Fault detection and fault state estimateion based on ensemble learning in industrial plants

Abstract

For the safe and stable operation of industrial and chemical plants, it is necessary to monitor and control their operating conditions. Because of the huge amount of operating data in plants, data-based process control systems have received considerable attention in recent years. Controlling each process variable independently is inefficient, because there are many process variables that must be controlled. One practical solution is multivariate statistical process monitoring (MSPM) which monitors multiple process variables and their relationships simultaneously. Principal component analysis is widely used as an MSPM method. However, it cannot consider nonlinearities between process variables and multimodal data distributions. In addition, although process faults can be detected, it is difficult to estimate process states in detail. Therefore we developed a new MSPM method to detect process faults and to estimate each process state in industrial plants simultaneously by combining PCA and ensemble learning. Many PCA models, each of which represents local process state, are prepared using initial database. Fault detection and process state estimation are performed by checking similarity between a query and each PCA model. We demonstrate the effectiveness of the proposed method using numerical simulation data in which actual industrial process is simulated.