2009 Volume 21 Issue 4 Pages 587-597
Skill science begins to attract much research attention recently. One of the most important issue in skill science is to discover rules of cooperative motion. In this problem, such relations as “if state B occurs in one sequence, state A will occur within a few minutes in another sequence” should be learned by analyzing time series data from sensors. However, in the actual applications, we often have to deal with time series data without knowing what the internal states are. For these cases, conventional correlation-based methods for time series analysis do not work well since they lack the capability of handling complicated structures such as the relations between time intervals. Based on the background, we propose a procedure to discover temporal relational patterns inductively after converting time series data into a set of state transition sequences. This paper proposes a new framework for analyzing time series multivariate data with non-stationary feature which are often observed in such data as EMG of human performance of skillful motion. Firstly, we determine a sequence of stationary stochastic models of a given time series datum for each component of multivariate measurement data. Secondly, we cluster the stochastic models and allocate a symbol to each cluster, resulting in a set of symbol sequences corresponding to each measurement. Thirdly we convert each symbol sequence into a sequence of time-interval-symbols (TIS) by associating each symbol with the start and end timestamps of the interval. Finally, we extract temporal relational rules (TRR) of TIS sequences using Inductive Logic Programming. Then, we can expect that the extracted rules represent some of the important features of the given multivariate time series data. In our experiments, we applied our proposed framework to the EMG of the skillful motion of playing Chellos, and showed that useful skill knowledge can be obtained.
