Automatic identification method for automobile handling stability test items based on data feature driving

Abstract

The invention discloses an automatic identification method for an automobile handling stability test item based on data feature driving, and aims to solve the problem that the scientificity and the preciseness are insufficient due to the fact that an existing identification method for the handling stability test item cannot quantify the correlation between a feature physical quantity and the test item. Firstly, test feature physical quantities are primarily selected, a maximum information coefficient algorithm is introduced, all the feature physical quantities collected in a test are screened, a feature physical quantum set meeting the correlation threshold requirement is obtained, then a random forest algorithm is adopted to establish a test item recognition model, the test item model is trained, and the test item recognition result is obtained. And automatic identification of the test items is realized. According to the method, the correlation between the characteristic physical quantity and the test item is quantified by using the maximum information coefficient algorithm, and then the characteristic physical quantity with high correlation is used as the input of the random forest model for test item identification, so that the identification accuracy is ensured, and the identification process is more scientific and rigorous.

Description

Technical Field

[0001] The invention relates to an automatic identification method for automobile handling stability test items, and in particular to an automatic identification method for automobile handling stability test items driven by data features. Background Art

[0002] In the process of automobile product development, real vehicle field tests are the key verification method to ensure product performance and quality. As a core factor affecting vehicle driving safety and driving experience, handling stability is an important part of product competitiveness. Therefore, a large number of real vehicle field tests on handling stability must be carried out during the product development stage. However, the handling stability test involves many working conditions, the amount of data is huge, and the traditional manual processing method is inefficient. In order to improve the test efficiency, OEMs actively explore the use of automation to solve the problem of processing massive data, and the automatic identification technology of test items is the basis and starting point for realizing automatic data processing.

[0003] According to literature research, there are mainly two methods for automatic identification of vehicle handling stability test items, as shown below: Zhong Zhaohui et al. established a convolutional neural network model by analyzing the image features of vehicle handling stability test item data in the paper "Automobile handling stability test type identification method based on convolutional neural network" to achieve the classification of test items. Lu Pingping et al. used dynamic time warping method and neural network method to achieve the classification of test items in patent CN107238500A "Method for establishing a rapid evaluation system for automobile handling stability test".

[0004] The existing methods are based on human experience, starting from the characteristics of each test item, selecting a characteristic physical quantity, and using its time domain curve characteristics as the feature of test item identification. However, this method relies heavily on human experience and does not quantify the correlation between the characteristic physical quantity and the test item, which makes the identification process lack sufficient scientificity and rigor. Summary of the invention

[0005] In order to solve the problem that the existing handling stability test item identification method lacks scientificity and rigor due to the inability to quantify the correlation between characteristic physical quantities and test items, the present invention provides a data feature-driven automobile handling stability test item automatic identification method, firstly, the test characteristic physical quantities are preliminarily selected, and the maximum information coefficient algorithm is introduced to screen all the characteristic physical quantities collected in the test to obtain a characteristic physical quantum set that meets the correlation threshold requirement, and then the random forest algorithm is used to establish a recognition model, and the model is trained to realize automatic identification of the test items.

[0006] The objective of the present invention is achieved through the following technical solutions:

[0007] A method for automatically identifying vehicle handling stability test items based on data feature drive includes the following steps:

[0008] S1. Preliminary selection of characteristic physical quantities for handling stability test and preparation of test data;

[0009] S2. Preprocess the test data prepared in step S1 and define data labels;

[0010] S3. Feature screening of characteristic physical quantities of handling stability test based on maximum information coefficient algorithm;

[0011] S4. Construct a test item identification model based on random forest algorithm;

[0012] S5. Perform model training on the test item recognition model constructed in step S4.

[0013] Furthermore, the step S1 comprises:

[0014] S11. Determine the characteristic physical quantities of the handling stability test, including: steering wheel angle, steering wheel torque, steering wheel angular velocity, lateral acceleration, yaw angular velocity, roll angle, vehicle speed, and longitudinal acceleration;

[0015] S12. Prepare test data for each test in the vehicle handling stability test according to the characteristic physical quantities of the handling stability test determined in step S11.

