AwsQgcNet: A Novel Soft Sensor Method for Quality Prediction Modeling in Suspended Magnetization Roasting Process

Quality prediction modeling is important when guiding and optimizing processes with advances in modern industrial technology. With the development of sensor technology and the complexity of production processes, deep learning-based soft sensor modeling methods enable the prediction of critical quality variables that are difficult to measure. Due to the time lag that exists in practical industrial processes, an appropriate selection of the lookback window can optimize prediction performance. In addition, quality-related features extracted by an upstream network can help downstream predictors perform better in terms of accuracy. Accounting for the above problems, the AwsQgcNet, which consists of feature extraction (Fe), adaptive window selection (Aws), and quality graph constraint (Qgc) module, is proposed for quality prediction modeling. The upstream feature extraction module is designed to extract features for the other two modules as the input. As a main branch of the framework, the adaptive window selection module is composed of multiple parallel subpredictors with different lookback windows, whose probability values are obtained by projecting the approximation errors. Then, the final prediction result is given by the mathematical expectation of all the subpredictors. The Qgc module works through its graph learning capability as an auxiliary branch. By setting proper quality graph labels, the Qgc module can force the network to extract quality-related features. A real industrial case of a suspension magnetization roasting (SMR) process verified the validity and superiority of the proposed method.