Research On Vehicle Driving State Recognition Technology Based On Motion Sensor

The vehicle state recognition technology based on machine learning can more effectively assist the driver to drive safely and reduce the traffic accident rate.In the future,it can provide a better treatment method for real-time monitoring of vehicle driving state and automatic driving.In this paper,the MPU-6500 sensor loaded on the vehicle is used to collect the attitude data such as vehicle acceleration and angular velocity,and the machine learning algorithms are used to recognize the driving state of the vehicle.The main research contents of this paper are as follows:In order to mine more features of sensor data,this paper uses the Mallat algorithm of wavelet time-frequency analysis to decompose the original sensor signal into contour coefficients and detail coefficients.Furthermore,this paper analyzes the time-frequency features of the signal,and extracts the fine-grained features of the coefficients which are fused with the coarse-grained time domain and frequency domain features of the original data signal as the feature vectors of the algorithm input.Finally,the feature analysis and selection are carried out with the ensemble learning algorithm.The experimental results verify the validity of fine-grained features.In the process of data collection,there are only a small number of labeled data samples that can be utilized,and a large number of unlabeled data samples consume human and material resources to label them.This paper proposes an extended training set strategy based on a semi-supervised learning algorithm to solve this problem.Combining Co-Forest algorithm and active learning algorithm QBC(Query by committee)to select unlabeled sample,the unlabeled sample is divided into white(high confidence),gray,black(low confidence)three-level sample.In the iterative process,the label selection of the blackness samples is carried out according to the clustering hypothesis,and the gray and black samples are fully utilized.Finally,the selection strategy of the unlabeled samples is optimized.The experimental results show that the improved algorithm has higher recognition accuracy than Co-Forest algorithm.Finally,in order to make the vehicle state recognition technology more effectively applied to real scenarios with higher real-time performance,this paper proposes an identification algorithm based on incremental update model,which uses Online random forest algorithm to predict the data stream whose data distribution changes and the model updates at the same time.The experimental results show that the Online random forest model has better training effect than the Offline random forest,and can be used for the recognition of time-varying vehicle state data streams in real scenes.