This paper addresses the industrial problem of online supervision of a process sequence on a machine without direct access to its controller. In the first stage, the external continuous signal is segmented and, using a evolving timeseries clustering algorithm, converted into a stream of discrete events. The novelty of this paper is the second stage: a DES model (discrete-event system) that continuously checks the logical correctness of the transition sequence and the temporal correctness of durations against four permitted process paths, with the time bounds determined by the minimum–maximum principle from the training set. On an independent test set, 13 932 successful transitions, 1 393 incorrect-order errors, and 51 timing-bound errors were detected; of 2 009 completed cycles, 565 were regular (28.1%). Since the test set lacks independent labels of actual faults, the proportion of deviating cycles (71.9%) does not necessarily correspond to the proportion of cycles that are actually faulty, as it may also include cycle variants not covered by the selected reference paths. Nevertheless, the results confirm the feasibility of online detection of logical and temporal deviations through the combination of signal-to-event conversion and DES-based supervision.
|
