.

[0048] This embodiment includes three parts, S1 to S3, as follows: Figure 1 As shown, the specific steps include: Step 1: S1, Feature parameter acquisition: Based on the Carsim simulation platform, data acquisition was completed using a sinusoidal sweep frequency steering condition. This condition, referencing ISO international standards, effectively avoids the limitations of insufficient diversity in the widely used oblique step steering condition in existing research. The steering wheel angle input curve is shown below. Figure 2As shown. The specific operating conditions are set as follows: the vehicle completes a sinusoidal steering with an amplitude of 180° within 0s-8s, a sinusoidal steering with an amplitude of 90° within 8s-12s, a sinusoidal steering with an amplitude of 45° within 12s-14s, and maintains a steering wheel angle of 0 for 14s-20s. The longitudinal vehicle speed is 30~120km / h with a step size of 10km / h, and the road surface adhesion coefficient is 0.4~1 with a step size of 0.1; a total of 70 simulation operating conditions are set, and the sampling frequency is set to 10Hz.

[0049] Collect yaw rate , centroid side slip angle angular velocity of center of mass deflection Body roll angle Body roll rate Longitudinal speed Lateral speed lateral acceleration , , Front wheel steering angle Road surface adhesion coefficient There are 12 parameters, among which, The output matrix can be represented as:

[0050] The collected data from 70 operating conditions were preprocessed to remove duplicate samples, resulting in 11,989 valid samples.

[0051] Step 2: S2, use the MIC algorithm to calculate the correlation between the 12 initial features. This algorithm can capture any form of linear, nonlinear, and non-monotonic correlation between features, and is more suitable for vehicle dynamics feature analysis than the traditional correlation coefficient. The specific calculation steps are as follows: (1) Gridding: Randomly select two features from the initial feature set. X and Y The two-dimensional scatter plot composed of the two features is divided along the coordinate axis direction into... i List j The grid of rows, calculate the values ​​under that grid. X and Y The mutual information matrix is ​​expressed as:

[0052] In the formula, is for and The joint probability density, and They are respectively and The marginal probability density.

[0053] (2) Normalization: The maximum mutual information value in the mutual information matrix is ​​normalized to eliminate the influence of grid size on the result. The expression is:

[0054] In the formula, Indicates size is The set of grid partitions.

[0055] (3) MIC value calculation: The maximum mutual information value under different grid sizes is selected as the final MIC value, and the expression is:

[0056] In the formula, It is the maximum possible number of mesh divisions, usually set to , This represents the sample size.

[0057] Based on the above steps, a MIC feature redundancy matrix of 12 initial features is constructed, and the result is as follows. Figure 3 As shown. The results indicate that the MIC value of the sideslip angle and lateral velocity is 1, and the lateral acceleration is... The MIC value is 1, indicating complete redundancy. To reduce dataset dimensionality and computational burden while retaining the ability to represent lateral stability, lateral vehicle speeds and lateral accelerations that are completely redundant with the centroid sideslip angle are removed. The final feature parameter set is obtained. .

[0058] Step 3: Steady-state partitioning based on K-means clustering: The K-Means clustering algorithm is used to analyze the feature parameter set. Perform steady state division Finally, the state label of each sample is obtained. Training dataset for constructing a vehicle lateral dynamic stability state identification model .

[0059] Considering that the dataset is distributed on both sides of the coordinate axis under positive and negative steering wheel angles, direct clustering would prevent samples with the same stable state from being classified into the same class. Therefore, feature parameter subsets with steering wheel angles greater than 0 and less than 0 are clustered separately. , Elbow rule analysis was performed to determine the optimal number of clusters for K-Means. The elbow analysis results are as follows: Figure 4 As shown. The slope changes under both positive and negative turning angles. The differences are quite obvious. Therefore, the lateral dynamic stability state of the vehicle is divided into four categories: stable state, approaching stable state, approaching unstable state, and unstable state. Combining the sample classification labels and data distribution, and The subsets of data with stable state labels are adjacent. After multiple experiments, the final number of clusters was set to 7, and the maximum number of iterations was 100. Finally, and Subdata sets with the same label were merged, and principal component analysis was used for dimensionality reduction visualization. The clustering results are shown in Figure 5.

