Novel Two-Stream Deep Fast-Slow Features Extraction for Chemical Process Soft Sensing Application

Chemical process data are the coexistence of fast-varying and slow-trend, which often exhibit strong nonlinearity and time-varying characteristics due to the complex physical and chemical mechanisms. The single model-based methods cannot show satisfactory performance, because the slow features and fast features are difficult to be extracted. This paper proposes a two-stream fast and slow feature fusing model (TS-FSFM), in which two-stream network structure including a slow feature stream and a fast feature stream is designed to extract slow and fast features in parallel. The slow feature stream is equipped with an encoder-decoder-based Siamese network and a fully connected (FC) layer for slow feature extraction, where long short-term memory (LSTM) networks are employed as the encoder and decoder units. Meanwhile, the fast feature stream consists of the conventional LSTM and FC networks for fast feature extraction. Finally, the features learned from the two streams are fused, and a supervised learning regression layer is employed for process soft sensing. The effectiveness and superiority of the proposed method are demonstrated on an industrial process case.