Soft sensor for non-invasive detection of process events based on Eigenresponse Fuzzy Clustering

Changes in process states and properties can be observed through measured variables. In this way, by classifying time series segments of measured data, changes in model parameters can be detected and the system state can be inferred. Time series classification methods are used in many fields, but the work presented here focuses mainly on the field of manufacturing. In the category of whole-series time series classifiers, the Nearest Neighbor classifier is often used. The aim of this work is to introduce an alternative supervised method for time series classification - Eigenresponse Fuzzy Clustering (EFC). We introduce class eigenresponses, which are time series prototypes of a class. We propose the learning eigenresponses for each class using a fuzzy clustering technique. Unlike some existing methods, we propose the use of multiple prototypes per class to better describe a wider range of values for each class. Moreover, the presented method is evaluated on several datasets. Using a dataset obtained on an industrial test bench on an e-bike drive assembly line, the method correctly classifies all time series. To further validate the performance, a set of publicly available datasets (UCR Archive) is used. For the category of datasets most similar to the target industrial application, an improvement over the benchmark approach is obtained. (c) 2022 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).