[0060] Step 4: Perform feature-label correlation analysis using the MIC algorithm to eliminate features with low correlation, providing a high-quality data foundation for subsequent model training. The correlation analysis results are as follows: Figure 6 As shown, regardless of whether the road surface conditions are high or low adhesion, vehicles traveling at high speeds are in an unstable state under large steering conditions. Therefore, the MIC score of the road adhesion coefficient is the lowest, only 0.002, which is close to no correlation. Considering the limitations of sensor technology, environmental interference, and high cost, the center of gravity sideslip angular velocity and the vehicle roll angular velocity are difficult to measure on real vehicles, which not only increases the data acquisition burden but also reduces the practicality of the model. Moreover, the MIC scores of both are less than 0.5. Therefore, we choose... The feature parameters, the final dataset can be represented as:

[0061] Step 5: S3, the Stacking ensemble learning model built in this embodiment adopts... Folded cross-validation is used to train multiple base models, and a meta-model is used to learn the complex mapping relationship between the prediction results of the base models and the true labels, achieving synergistic optimization of model performance and complex data processing capabilities. Compared with traditional single models, stacked generalization models can integrate the advantages of multiple heterogeneous base models and effectively reduce the risk of overfitting, thereby improving model accuracy and generalization ability.

[0062] The selection of base models must meet the core requirements of accuracy and diversity to ensure the learning quality and ensemble effect of the meta-model. First, five classic supervised learning algorithms were selected as candidate base models: Classification and Regression Tree (CART), Gaussian Naive Bayes, Support Vector Machine (SVM), k-Nearest Neighbors (k-NN), and Linear Logistic Regression.

[0063] The candidate base model was trained using a 5-fold cross-validation method, and the dataset was divided into training and training sets in an 8:2 ratio. and test set and the training set The data was split into 5 subsets, each used alternately as the validation set, and the other 4 subsets used as the training set. The model score was denoted as... Finally, the model with the best mean score is selected to calculate its performance on the test set, which can be expressed as:

[0064] in, This represents the number of models trained based on different hyperparameters.

[0065] The training results are shown in Table 2.

[0066] Table 2 Comparison of Base Model Performance

[0067] The results showed that the SVM model performed best, with a validation accuracy of 98.7%. CART, k-NN, and Gaussian Naive Bayes all achieved validation accuracies above 85%. Linear logistic regression performed the worst and was therefore eliminated. Ultimately, CART, Gaussian Naive Bayes, SVM, and k-NN were initially selected as candidate base models.

[0068] The core of a meta-model is learning the complex mapping relationship between the prediction results of the base model and the true labels. To avoid overfitting, an overly complex structure should be avoided. This embodiment uses an MLP as the meta-model, which can capture complex nonlinear relationships in the input data through hidden layers and activation functions, automatically learn the high-order features of the input data, and has the advantages of simple structure and high computational efficiency. At the same time, MLPs do not make structural assumptions about the input data, making them more suitable for the dataset established above.

[0069] Ablation experiments were conducted using a stepwise removal method, with accuracy as the primary evaluation metric and training time and single-step computation time as the evaluation standards. The base models output four class probabilities as input to the meta-model. The hidden layers of the MLP meta-model consisted of two fully connected layers with 64 and 32 neurons respectively, employing ReLU activation and batch normalization. The output layer used a softmax activation function to transform the original output of the neural network into a probability distribution, with a dimension of 4. The maximum number of training epochs was 300, with an initial learning rate of 0.002 and a learning rate decay factor of 0.5. Early stopping and regularization techniques were implemented to prevent overfitting, with a regularization coefficient of 0.001. Considering the differences in model performance across training iterations, each model was trained five times, and its optimal results were displayed. The average single-step computation time was used. The ablation experiment results are shown in Table 3. A, B, C, and D correspond to the four base models: CART, SVM, k-NN, and Gaussian Naive Bayes, respectively.

