Pattern Matching of In-vehicle Acceleration Time Series Data

This paper presents a novel approach to find patterns in vehicle x-y-z acceleration data for use in prognostics and diagnostics. In this problem, vehicles are assumed to travel on the same routes and often times as a part of convoys but their GPS and other position information has been removed for privacy reasons. The goal of the pattern matching scheme is to identify the route or convoy associations within vehicles by using the acceleration data collected onboard these vehicles. A crucial step in solving this problem is to choose the right feature vector, as raw matching of acceleration signals is inappropriate due to different velocities, driving behaviors, vehicle loading, etc. In this paper, we demonstrate the feasibility of using ‘Multi-Scale Extrema Features’ for this application. The paper also addresses implementation details to enhance performance for in-vehicle acceleration data, corrupted by different sources of noise. A novel ‘Multi-Scale Encoding’ method is also proposed for the above feature vector and it leads to a significant improvement in the performance over traditional pattern matching methods. While the main focus of the paper is towards identifying feature vectors that effectively describe in-vehicle acceleration data, the feature vector could potentially be used with acceleration data obtained from other applications.