[0016] Furthermore, the step S2 comprises:

[0017] S21. Low-pass filtering and standardization of the test data prepared in step S1;

[0018] S22. Define data labels: define the test data under different test items as different labels, and use the data after defining the labels as the candidate data set in the feature screening algorithm of step S3, and each label represents a candidate input variable.

[0019] Furthermore, the step S21 includes:

[0020] First, a one-dimensional median filter method is used to remove outliers; assuming that the original data sequence is {p1, p2, …p i ,…p q}, the filter window length is set to L = 2k + 1, and the data in the window is {p i-k ,…,p i ,…p i+k}, sort the L data in the window by size, and the data after filtering output is the center point data after sorting in the window; the filter output expression is formula (1):

[0021] yi =median[{p i-k ,…,p i ,…p i+k ] (1)

[0022] In the formula, i∈[ik,…,i,…,i+k]; y i is the median filter output; median means median filter;

[0023] Then, a FIR filter is used for low-pass filtering, and the filtered test data is phase corrected to eliminate the time domain lag caused by filtering;

[0024] Finally, the test data is standardized so that the characteristic physical quantities have the same magnitude. The expression is:

[0025] x′ ik =(x ik -x min ) / (x max -x min ) (2)

[0026] In the formula, x′ ik is the standardized data, x ik is the observed value, x min is the minimum value of the observed value, x max is the maximum value of the observations.

[0027] Furthermore, the step S3 comprises:

[0028] S31. Construct a candidate variable set, and define the eight characteristic physical quantities to be selected as the candidate variable set X = {X, X2…X α}, i∈[1,8], define the characteristic physical quantum set S as an empty set;

[0029] S32. The maximum information coefficient algorithm is used to calculate the correlation between the characteristic physical quantity and the data label in the test data. The maximum correlation coefficient expression is as follows:

[0030]

[0031] In the formula, X represents the characteristic physical quantity; Y represents the data label; p(x,y) represents the joint probability distribution function of X and Y; p(x) represents the marginal probability distribution function of X; p(y) represents the marginal probability distribution function of Y; m, n represent the number of division grids in the x and y directions of the two-dimensional space; B is a variable, and the size of B is set to 0.6 power of the data amount;

[0032] S33. According to formula (3), in the variable set X = {X1, X2...X α}, α∈[1,8] is used to calculate each variable X α The MIC value of the data label is recorded, and the largest value of the candidate characteristic physical quantity X is recorded. α , count the characteristic physical quantity into the sample set S, and remove the characteristic physical quantity X from the candidate variable set X α ;

[0033] S34. Determine the remaining characteristic physical quantities X α The correlation with the test item type Y, when the correlation judgment condition is not met, returns to step S33 to continue screening; when the correlation judgment condition is met, it means that the remaining characteristic physical quantities of the set X and the test item type do not meet the set correlation threshold, and the characteristic physical quantum set S is output;

[0034] The correlation determination condition expression is as follows:

[0035] MIC(Y;X a )<η·MIC(Y;Y) (4)

[0036] In formula (4), η is the threshold of correlation. If MIC(Y; X a )<η·MIC(Y;Y), then the variable is considered to contain no information of category Y and is judged as an irrelevant variable;

[0037] The threshold η of the correlation is selected to satisfy formula (5):

[0038] minMIC(X a ;Y) / MIC(Y;Y)≤η≤maxMIC(X a ;Y) / MIC(Y;Y) (5)

[0039] Where minMIC(X a ; Y) / MIC(Y; Y) represents the lower limit of η, maxMIC(X a ; Y) represents the upper limit of η, η∈[0,1].