[0070] Table 3 shows the comparison of ablation test results.

[0071] Experimental Results Analysis: After removing k-NN (C) and Gaussian Naive Bayes (D), the validation accuracy decreased by only 0.05% compared to the complete model, but the single-step computation time decreased to 0.31ms, and the training time was reduced by 7s, achieving optimal overall performance while maintaining high accuracy. If CART or SVM models were removed, the validation accuracy decreased by 3.17% and 0.38% respectively, showing a significant deterioration in model performance, indicating that these two models play a crucial role in overall performance. In summary, the optimal base model combination was determined to be CART and SVM. These two models possess excellent nonlinear fitting ability and small-sample generalization ability, respectively, and are highly complementary, providing high-quality input for the meta-model. The final Stacking ensemble learning structure was determined to be CART-SVM-MLP, as shown below. Figure 7 As shown.

[0072] Step 6: Performance Verification of the Stacking Ensemble Learning Model: The performance of the Stacking model is shown in Figure 8. Figures 8(a) and (b) compare the training results of the training and test sets, respectively. Labels 1-4 in the figures correspond to the stable state, the state approaching stability, the state approaching instability, and the unstable state, respectively. The confusion matrix is ​​shown in Figures 8(c) and (d). The recall and precision of the model for the four categories are both above 99%, further demonstrating that the Stacking model designed in this invention has excellent classification performance and generalization ability.

[0073] To highlight the superior performance of the proposed Stacking model, the performance of the Cart, SVM, and MLP models was compared. Notably, no feature-label correlation analysis was performed on the MLP2 model dataset. The model comparison results are shown in Table 4. The validation accuracy of the Stacking model is 5%, 0.7%, and 0.7% higher than that of the Cart, SVM, and MLP1 models, respectively, outperforming the performance of a single model consisting of a base model and a meta-model. The prediction and validation accuracy of the MLP1 model are 0.4% and 0.3% lower than that of the MLP2 model, respectively. This is because features with a correlation of less than 0.5 with the classification label were removed, reducing the model's performance. The training accuracy of the Stacking and MLP2 models differs by only 0.1%, and the validation accuracy of the Stacking model is 0.4% higher than that of the MLP2 model. This effectively prevents the loss of model accuracy due to feature selection and also demonstrates that the proposed model has stronger generalization ability.

[0074] Table 4 shows the model comparison results.

[0075] For example, such as Figure 10As shown, this embodiment also provides a vehicle lateral dynamic stability state identification system, including: a data acquisition module for acquiring the lateral dynamic characteristics of the vehicle to be identified; The state identification module is used to input the lateral dynamic features of the vehicle to be identified into the pre-trained vehicle lateral dynamic stability state identification model and output the vehicle lateral dynamic stability state identification result. The vehicle lateral dynamic stability state discrimination model is based on a Stacking ensemble learning framework with CART and SVM as base models and MLP as meta-model. The vehicle lateral dynamic stability state discrimination model is trained on a training dataset based on the Stacking ensemble learning framework. The specific steps for constructing the training dataset are as follows: Based on the sinusoidal sweep frequency steering condition, the vehicle's lateral dynamic characteristic parameters are collected, and sample preprocessing is performed to remove duplicate samples to obtain the initial feature dataset. The progressive maximum mutual information coefficient method is used to sequentially perform feature redundancy removal, K-Means clustering state partitioning and labeling, and feature-label correlation filtering operations on the initial feature dataset to construct the training dataset.

[0076] The present invention also provides a vehicle lateral dynamic stability state identification device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the vehicle lateral dynamic stability state identification method.

[0077] The present invention also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the vehicle lateral dynamic stability state identification method.

