[1Yin-B] Poster 2

[1Yin-B-06] Effectiveness of Document Vector Representations in Clustering Vehicle Time-Series Data

〇Masaya Mita¹, Kosuke Nakagaki¹, Toshiki Tanaka¹, Masahito Sakaguchi¹ (1. Suzuki Motor Corpolation)

Keywords:

Multivariate Time Series Clustering,Document Vector Representation,Vector Quantization,Transformer,Vehicle Driving Data Analysis

Analyzing driving time-series data collected from vehicles is essential for understanding usage patterns and improving vehicle quality. Time-series clustering is a key approach for extracting patterns from such data. In this study, we focus on the similarity of driving time-series data, which requires global interpretation of entire sequences while taking local patterns such as acceleration into account, and argue that this aligns well with the design principles of document vector representations. Based on this idea, we apply document vector representations to clustering of driving time-series data and evaluate their effectiveness through a route discrimination task using real-world fleet data. Specifically, we propose a method that converts partial time-series segments into discrete token sequences and clusters driving time-series data using document vector representations obtained from Transformer models trained with Masked Language Modeling (MLM) and related approaches as features. Experimental results show that the proposed method achieves higher clustering scores than Dynamic Time Warping (DTW) and self-supervised time-series Transformer models, demonstrating the effectiveness of document vector representations for analyzing driving time-series data.