[0040] Furthermore, the step S4 comprises:

[0041] Random forest uses a self-service sampling method, taking a data set containing physical quantity features such as vehicle speed, roll angle, yaw angular velocity, steering wheel angle, and steering wheel torque as the original data set, and extracting multiple samples from the original data set to form multiple sub-data sets;

[0042] Each sub-dataset is used to build a decision tree. In the process of building the decision tree, a part of the features are randomly selected for splitting. Each time the tree splits, each split is quantitatively scored according to the Gini impurity, and the feature with the smallest Gini impurity is selected as the node; the Gini impurity formula is as follows:

[0043] Gini(U) = ∑p(u)(1-p(u i )) (6)

[0044] In formula (6), p(u i ) represents the probability that a random sample belongs to category i;

[0045] Set the specified number of decision trees until the specified number of decision trees are constructed;

[0046] The prediction results of all decision trees are counted by a voting strategy, and the model output after statistics is the category of the test item, that is, the data label defined in step S2.

[0047] Furthermore, the step S5 comprises:

[0048] Divide the data with labels defined in step S2 into a training set and a test set;

[0049] Using the training set to identify the test item model constructed in step S4;

[0050] After the model training is completed, the test set data is used to verify the model training effect to ensure the accuracy of the model recognition results.

[0051] The present invention can bring the following beneficial effects:

[0052] The present invention provides an automatic identification method for automobile handling stability test items based on data feature drive, firstly, the test characteristic physical quantity is preliminarily selected, and the maximum information coefficient algorithm is introduced to screen all the characteristic physical quantities collected in the test, and the characteristic physical quantum set that meets the correlation threshold requirement is obtained, and then the random forest algorithm is used to establish a test item identification model, and the test item model is trained to realize automatic identification of the test items. The present invention uses the maximum information coefficient algorithm to quantify the correlation between the characteristic physical quantity and the test item, and then uses the characteristic physical quantity with high correlation as the input of the random forest model to identify the test item, thereby ensuring the accuracy of the identification and making the identification process more scientific and rigorous. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a flow chart of a method for automatically identifying vehicle handling stability test items based on data feature drive according to an embodiment of the present invention;

[0054] Figure 2 This is a flow chart of feature screening based on maximum information coefficient in an embodiment of the present invention;

[0055] Figure 3 A test item identification model training error curve constructed for an embodiment of the present invention;

[0056] Figure 4 A confusion matrix for training a test item recognition model constructed for an embodiment of the present invention;

[0057] Figure 5 A test confusion matrix of the test item recognition model after training in an embodiment of the present invention;

[0058] Figure 6 This is the test result of the test item recognition model after training according to the embodiment of the present invention. DETAILED DESCRIPTION

[0059] The actual implementation of the method of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0060] This embodiment is a method for automatically identifying automobile handling stability test items based on data feature drive. First, the test characteristic physical quantities are preliminarily selected, and the maximum information coefficient algorithm is introduced to screen all characteristic physical quantities collected in the test to obtain a characteristic physical quantum set that meets the correlation threshold requirement. Then, the random forest algorithm is used to establish a test item recognition model, and the test item model is trained and evaluated. Figure 1 As shown, it mainly includes the following steps:

[0061] S1. Preliminary selection of characteristic physical quantities for handling stability test and preparation of test data:

[0062] S11. Determine the characteristic physical quantities of the handling stability test:

[0063] The existing GB / T 6323-2014 "Automobile Handling Stability Test Method" includes seven tests, namely: serpentine test, steering wheel center area handling stability test, steering return test, steering lightness test, steady-state rotation test, steering wheel angle step test, and steering wheel angle pulse test. In addition, the ISO standard also includes double lane change test, steering release open loop test, and single sine test;

[0064] According to the data collection requirements specified in national standards and ISO standards, the characteristic physical quantities that need to be collected for the handling stability test are determined, including: steering wheel angle, steering wheel torque, steering wheel angular velocity, lateral acceleration, yaw angular velocity, roll angle, vehicle speed, and longitudinal acceleration.

[0065] S12. Test data preparation: Prepare test data for each test according to the characteristic physical quantities of the handling stability test determined in step S11.

