Dynamic process monitoring using adaptive local outlier factor

• A fast moving window LOF algorithm is proposed for time-varying and multi-mode industrial processes.
• The proposed method can deal with complex data distributions of both with-in mode data and multi-mode data.
• An updating termination rule is developed to prevent the model fouled by disturbances or faults.
• Any new operating mode can be adapted by applying a switch rule.
• No process prior knowledge is needed.

A numerically efficient moving window local outlier factor (LOF) algorithm is proposed in this paper for monitoring industrial processes with time-varying and multimode characteristics. The key feature of the proposed algorithm can be identified as its underlying capability to handle complex data distributions and incursive operating condition changes including both slow dynamic variations and instant mode shifts. With some updating of the rules developed for accelerating the computation speed, a two-step adaption approach is introduced to keep the monitoring model up-to-date. Then, a switch strategy and an update termination rule are designed to deal with operating mode changes. Due to the utilization of local information, the proposed algorithm has a superior ability both in detecting faulty conditions and fast adapting to new operating modes. Finally, the utility of the proposed method is demonstrated through a numerical example and a non-isothermal continuous stirred tank reactor (CSTR).