Feature extraction for time series classification using univariate descriptive statistics and dynamic time warping in a manufacturing environment
Nikolai West, Thomas Schlegl, Jochen Deuse
Abstract
The decade-long trend toward process automation and end-to-end machine connectivity has fueled an enormous growth of data recorded in the manufacturing industry. Leveraging this potential requires manufacturing companies to extract actionable insights from the data sources. In particular, handling time series data on a large scale requires the use of feature extraction for dimensionality reduction. For this purpose, we propose a new algorithmic approach that uses Dynamic Time Warping to extract maximally discriminative features in a multivariate data set. A benchmark against `time series feature extraction based on scalable hypothesis tests' and state-of-the-art methods, such as InceptionTime, Convolutional Neural Network or ResNet classifier, to evaluate the overall effectiveness. While the proposed algorithm underperforms for an example data set with comparably low dimensionality, scoring 16.67% and 22.80% lower average accuracy than the benchmarks, it achieves competitive results for the real-world application in a manufacturing environment. Here, the average accuracy reaches a delta of just 12.20 % and simultaneously reduces computational effort by 97.90 %.