Machine learning with segment-aligned multisensor trace data

Abstract

A machine-learning system receives from multiple sensors a set of trace-data time series. Each time series contains a chronological sequence of sensor measurements of one attribute of one instance of a manufacturing product or process. The system partitions each series into a set of contiguous segments and selects one received series to be a standard series for each attribute. The starting and ending measurements of each non-standard time series are then time-aligned to the starting and ending points of the non-standard series' corresponding standard series, using a dynamic time-warping procedure. One or more segments of each aligned non-standard series are then aligned to each segment of the corresponding standard series. The resulting time-aligned, segmented time series are then incorporated into a corpus that is used by a machine-learning module to train a self-learning application.