Abstract:
Driving behavior data are collected as multi-dimensional time-series data, which are measured by various sensors installed in a vehicle.
These data often include redundant information, e.g., both the wheel speed and the engine speed represent the velocity of the vehicle.
Such redundant information complicates data analysis.
Hence, additional factors need to be analyzed, further calculation is required, and varying levels of redundancy cause the results of the analysis to differ.
Therefore, the multi-dimensional driving behavior data should be converted into latent low-dimensional time-series data that are equivalent to the original data.
Furthermore, driving behavior data are often defective owing to sensor failures.
Therefore, another important task is to reduce the adverse effects of defective data when extracting low-dimensional time-series data.
This study proposes a defect-repairable feature extraction method based on a deep sparse autoencoder (DSAE) to extract low-dimensional time series of latent features from driving behavior data containing defects.
The proposed method is called DSAE with back-propagation~(DSAE-BP).
Through experiments, DSAE is shown to achieve high-performance latent feature extraction for driving behavior analysis.
It extracts highly correlated low-dimensional time series of latent features from different multi-dimensional time-series data of driving behavior by reducing various degrees of redundancy.
In addition, it represents a variety of simple and complex driving behaviors, as illustrated by an application for the visualization of driving behaviors on the basis of the extracted low-dimensional time series of latent features.
Moreover, DSAE-BP reduces the adverse effects of defects on the extracted low-dimensional time series of latent features.
This is demonstrated by applying DSAE-BP to a driving behavior segmentation task involving defects.