The Study Of Terrain Identification Algorithm Based On Model Interpretation Of Random Forest

There is a demand for driving on unpaved roads.In order to realize automatic driving in these relatively harsh environments,it is necessary to identify the terrain and use it as a basis to adjust each control system of the vehicle to achieve the purpose of increasing vehicle performance.At present,there are two main types of terrain recognition methods: vision based methods and vehicle dynamic characteristics based methods.Vision based methods are easier to be affected by the external environment.In contrast,dynamic characteristics based methods are more reliable and easier to be applied to vehicle control.Because of the complex interaction mechanism between different terrains and wheels and the wheels’ nonlinear characteristics,it is difficult to determine the interaction mechanism between wheel and pavement directly through theory.To solve this problem,most scholars realize terrain recognition through machine learning because of the good performance of dealing with nonlinear problems of machine learning.Although the performance of the machine learning is good,it is often treated as a "black box" because of its complex structure,which makes it difficult for people to understand the reasons why the model makes decisions and when the model is prone to make a mistake,therefore it is necessary to interpret the model.Aiming at the complex interaction between terrain and wheels and the black box characteristics of machine learning,this paper studies the terrain recognition algorithm.Firstly,the real vehicle is used to collect the terrain driving data.Based on these data,the features are interpreted,analyzed and picked by the SHAP interpretation method,and then the picked features are used to train the random forest terrain classifier.Finally,the false recognition boundary is extracted by LIME interpretation method,and the classification results are corrected by HMM model.The main research contents of this paper including:(1)Data acquisition and characteristic calculationAccording to the main characteristics of the terrain,four typical terrains are selected:compacted dirt,sand,good asphalt pavement and ice road.Then the data acquisition system has been built by VN1630 and RT3000,and the test vehicle is used to complete the data acquisition under the four kinds of terrains.Through the simple analysis of the influence of terrain parameters on vehicles and taking the vehicle speed into account,the vertical acceleration coefficient and the variance of wheel speed fluctuation are defined,which are combined with the rolling resistance of the vehicle as the main characteristics for terrain recognition.(2)Feature extraction based on SHAP interpretationThe time domain and frequency domain features of the main features are calculated,and the driving feature sample set is established.Using the same algorithm as the terrain classifier,all the features in the sample set are taken as the input,and the random forest algorithm is used to establish a pre-analysis model to extract all the useful message.Then the SHAP interpretation method is used on the pre-analysis model globally to get shapley values of each feature,and the relationship between the features and the terrains in different cases has been analyzed through the partial dependence diagram.Finally,according to the interpretation results,the features with the top 20 importance have been selected as the input of the random forest terrain classifier.(3)The study of terrain identification based on random forestBased on the characteristics of the problem and the complexity of the model,the random forest multi-classifier is selected as the terrain recognition algorithm.The features picked based on the SHAP interpretation method are used as the reference of road recognition,and the simplified sample set is randomly divided into training set and test set.In the training process,the growth of the tree is constrained to complete the training of the random forest terrain classifier model.The real vehicle data in used in the off-line simulation to verify the accuracy of random forest terrain classifier model.(4)Model improvement of random forest based on LIME interpretationAiming at the problem of jitter in the output of the model,three hypotheses are proposed,and under these three hypotheses,the state transition matrix and observation matrix are obtained.The HMM model is established,and the Viterbi algorithm is used to decode the model to obtain the most possible terrain.Then the LIME interpretation method is used to explain the wrong samples after HMM processed,extract the wrong recognition boundary of the model,count the number of samples in the wrong recognition boundary.The internal parameters of HMM model is modified according to the statistical results and whether the sample is in the wrong recognition boundary.Finally,the improved HMM model post processing method has been used to verify that this method can improve the accuracy of the model output results.