[0066] S2. Preprocess the test data prepared in step S1 and define data labels:

[0067] S21. Test data preprocessing: low-pass filtering and standardization processing are performed on the test data prepared in step S1.

[0068] First, a one-dimensional median filter method is used to remove outliers. Assume that the original data sequence is {p1, p2, …p i ,…p q}, the filter window length is set to L = 2k + 1, and the data in the window is {p i-k ,…,p i ,…p i+k}, sort the L data in the window by size, and the data after filtering output is the center point data after sorting in the window; the filtering output can be expressed by the formula:

[0069] y i =median[{p i-k ,…,p i ,…p i+k ] (1)

[0070] In the formula, i∈[ik,…,i,…,i+k]; y i is the median filter output; median represents median filtering.

[0071] Then, a FIR filter is used for low-pass filtering with a cutoff frequency of 5 Hz. The filtered test data is phase corrected to eliminate the time domain lag caused by filtering.

[0072] Thirdly, the test data is standardized to eliminate the influence of the amplitude and mean of different test data, so that the characteristic physical quantities have the same magnitude. The expression is as follows:

[0073] x′ ik =(x ik -x min ) / (x max -x min ) (2)

[0074] In the formula, x′ ik is the standardized data, x ik is the observed value, x min is the minimum value of the observed value, x max is the maximum value of the observations.

[0075] S22. Define data labels:

[0076] The test data under different test items are defined as different labels, for example, the data label of the snake test is defined as "1", and the data label of the steering wheel center area handling stability test is defined as "2". The data after defining the labels are used as the candidate data set in the feature screening algorithm of step S3. Each label represents a candidate input variable. The corresponding relationship between the test items and the data labels is shown in Table 1:

[0077] Table 1 Correspondence between test items and data labels

[0078] Test Project Data Labels Snake test 1 Steering wheel center area handling stability test 2 Steering return test 3 Steering ease test 4 Steady-state rotation test 5 Double lane change test 6 Steering release open loop test 7 Single sine test 8

[0079] S3. Based on the maximum information coefficient (MIC) algorithm, characteristic physical quantities of the handling stability test are screened, such as Figure 2 As shown:

[0080] S31. Construct a candidate variable set, and define the eight characteristic physical quantities such as the steering wheel angle, steering wheel torque, and steering wheel angular velocity as the candidate variable set X = {X, X2…X α}, i∈[1,8]. Define the characteristic physical quantum set S as an empty set.

[0081] The correspondence between the candidate variable names and variable symbols is shown in Table 2:

[0082] Table 2 Correspondence between candidate variables and variable symbols

[0083]

[0084]

[0085] In this embodiment, the characteristic physical quantity set X={X1, X2…X8}, α∈[1,8] of the 8 test items in Table 1 is used as the input of the maximum information coefficient algorithm, and the data labels of the corresponding items are used as output.

[0086] S32. Use the maximum information coefficient algorithm to calculate the correlation between the characteristic physical quantity and the data label in the test data. The maximum correlation coefficient expression is as follows:

[0087]

[0088] In formula (3), X is a random variable, which represents the characteristic physical quantity; Y is a random variable, which represents the data label; p(x, y) is the joint probability distribution function of X and Y; p(x) is the marginal probability distribution function of X; p(y) is the marginal probability distribution function of Y; m, n are the number of division grids in the x and y directions of the two-dimensional space; B is a variable, and the size of B is set to 0.6 of the data amount.

[0089] S33. According to formula (3), in the variable set X = {X1, X2...X α}, α∈[1,8] calculates each variable X α The MIC value of the data label is recorded, and the largest value of the candidate characteristic physical quantity X is recorded. α , count the characteristic physical quantity into the sample set S, and remove the characteristic physical quantity X from the candidate variable set X α .

[0090] In this embodiment, the maximum information coefficient of each characteristic physical quantity and the project category is calculated according to formula (3), sorted from large to small, and the cumulative contribution rate of the characteristic physical quantity to the model classification is obtained, see Table 3.

