Automatic signal segmentation using unsupervised learning techniques for process fingerprint identification in smart manufacturing systems

The use of process fingerprints in manufacturing is popular as it provides insight about the processes, and it enables the development of smart applications. Building a fingerprint requires signal segmenting, which can be a cumbersome task as not all the information required is always available. The present paper evaluates different algorithms for pattern recognition in univariate time series using unsupervised learning techniques for event identification. This comparison is performed through extrinsic and intrinsic evaluation metrics obtained from simulated data, representing industrial data. The results indicate that the matrix-profiles based algorithm can perform better than those based on changepoint detection.