Driver Identification Using Vehicle Diagnostic Data with Fully Convolutional Neural Network

Abstract

In this study, driver identification task was performed using artificial neural networks and vehicle diagnosis data without feature engineering which is one of the most challenging steps of machine learning methods. Along with the data sets used in the literature, the data set including rural road, urban and highway routes with 5 different drivers was collected and used experimentally. Driver identification problem is considered as multivariate time series classification problem. With the FCNN (Fully Convolutional Neural Network) architecture established to solve the problem, the success rate of RNN (Recurrent Neural Network)-based architectures and traditional ML methods in the literature has been exceeded. The window size in the sliding window method has been reduced in training and validation. At the same time, the space occupied by the model on the memory and the number of floating point operations were reduced, increasing the possibility of use in in-vehicle systems. While the results are obtained with the stratified fold validation, they are compared with the time series segmentation method in data sets containing a single drive on the route.