[0091] Table 3 MIC values ​​of characteristic physical quantity sets

[0092]

[0093]

[0094] According to formula (3), the MIC value of each feature variable and category label is calculated. It can be seen from Table 3 that:

[0095] minMIC(X a ; Y)=0.039,maxMIC(X a ;Y)=0.228. MIC(Y;Y)=1

[0096] S34. Determine the remaining characteristic physical quantities X α When the correlation with the test item type Y does not meet the discrimination condition of formula (4), it means that the remaining characteristic physical quantities in the set X are still related to the test item type, so return to step S33 and continue screening; when the discrimination condition is met, it means that the remaining characteristic physical quantities in the set X do not meet the set correlation threshold with the test item type, and the characteristic physical quantum set S is output.

[0097] The correlation discriminant relationship is as follows:

[0098] MIC(Y;X a )<η·MIC(Y;Y) (4)

[0099] In formula (4), η is the threshold of correlation. If MIC(Y; X a )<η·MIC(Y;Y), it is considered that the variable does not contain information of category Y and is determined to be an irrelevant variable.

[0100] Formula (5) can provide guidance for selecting the correlation threshold η:

[0101] minMIC(X a ;Y) / MIC(Y;Y)≤η≤maxMIC(X a ;Y) / MIC(Y;Y) (5)

[0102] Where minMIC(X a ; Y) / MIC(Y; Y) represents the lower limit of η, maxMIC(X a ; Y) represents the upper limit of η, η∈[0,1].

[0103] In this embodiment, according to formula (5), the specific value range of the correlation threshold η is η∈[0.039,0.228].

[0104] According to the specific value range of the correlation threshold η, η is set to 0.1, and the characteristic physical quantity that meets formula (3) is obtained as the characteristic physical quantity subset S. The characteristic subset S includes the characteristics {vehicle speed, roll angle, yaw angular velocity, steering wheel angle, steering wheel torque}.

[0105] In this embodiment, the calculation results of the feature subset S are combined and the following four factors are comprehensively considered:

[0106] 1. In the process of feature selection, the individual contribution rate of each feature and their contribution rate to the overall accumulation are mainly considered. The individual contribution rates of all features are analyzed in Table 3. Among them, the individual contribution rates of three features are lower than the correlation threshold of 0.1, which indicates that they contribute less to the classification ability of the model. Eliminating the three features can simplify the model and reduce the amount of calculation.

[0107] 2. After removing the above features, the cumulative contribution rate of the remaining features was analyzed. The results showed that after removing these three features, the cumulative contribution rate of the remaining features could still reach 80.1%, which could maintain the prediction accuracy of the model.

[0108] 3. In terms of longitudinal acceleration, since the steering tests are all carried out under uniform speed conditions, its values ​​fluctuate around zero and have little correlation with the test items.

[0109] 4. There is a strong coupling relationship between lateral acceleration and yaw rate. According to the maximum information coefficient algorithm analysis, yaw rate has a higher sensitivity in capturing the dynamic response characteristics of the vehicle. Therefore, yaw rate is given priority in feature selection, and lateral acceleration is eliminated.

[0110] Finally, vehicle speed, roll angle, yaw rate, steering wheel angle, and steering wheel torque are selected as the features of the classification model. These features not only have a high contribution rate, but also can fully reflect the vehicle's handling stability, providing effective guarantee for the accuracy of the classification model.

[0111] S4. Constructing a test item identification model based on random forest algorithm:

[0112] Random forest uses a bootstrap sampling method (random sampling with replacement) to extract multiple samples from the original data set to form multiple sub-data sets. In this embodiment, a data set containing physical quantity features such as vehicle speed, roll angle, yaw angular velocity, steering wheel angle, and steering wheel torque is used as the original data set, and multiple samples are extracted from the original data set to form multiple sub-data sets.

