Supervised Time Series Segmentation as Enabler of Multi-Phased Time Series Classification: A Study on Hydraulic End-of-Line Testing

Abstract

Multi-phased time series are found in many industrial processes. Their classification still poses a big challenge for algorithms compared to single-phased time series forms. To overcome this issue, this paper suggests using deep learning to generate timestamp-wise state labels that serve as semantic annotations for all measured data points. We investigate whether the availability of state labels can boost the performance of machine learning classifiers by enabling state-wise feature extraction in multi-phased time series. The study is performed on a real-world industrial classification problem in a hydraulic pump factory. Various state label predictions with different accuracy scores are created via deep learning-based time series segmentation. We evaluate how the accuracies of the state label predictions affect the results of the binary classification. Our results show that in settings where accurate state labels are present the classification Fl-scores were significantly higher compared to baseline approaches. Therefore, we emphasized the need to find well performing time series segmentation methods.