Research On Anomaly Detection Method For Driving Data From Internet Of Vehicle Based On Isolation Forest

Limited by the collection equipment and technology,the traditional Driving Cycle data collection often has the problem of small data sample size and low data quality.With the development of Internet of Vehicle,it is not difficult nowadays to obtain the vehicle driving data in real time,for which the vehicle monitoring platform can realize the data collection of vehicle terminal equipment and vehicle sensors,data transmission of wireless vehicle networking and data storage of remote database through wireless equipment and 5G network.However,how to detect abnormal data efficiently and improve the quality of data presents a new challenge to the researchers.At present,in the research of anomaly detection for driving data from Internet of Vehicle,the traditional methods are mostly setting a uniform threshold to detect abnormal data,which causes low utilization of equipment information,one-sided evaluation results,limited detection effect affected by the initial dataset’s characteristics and other problems.Although Machine Learning is now introduced in the anomaly detection algorithm,it inevitably leads to the problem of dimensional disaster due to the high dimensional characteristics of the dataset.What’s more,the traditional data dimension reduction method has the problem that it can not be able to reflect the vehicle’s driving state characteristics and the essential characteristics of dynamics.In order to solve these problems,this paper constructs dynamic observation deviation indexes by building dynamic force observation model and dynamic speed observation model,and constructs indexes from the perspective of vehicle dynamics to reduce the dimensionality of multi-dimensional vehicle driving dataset,which solves the problems that the traditional vehicle driving data dimensionality reduction cannot reflect the essential characteristics of vehicle dynamics model causing the limited effect of non-linear dependency data detection.Besides,combining with the Isolation Forest,it can detect abnormal data more efficiently and accurately for massive highdimensional vehicle driving dataset,and greatly reduce its computational cost and time cost.The main work of this paper is as follows:(1)Construct the simulation dataset based on Truck Sim and complete the addition of abnormal dataConstruct the characteristics of the vehicle simulation model based on Truck Sim,select the control strategy,and simulate different load conditions(no load,half load,full load),and complete the construction of 3D road simulation model.The simulation tracking data is inputted from the China heavy-duty commercial vehicle test cycle,and the simulation tracking dataset containing various data channels is output.The obtained dataset is used as the truth-value dataset for the subsequent study,and it is input to MATLAB for articulation and merging to realize the expansion of data volume.The point anomaly data and segment anomaly data are simulated and added by generating random anomaly offsets,after which the anomaly data is labeled.In addition,data types with different degrees of abnormality are simulated by random numbers obeying different distributions,and finally weak abnormal,moderate abnormal and strong abnormal simulation datasets containing abnormal data labels are obtained.(2)Construct kinetic feature indicators to realize kinetic dimensionality reductionFrom the perspective of vehicle dynamics,the motion of the vehicle is studied according to the vehicle longitudinal dynamics,and the longitudinal driving force and driving resistance are modeled from the equilibrium relationship of the longitudinal dynamics of the vehicle,thus the construction of the force observation model is completed.Combining the concepts of kinematics and dynamics theories,this paper defines the kinematic speed and the dynamical speed of the vehicle,and complete the construction of the speed observation model.The deviation between the observed values and the theoretical model is calculated by integrating the time-varying vehicle state observation data into the constructed force observation model and speed observation model,and the characteristic indexes of dynamic force observation deviation and dynamic speed observation deviation are constructed.And according to the dynamic observation deviation characteristic indexes,the two-dimensional characteristic simulation dataset is calculated,and the dimensionality reduction from the dynamics perspective is realized.(3)Use the framework of Isolation Forest algorithm to realize anomaly detectionRealize the detection of anomaly data for the constructed two-dimensional simulation dataset by using Isolation Forest algorithm.In order to further illustrate the effectiveness of the method proposed in this paper,we obtain the real vehicle dataset from Internet of Vehicle and carry out the manual anomaly labeling process to obtain the real vehicle dataset containing labels.Then the classical anomaly detection algorithm 3 method is reproduced for the real vehicle dataset and the simulation dataset with different anomaly degrees,and the effectiveness of 3 method and Isolation Forest is compared and analyzed based on the results of several indexes,such as Prediction,Recall and F1 score.