A novel semi-supervised prediction modeling method based on deep learning for flotation process with large drift of working conditions

Deep neural networks have been broadly utilized for soft sensing modeling for the process performance which is significant for process control but cannot be measured online. However, the popular deep learning models still cannot adapt to large drift of working conditions in the process industry, which causes the model accuracy to become worse and worse with the time go on. Moreover, the cost of acquiring sufficient labeled data is very high. Therefore, in this study, a semi-supervised deep learning method called dynamic multi-scale selective kernel network (DMS-Sknet) with novel loss function is proposed by taking the flotation process as the case. In DMS-SKnet, multiscale features are extracted from froth images by using multi-scale dilated convolution kernel, and then fused with other process data in time series. A channel attention module with soft attention is designed to learn the important relationships between multi-scale feature maps and process features. Finally, based on the semi-supervised Mean-teacher (MT) learning framework, a new loss function is proposed, in which temporal distance is considered to improve the generalization ability and the long-term accuracy of the network. The experimental results using industrial flotation process data show that this method can effectively improve the grade prediction accuracy after a long period of significant changes in the working conditions.