[0078] When the processor executes the computer program, it implements the above-mentioned steps for identifying the lateral dynamic stability state of the vehicle, for example: collecting the lateral dynamic features of the vehicle to be identified; The lateral dynamic features of the vehicle to be identified are input into the pre-trained vehicle lateral dynamic stability state discrimination model, and the vehicle lateral dynamic stability state identification result is output. The vehicle lateral dynamic stability state discrimination model is based on a Stacking ensemble learning framework with CART and SVM as base models and MLP as meta-model. The vehicle lateral dynamic stability state discrimination model is trained on a training dataset based on the Stacking ensemble learning framework. The steps for constructing the training dataset are as follows: Based on the sinusoidal sweep frequency steering condition, the vehicle's lateral dynamic characteristic parameters are collected, and sample preprocessing is performed to remove duplicate samples to obtain the initial feature dataset. The progressive maximum mutual information coefficient method is used to sequentially perform feature redundancy removal, K-Means clustering state partitioning and labeling, and feature-label correlation filtering operations on the initial feature dataset to construct the training dataset.

[0079] For example, the computer program can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing preset functions, wherein the instruction segments describe the execution process of the computer program in the vehicle lateral dynamic stability identification device. For example, the computer program can be divided into a data acquisition module and a state identification module; the specific functions are as follows: the data acquisition module is used to acquire the lateral dynamic features of the vehicle to be identified; the state identification module is used to input the lateral dynamic features of the vehicle to be identified into a pre-trained vehicle lateral dynamic stability state identification model and output the vehicle lateral dynamic stability state identification result; wherein, the basic model of the vehicle lateral dynamic stability state identification model is a Stacking ensemble learning framework with CART and SVM as base models and MLP as meta-model, and the vehicle lateral dynamic stability state identification model is trained on the training dataset based on the Stacking ensemble learning framework; the specific implementation steps of constructing the training dataset are as follows: acquire vehicle lateral dynamic feature parameters based on sinusoidal sweep frequency steering conditions, perform sample preprocessing operations, remove duplicate samples to obtain an initial feature dataset; use the progressive maximum mutual information coefficient method to sequentially perform feature redundancy removal, K-Means clustering state division and labeling, and feature-label correlation filtering operations on the initial feature dataset to construct the training dataset.

[0080] The vehicle lateral dynamics stability identification device can be a computing device such as a desktop computer, laptop, handheld computer, or cloud server. The device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above are examples of vehicle lateral dynamics stability identification devices and do not constitute a limitation on such devices. The device may include more components than described above, or combine certain components, or use different components. For example, the vehicle lateral dynamics stability identification device may also include input / output devices, network access devices, buses, etc.

[0081] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor, or any conventional processor. The processor is the control center for the vehicle lateral dynamics stability state identification, connecting various parts of the entire vehicle lateral dynamics stability state identification device through various interfaces and lines.

[0082] The memory can be used to store the computer program and / or modules. The processor realizes various functions of the vehicle lateral dynamic stability identification device by running or executing the computer program and / or modules stored in the memory and calling the data stored in the memory.

[0083] The memory may primarily include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a function (such as sound playback, image playback, etc.). The data storage area may store data created based on the use of the mobile phone (such as audio data, phonebook, etc.). Furthermore, the memory may include high-speed random access memory and non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital cards (SD cards), flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0084] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the vehicle lateral dynamic stability state identification method.

[0085] If the modules / units integrated in the vehicle lateral dynamic stability state identification system are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0086] Based on this understanding, the present invention can implement all or part of the processes in the above-described vehicle lateral dynamic stability state identification method, or it can be accomplished by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the above-described vehicle lateral dynamic stability state identification method. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or a preset intermediate form, etc.

[0087] The computer-readable storage medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0088] It should be noted that the content contained in the computer-readable storage medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable storage medium does not include electrical carrier signals and telecommunication signals.

