Intelligent online process fault diagnosis through integrating Andrews function, autoencoder, and neural networks

Abstract This paper proposes an online process fault diagnosis scheme that integrates principal component analysis, Andrews function, autoencoder, and multilayer feedforward neural network to enhance the fault diagnosis performance. Useful features are extracted from the online monitoring data by using Andrews function (also known as Andrews plot). To overcome the influence of the arrangement order of variables on the outcomes of Andrews function, a principal component analysis (PCA) model is developed from the normal process operation data. The principal components of the original process data, arranged in the descending order of data variation explained, are used in Andrews function calculation. To address the issue of feature selection in Andrews function, a large number of Andrew function outputs are retained and then compressed using an autoencoder. The compressed features from the encoder are then utilized as the inputs to a multi‐layer feedforward neural network for fault diagnosis. The proposed method is demonstrated on a simulated continuous stirred tank reactor (CSTR) and the diagnostic performance is compared with those of a conventional neural network and an Andrews function based online fault diagnosis schemes. A wide range of process faults in abrupt and incipient fault forms are tested. The results demonstrate that the proposed online process fault diagnosis scheme gives improved fault diagnosis speed and reliability.