Batch Process Fault Diagnosis Based on The Combination of Deep Belief Network and Long Short-Term Memory Network

Abstract

With the rapid development of deep learning in recent years, more and more deep architecture models have been used for batch process fault diagnosis. Deep Belief Network (DBN) has advantages in extracting features and processing high-dimensional, non-linear data, but the relevance of time series is not fully considered in training with time-dependent signals. The batch process has the characteristics of non-linearity, multiple working conditions and multiple time periods. Hence, DBN does not perform well in batch process. For this purpose, a method based on the combination of Long Short-Term Memory (LSTM) network and Deep Belief Network (DBN) is proposed. The method first adopts the preprocessing method of variable expansion and continuous sampling, and then uses DBN-LSTM network for feature extraction, time correlation analysis, and fault diagnosis. This method is applied to a class of semiconductor etching process. The experimental results show that the proposed method can effectively extract time-ordered nonlinear fault features from the original batch process data and has high fault diagnosis accuracy.