[0089] Compared with existing identification methods, the present invention provides a method and related device for identifying the lateral dynamic stability state of a vehicle, which has the following advantages: This invention first collects lateral dynamics feature parameters of vehicles at different speeds and road surface adhesion coefficients based on sinusoidal frequency sweep steering conditions. Second, it proposes a progressive maximum mutual information coefficient (MIC) dataset construction method. In the first stage, the MIC algorithm is used to perform correlation analysis on the feature parameters and eliminate redundant features. In the second stage, K-means clustering is combined to complete the vehicle stability state classification and generate classification labels. In the third stage, the MIC algorithm is used to complete feature-label correlation analysis, eliminate features with low correlation, and construct a high-quality training dataset. Finally, through ablation experiments, the optimal ensemble framework with Classification Regression Tree (CART) and Support Vector Machine (SVM) as base models and Multilayer Perceptron (MLP) as meta-model is determined. A vehicle lateral dynamics stability state identification model based on stacking ensemble learning is constructed. Compared with single models and traditional phase plane methods, this invention significantly improves identification accuracy, predictability, and generalization ability, providing accurate and predictive decision-making basis for vehicle lateral stability control and effectively improving the active safety performance of vehicles under extreme conditions.

[0090] The above embodiments are merely one of the implementation methods for achieving the technical solution of the present invention. The scope of protection claimed by the present invention is not limited to this embodiment, but also includes any variations, substitutions and other implementation methods that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention.

[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for identifying the lateral dynamic stability state of a vehicle, characterized in that, include: Collect the lateral dynamic characteristics of the vehicle to be identified; The lateral dynamic features of the vehicle to be identified are input into the pre-trained vehicle lateral dynamic stability state discrimination model, and the vehicle lateral dynamic stability state identification result is output. The vehicle lateral dynamic stability state discrimination model is based on a Stacking ensemble learning framework with CART and SVM as base models and MLP as meta-model. The vehicle lateral dynamic stability state discrimination model is trained on a training dataset based on the Stacking ensemble learning framework. The steps for constructing the training dataset are as follows: Based on the sinusoidal sweep frequency steering condition, the vehicle's lateral dynamic characteristic parameters are collected, and sample preprocessing is performed to remove duplicate samples to obtain the initial feature dataset. The progressive maximum mutual information coefficient method is used to sequentially perform feature redundancy removal, K-Means clustering state partitioning and labeling, and feature-label correlation filtering operations on the initial feature dataset to construct the training dataset.

2. The method for identifying the lateral dynamic stability state of a vehicle according to claim 1, characterized in that, The method involves collecting vehicle lateral dynamics characteristic parameters based on sinusoidal frequency sweep steering conditions, performing sample preprocessing, and removing duplicate samples to obtain an initial feature dataset, including: A vehicle model was built based on a vehicle dynamics simulation platform, and data was collected using a sinusoidal sweep frequency steering condition. The steering wheel angle was executed sequentially with decreasing amplitude sinusoidal steering in preset time periods, and the steering wheel angle was kept at zero in subsequent time periods. The simulation conditions covered a reasonable range of longitudinal vehicle speed and road surface adhesion coefficient and were divided according to preset step sizes to form several simulation conditions. Multiple characteristic parameters related to vehicle lateral dynamics are collected at a preset sampling frequency; after removing duplicate samples, a valid sample set is obtained, which constitutes the initial feature dataset; wherein, the multiple characteristic parameters include yaw rate, center of gravity sideslip angle, center of gravity sideslip rate, vehicle roll angle, vehicle roll rate, longitudinal vehicle speed, lateral vehicle speed, lateral acceleration, front axle lateral load transfer rate, rear axle lateral load transfer rate, front wheel steering angle, and road adhesion coefficient.

3. The method for identifying the lateral dynamic stability state of a vehicle according to claim 1, characterized in that, The execution feature redundancy removal includes: The maximum mutual information coefficient algorithm is used to calculate the correlation between features in the initial feature dataset. The process of gridding, normalization and selection of maximum mutual information is executed in sequence to construct the feature redundancy matrix. Based on the feature redundancy matrix, completely redundant features are removed to obtain the feature parameter set after redundancy removal.

