Unsupervised Domain Adaptation for Grade Prediction of Froth Flotation Based on Wasserstein Distance and Transformer

In a zinc flotation process, the concentrate grade is an important indicator which cannot be measured online. To ensure the stability of the process, deep learning has been widely used for the prediction of concentrate grade. However, grade prediction based on deep learning leads to the dataset bias problem, which caused by illumination change, environment noise, etc. A popular network for addressing this problem is the domain adversarial neural network (DANN), which cannot predict the domain label according to data distribution and extract global information. To overcome these limitations, we propose a new method, the wasserstein-transformer domain adversarial neural network (WT-DANN), which uses transformer to extract global features and calculates domain loss with wasserstein distance. We evaluated our approach on public dataset (Office31) and flotation dataset. On the Office31 dataset, WT-DANN achieved a better cross-domain accuracy than that of DANN. For the flotation dataset, WT-DANN's cross-domain prediction accuracy was 60.7%, compared to DANN's accuracy of 51.2%.