[0113] Each sub-dataset is used to build a decision tree. In the process of building the decision tree, a part of the features are randomly selected for splitting. Each time the tree splits, each split is quantitatively scored according to the Gini impurity, and the feature with the smallest Gini impurity is selected as the node. The Gini impurity formula is as follows:

[0114] Gini(U) = ∑p(u)(1-p(u i )) (6)

[0115] In formula (6), p(u i ) represents the probability that a random sample belongs to class i.

[0116] A specified number of decision trees are set until a specified number of decision trees are constructed. In this embodiment, the specified bibliography of the decision tree is 100, and the minimum number of leaf node samples is set to 1.

[0117] The prediction results of all decision trees are counted using a voting strategy, and the model output after statistics is the category of the test item, that is, the label defined in step S2.

[0118] S5. Perform model training on the test item recognition model constructed in step S4.

[0119] Divide the labeled data in step 2 into training set and test set in a ratio of 7:3.

[0120] The test item recognition model constructed in step S4 is trained using the training set. After the model training is completed, the model training effect is verified using the test set data to ensure the accuracy of the model recognition result.

[0121] In this embodiment, the error curve of model training is as follows: Figure 3 As shown, observe Figure 3 Error curve, the error curve of the model drops rapidly during the training phase, and the error gradually decreases as the number of decision trees increases. After the number of decision trees reaches 50, the error curve begins to converge, indicating that the model has been able to stably learn the key features in the data, and additional decision trees no longer significantly improve the performance of the model. 100 decision trees provide the model with enough capacity to capture the complexity and nonlinear relationships in the data, and the setting of the minimum number of leaf node samples to 1 allows the model to generate more detailed decision boundaries, thereby improving classification accuracy.

[0122] Use the confusion matrix to visualize the training results and the reverse, and observe Figure 4 From the training confusion matrix of the classification model, we can see that the model training set has a total of 1,305 samples, of which 1,305 samples are correctly classified, and the model training accuracy is 100%.

[0123] After the model training is completed, the test set data is used to verify the model training effect to ensure the accuracy of the model recognition results. Figure 5 The test confusion matrix and Figure 6 The classification model test results show that there are 559 test set samples, of which 555 are correctly classified and 4 are incorrectly classified. The test set accuracy is 99.28%, and the model has a very high recognition accuracy, which meets the usage requirements.

[0124] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for automatic identification of vehicle handling stability test items based on data feature drive, characterized in that: The following steps are involved: S1. Preliminary selection of characteristic physical quantities for handling stability test and preparation of test data; S2. Preprocess the test data prepared in step S1 and define data labels; S3. Feature screening of characteristic physical quantities of handling stability test based on maximum information coefficient algorithm; S4. Construct a test item identification model based on random forest algorithm; S5. Perform model training on the test item recognition model constructed in step S4.

2. The method for automatically identifying vehicle handling stability test items based on data feature drive according to claim 1, characterized in that: The step S1 comprises: S11. Determine the characteristic physical quantities of the handling stability test, including: steering wheel angle, steering wheel torque, steering wheel angular velocity, lateral acceleration, yaw angular velocity, roll angle, vehicle speed, and longitudinal acceleration; S12. Prepare test data for each test in the vehicle handling stability test according to the characteristic physical quantities of the handling stability test determined in step S11.

3. The method for automatically identifying vehicle handling stability test items based on data feature drive according to claim 1, characterized in that: The step S2 comprises: S21. Low-pass filtering and standardization of the test data prepared in step S1; S22. Define data labels: define the test data under different test items as different labels, and use the data after defining the labels as the candidate data set in the feature screening algorithm of step S3, and each label represents a candidate input variable.

4. The method for automatically identifying vehicle handling stability test items based on data feature drive according to claim 3, characterized in that: The step S21 comprises: First, a one-dimensional median filter method is used to remove outliers; assuming that the original data sequence is {p1, p2, …p i ,…p q }, the filter window length is set to L = 2k + 1, and the data in the window is {p i-k ,…,p i ,…p i+k }, sort the L data in the window by size, and the data after filtering output is the center point data after sorting in the window; the filter output expression is formula (1): y i =median[{p i-k ,…,p i ,…p i+k ] (1) In the formula, i∈[ik,…,i,…,i+k]; y i is the median filter output; median means median filter; Then, a FIR filter is used for low-pass filtering, and the filtered test data is phase corrected to eliminate the time domain lag caused by filtering; Finally, the test data is standardized so that the characteristic physical quantities have the same magnitude. The expression is: x′ ik =(x ik -x min ) / (x max -x min ) (2) In the formula, x′ ik is the standardized data, x ik is the observed value, x min is the minimum value of the observed value, x max is the maximum value of the observations.