4. The method for identifying the lateral dynamic stability state of a vehicle according to claim 3, characterized in that, The K-Means clustering state partitioning and labeling includes: The feature parameter set after redundancy removal is split into two feature parameter subsets corresponding to the forward and reverse steering angles of the steering wheel. Perform elbow rule analysis on the two subsets separately to determine the optimal number of clusters. The lateral dynamic stability state of vehicles is classified into four categories: stable, tending towards stability, tending towards instability, and unstable. K-Means clustering is performed based on preset clustering parameters to assign state classification labels to all samples; Merge the subsets with the same label within the positive and negative corner subsets to obtain the feature dataset with state labels.

5. The method for identifying the lateral dynamic stability state of a vehicle according to claim 4, characterized in that, The feature-label relevance filtering includes: The maximum mutual information coefficient algorithm is used to calculate the correlation between each feature and the state label in the feature dataset with state labels, and the correlation score of each feature is obtained. Features with a relevance score lower than the first preset score and no substantial correlation with the label, as well as features with a relevance score lower than the second preset score and difficult to measure in real vehicles, are removed. After removing features, the final effective feature set is determined, and the training dataset is constructed.

6. The method for identifying the lateral dynamic stability state of a vehicle according to claim 1, characterized in that, The Stacking ensemble learning framework includes: The framework is set up with a two-layer structure: the first layer is the base model layer, and the second layer is the meta-model layer. The base model layer takes effective features from the training dataset as input and outputs predicted probabilities for four types of stable states. The meta-model layer takes as input the predicted probabilities from the base model layer and outputs as the classification results of the vehicle's lateral dynamic stability state.

7. The method for identifying the lateral dynamic stability state of a vehicle according to claim 1, characterized in that, The training process of the vehicle lateral dynamic stability state discrimination model is as follows: The training dataset is divided into a training subset and a test subset according to a preset ratio; A five-fold cross-validation method was used to pre-train multiple candidate base models, and candidate base models with satisfactory accuracy were selected based on the validation accuracy. Through ablation experiments, CART and SVM were identified as the optimal ensemble framework for Stacking ensemble learning, with MLP as the meta-model. The model training and validation are completed based on the training subset and test subset and the optimal ensemble framework, resulting in a well-trained vehicle lateral dynamic stability state discrimination model. The candidate base models include Classification and Regression Tree (CART), Multilayer Perceptron (MLP), Gaussian Naive Bayes, Support Vector Machine (SVM), k-Nearest Neighbors (k-NN), and Linear Logistic Regression (LRRP).

8. A vehicle lateral dynamic stability state identification system, characterized in that, include: The data acquisition module is used to collect the lateral dynamic characteristics of the vehicle to be identified; The state identification module is used to input the lateral dynamic features of the vehicle to be identified into the pre-trained vehicle lateral dynamic stability state identification model and output the vehicle lateral dynamic stability state identification result. The vehicle lateral dynamic stability state discrimination model is based on a Stacking ensemble learning framework with CART and SVM as base models and MLP as meta-model. The vehicle lateral dynamic stability state discrimination model is trained on a training dataset based on the Stacking ensemble learning framework. The specific steps for constructing the training dataset are as follows: Based on the sinusoidal sweep frequency steering condition, the vehicle's lateral dynamic characteristic parameters are collected, and sample preprocessing is performed to remove duplicate samples to obtain the initial feature dataset. The progressive maximum mutual information coefficient method is used to sequentially perform feature redundancy removal, K-Means clustering state partitioning and labeling, and feature-label correlation filtering operations on the initial feature dataset to construct the training dataset.

9. A vehicle lateral dynamic stability state identification device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the vehicle lateral dynamic stability state identification method according to any one of claims 1-7 when executing the computer program.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it is used to implement the vehicle lateral dynamic stability state identification method according to any one of claims 1-7.

Images