An Industrial Process Monitoring Method Based on Entropy Projection Transformation Analysis

The feature extraction and selection of process variables and the calculation of monitoring statistics are inevitable problems in industrial process monitoring. A new process detection method is proposed in this article that makes full use of the historical information of process variables for feature extraction. First, a projection transformation grid is proposed for the feature extraction and dimensionality reduction of process variables. Using historical data to construct a joint matrix and paying attention to the correlation between process variables and the joint matrix, an entropy projection transformation component (EPTC) is extracted, and the feature weight is calculated to complete the extraction of the process variable information. On this basis, the feature differences of neighboring samples are further evaluated to select the optimal information entropy. In addition, a statistic based on smooth relative entropy (SRE) and the corresponding threshold calculation method are proposed, improving the model's sensitivity to early fault monitoring. Experiments demonstrate that incorporating optimal information entropy into industrial process monitoring can improve the reliability of fault detection compared with a traditional process monitoring model.