5. The method for automatically identifying vehicle handling stability test items based on data feature drive according to claim 1, characterized in that: The step S3 comprises: S31. Construct a candidate variable set, and define the eight characteristic physical quantities to be selected as the candidate variable set X = {X, X2…X α }, i∈[1,8], define the characteristic physical quantum set S as an empty set; S32. The maximum information coefficient algorithm is used to calculate the correlation between the characteristic physical quantity and the data label in the test data. The maximum correlation coefficient expression is as follows: In the formula, X represents the characteristic physical quantity; Y represents the data label; p(x,y) represents the joint probability distribution function of X and Y; p(x) represents the marginal probability distribution function of X; p(y) represents the marginal probability distribution function of Y; m, n represent the number of division grids in the x and y directions of the two-dimensional space; B is a variable, and the size of B is set to 0.6 power of the data amount; S33. According to formula (3), in the variable set X = {X1, X2...X α }, α∈[1,8] is used to calculate each variable X α The MIC value of the data label is recorded, and the largest value of the candidate characteristic physical quantity X is recorded. α , count the characteristic physical quantity into the sample set S, and remove the characteristic physical quantity X from the candidate variable set X α ; S34. Determine the remaining characteristic physical quantities X α The correlation with the test item type Y, when the correlation judgment condition is not met, returns to step S33 to continue screening; when the correlation judgment condition is met, it means that the remaining characteristic physical quantities of the set X and the test item type do not meet the set correlation threshold, and the characteristic physical quantum set S is output; The correlation determination condition expression is as follows: MIC(Y;X a )<η·MIC(Y;Y) (4) In formula (4), η is the threshold of correlation. If MIC(Y; X a )<η·MIC(Y;Y), then the variable is considered to contain no information of category Y and is judged as an irrelevant variable; The threshold η of the correlation is selected to satisfy formula (5): minMIC(X a (Y) / MIC(Y)≤η≤maxMIC(X) a (Y) / MIC(Y;Y) (5) Where minMIC(X a ; Y) / MIC(Y; Y) represents the lower limit of η, maxMIC(X a ; Y) represents the upper limit of η, η∈[0,1].

6. The method for automatically identifying vehicle handling stability test items based on data feature drive according to claim 1, characterized in that: The step S4 comprises: Random forest uses a self-service sampling method, taking a data set containing physical quantity features such as vehicle speed, roll angle, yaw angular velocity, steering wheel angle, and steering wheel torque as the original data set, and extracting multiple samples from the original data set to form multiple sub-data sets; Each sub-dataset is used to build a decision tree. In the process of building the decision tree, a part of the features are randomly selected for splitting. Each time the tree splits, each split is quantitatively scored according to the Gini impurity, and the feature with the smallest Gini impurity is selected as the node; the Gini impurity formula is as follows: Gini(U)=∑p(u)(1-p(u i )) (6) In formula (6), p(u i ) represents the probability that a random sample belongs to category i; Set the specified number of decision trees until the specified number of decision trees are constructed; The prediction results of all decision trees are counted by a voting strategy, and the model output after statistics is the category of the test item, that is, the data label defined in step S2.

7. The method for automatically identifying vehicle handling stability test items based on data feature drive according to claim 1, characterized in that: The step S5 comprises: Divide the data with labels defined in step S2 into a training set and a test set; Using the training set to identify the test item model constructed in step S4; After the model training is completed, the test set data is used to verify the model training effect to ensure the accuracy of the model recognition results.

Images