Adaptive Segmentation of IIoT Time Series Data via Change Point Detection for Machinery Fault Classification

Abstract

Predictive maintenance (PdM) is a critical concept in Industry 4.0 that aims to improve manufacturing processes by predicting the remaining useful time of machinery. The development of PdM models relies on access to sufficient data, including condition monitoring and maintenance data from industrial applications. One of the critical aspects of modern PdM approaches is the classification of potential fault signals using time series data collected from IoT devices. However, in most cases, the non-stationary nature of these time series data often causes difficulties with the validity of traditional segmentation techniques when applied to such dynamically changing data patterns. In this work, we propose an adaptive segmentation approach through change point detection to address the inherent non-stationarity of time series, thereby improving the classification performance of traditional classifiers for fault detection problems. By using an adaptive segmentation scheme, we aim to extract more relevant features that will lead to improved classification performance. Taking the time-sensitive nature of the problem into account, we employed three well-known change point detection algorithms (Pruned Exact Linear Time -PELT- algorithm, Binary Segmentation, and Bottom-up segmentation). The effectiveness of the proposed methods is demonstrated by experiments using two different datasets widely used in the PdM literature.