Time Series Slow Feature Extraction for Dow Data Problem

Handling process data characterized by strong nonlinearity, high dimensionality, cross-correlations, and auto-correlations presents a considerable hurdle for data-driven soft sensor modeling. While traditional slow feature analysis (SFA) adeptly captures slow and static features from linear data, it may falter in capturing the nonlinear, high-dimensional, and dynamic features inherent in time series data. Conversely, although Long Short-Term Memory (LSTM) networks are designed to address long-term dependencies within sequences, they encounter challenges, particularly with very lengthy data sequences, in effectively capturing these dependencies. Consequently, relying solely on SFA or LSTM may prove inadequate for addressing the complexities of time series data in industrial processes. To confront these challenges, this study proposes an innovative approach termed Slow LSTM (SLSTM), which amalgamates the hidden layers of LSTM and SFA to enhance feature extraction. These extracted features are subsequently fed into a fully connected layer for prediction. The efficacy of this proposed method is validated through comparisons with various methods, including SFA-FC, LSTM, and RNN.