Data-driven Causal Association Discovery in Manufacturing Industries

With the development of smart sensors, information storage, processing technology and computer performance, large amounts of operating data collected from production process provide opportunities as well as challenges in remaining useful life (RUL) estimation. On one hand, data-driven analysis approaches are experiencing a fast development. On the other hand, the collected variables may be redundant, noisy and high-dimensional for RUL. Thus, data dimension reduction is applied for eliminating useless information. Different from correlation-based methods, causal inference methods can obtain reliable models reflecting causal relationships among interesting variables. Thus, the latter is more suitable in data dimension reduction. In this study, we use PCMCI+, a causal discovery method based on graph model, that handles both lagged and contemporaneous relationships in multi-variable time series. We validate this method on time series data extracted directly from a medium frequency quenching machine. The obtained results confirm that PCMCI+ is able to recognize causal associations among various sensor data. For instance, variables in the same process have relatively larger causal relationships than those in different processes.