It has been a challenge to find patterns in a time series of sensor data for fault detection in a system. Since it is usually not straightforward to discover meaningful features and rules directly from complex time series, data discretization has been popularly employed to reduce data size while preserving meaningful features from the original data, for which the choice of appropriate discretization parameters is crucial. We thus present a systematic discretization procedure of multivariate time series data that includes: 1) label definition in consideration of the estimated distribution functions of sensor signals and the trends of signal's short-term variation and 2) label specification to a set of time segments in order to describe the state of a given system for the time segment as a discretized state vector. Formal definitions of fault patterns and discretization problems are made to conduct empirical sensitivity analysis of discretization parameters in finding the most informative fault patterns. We then investigate the relationship between the parameters and the key characteristic indicators of sensor signals. The computational results with the ten real-world data sets provide a practical advice to select appropriate parameters.