A Novel Soft Sensor Model for Alumina Concentration Based on Collaborative Training SDBN

To address the challenge of insufficient labeled sample data and the difficulty in obtaining labeled samples in industrial processes, we propose a co-training based Soft Deep Belief Network (SDBN) model for soft measurement, named Co-training SDBN (Coreg-SDBN). The original labeled samples are evenly divided, and two initial SDBN models are established based on the two sets of labeled samples respectively. These models are then used to progressively augment the dataset with unlabeled data, eventually leading to the development of two independent models on the augmented dataset. For online predictions, the outputs of the two final models are integrated to produce the prediction output. The model's effectiveness is demonstrated through simulation validation on aluminum electrolysis industry data, showing that the inclusion of a large volume of unlabeled samples can significantly enhance the performance of the SDBN model compared to both SDBN and DBN alone.