Quality Trend Evaluation Based on Gated Stacked Variable-Wise Weighted Quality-Related Autoencoder for Flotation Process

To save energy, reduce consumption, and improve the efficiency of enterprises, it is necessary to timely detect key quality indicators and evaluate the quality development trend of the industrial productions in real time. With the development of industrial internet and industrial big data, deep learning methods represented by stacked autoencoder (SAE) are widely used. However, traditional SAE-based quality prediction models lack the ability to extract quality-related features and only use the feature of the last hidden layer for the final prediction, discounting its predictive performance. In view of this, a new gated stacked variable-wise weighted quality-related autoencoder (GSVWQAE) is proposed to improve the modeling performance in this study. Through correlation analysis with quality indicator, important process variables are identified in the input layer of each variable-wise weighted quality-related autoencoder (VWQAE), and different variables are given different weights. Subsequently, incorporating the prediction error of the quality indicator into the loss function facilitates the utilization of quality-related information to guide feature extraction. Moreover, the integration of gated neurons enables the comprehensive exploitation of distinct hidden layer features, quantifying their contributions and regulating the information flow from various levels for the ultimate quality prediction and trend evaluation. The proposed method's effectiveness and feasibility are ultimately confirmed through verification in the flotation process.