Condition monitoring based on corrupted multiple time series with common trends

Condition monitoring is a fundamental task in the reliability engineering and operation management of a complex industrial system. It aims to detect faults based on sensing data but poses significant challenges when dealing with corrupted multiple time series data in many real-world applications. These time series typically exhibit similar changing patterns influenced by common trends (e.g., workload, ambient condition) and physical relationships among corresponding variables, and are often significantly corrupted large outliers (e.g., transmission interruption). Although several traditional common trend models have been employed to analyze such condition monitoring data in the literature, they are fully parametric, constrained by restrictive assumptions, and not robust to outliers. In this article, we propose a novel semiparametric decomposition model to analyze a set of monitored time series and separate it into common, idiosyncratic, and sparse components, under relatively mild assumptions. We also introduce effective algorithms for model estimation and a monitoring scheme for fault detection. The numerical and real case studies demonstrate the superiority of the proposed method over existing approaches, in terms of both decomposition accuracy and detection performance for system faults.