Novel Multiscale Trend Decomposition LSTM Based on Feature Selection for Industrial Soft Sensing

Industrial soft sensing plays a crucial role in process modeling, optimization, and control, and is extensively utilized to predict hard-to-measure process variables. To effectively manage complex process data with varying time scales, a novel multiscale trend decomposition long short-term memory model based on feature selection (FS-MSTD-LSTM) is proposed. In the FS-MSTD-LSTM model, the random forest method is first employed to screen the feature sequences, thereby constructing the optimal input feature set and ensuring the model is provided with the most informative features. Following this feature selection, the seasonal-trend decomposition using loess algorithm is used to perform multi-scale sampling, capturing input features with short-, medium-, and long-term scales. Next, three LSTM subnetworks are constructed using these three-scale input features to perform predictions in a shared learning format, efficiently handling complex process data across different time scales. Consequently, the FS-MSTD-LSTM model for industrial soft sensing is developed. The performance of the proposed FS-MSTD-LSTM model is evaluated using two industrial datasets, and simulations on these datasets demonstrate that the FS-MSTD-LSTM model achieves higher soft sensing accuracy compared to other related methods, indicating its superior performance in managing complex process data with varying time scales.