Data‐Driven Fault Detection for Multimode Industrial Processes Based on Mixture of Autoencoders

ABSTRACT Data‐driven fault detection technique has been widely applied to industrial processes to reduce production losses caused by faults. Due to complicated chemical reactions and phase changes, the relationships between process variables are nonlinear. In particular, many industrial processes often operate under multiple conditions, resulting in multimode characteristics in the collected data and further complicating the nonlinearities between process variables, rendering challenging difficulties for traditional fault detection methods. In view of this, this paper proposes a novel fault detection method for multimode industrial processes by developing a mixture of autoencoders (MixAE) model. The MixAE is first designed with Gaussian mixture model and multiple single‐hidden layer autoencoders; then, an efficient learning algorithm, based on expectation–maximization algorithm and gradient ascent algorithm, is developed to train the MixAE. By using the confidence index provided by each autoencoder, a comprehensive evaluation statistic is formulated for detecting anomalies. Finally, the proposed fault detection method is validated with a numerical example and an experimental three‐tank liquid level control system. The experimental results demonstrate that the proposed method achieves better fault detection performance for multimode industrial processes compared to traditional methods.