(173j) A Deep Learning-Based Feature Extraction Framework for Monitoring High-Order Nonstationary Industrial Processes

Early detection of abnormal process deviations is an urgent need for modern manufacturing plants. When distributed control system is widely used to reach the effective control of process, a large number of process variables can be measured and stored, which greatly facilitates the development of data-driven process monitoring technologies. Numerous multivariate statistical and artificial intelligence methods have been employed to establish process monitoring models by extracting internal correlations among variables. However, most reported methods are developed by assuming that the process is stationary, while industrial operation data show high-order nonstationary characteristics, as the statistical indicators, such as the mean and standard deviation of certain variables, change with time. Differencing is a common and effective way of dealing with nonstationary sequences in the field of economics, but the data processed after more than second-order differencing will lack interpretability. Moreover, useful process information, especially the small changes at the early stage of faults, can be removed simultaneously by differencing, making the fault difficult to be early detected. Cointegration theory is an alternative to nonstationary modeling by extracting the long-term equilibrium relationship among nonstationary variables. However, it is still not applicable to industrial processes because it is still limited to the extraction of linear relationships.

In this work, a novel process monitoring framework is proposed by integrating the idea of differencing with a deep encoder-decoder recurrent neural network. Considering that the forget gate of most memory neural networks is saturated with long-term stationary information, and the short-term nonstationary characteristics cannot be effectively extracted, nonstationary information obtained by differential transformation is input to the next memory neural network structure in this work, and therefore the forget gate can be used to choose how much nonstationary information to forget or remember. A stacking multilayer network is constructed to effectively extract high-order nonstationary characteristics in industrial processes. The process nonlinearity and dynamic can be extracted as well by utilizing recurrent neural network. The encoder-decoder structure is employed to perform unsupervised fault detection. The mean square error and KL divergence are integrated as the monitoring index of the proposed method. Case studies on both the Tennessee Eastman process and an industrial continuous catalytic reforming heat exchange unit are presented to illustrate the effectiveness of the proposed method. The results confirm that the proposed method can be applied to early detect the faults with a low false alarm rate by effectively extracting the high-order nonstationary characteristics of industrial processes.