A Two-step Feature Selection Algorithm based on Mutual Information for Industrial Soft Sensor Applications

The development of data-driven industrial soft sensors is susceptible to encountering datasets that contain unnecessary, excessive, and redundant information. This study proposes a two-step feature selection method, where mutual information is used in the initial screening phase to select features with strong mutual information, followed by a secondary screening phase that considers both mutual information and redundancy index. This approach ensures the accurate selection of a streamlined feature set that preserves essential information while discarding redundant variables. Subsequently, the essential features obtained from the two-step feature selection algorithm are employed to develop an attention-LSTM network for industrial soft sensor applications. The efficacy and precision of the suggested approach are validated in an industrial hydrocracking process.