Accurate prediction of mill load parameters is crucial to improving grinding efficiency and saving energy. Traditional prediction models have challenges such as poor interpretability, low prediction efficiency and differences in data distribution. This study innovatively proposed a multi-task prediction model that integrates physical information and domain adaptation. By constructing a physical-data-driven hybrid model, the physical relationship between mill load parameters is embedded into the model as prior knowledge to improve the prediction accuracy of the model. At the same time, multi-task learning is used to predict the material-to-ball volume ratio and the pulp density at the same time, which improves efficiency and reduces repetitive work. The domain adaptation method is introduced to ensure that the model maintains stable prediction performance when the data distribution changes. Laboratory ball mill data verification shows that the proposed model not only improves the prediction accuracy, but also adapts well to variable working conditions, showing significant superiority.

The problem of misalignment of the original measurement model is caused by nonlinear, time-varying characteristic of the batch process. In this paper, a method based on geodesic flow kernel (GFK) for feature transfer is proposed. By mapping data into the manifold space, the feature transfer from source domain to target domain is implemented. Distribution adaptation of real-time data and modeling data is performed to reduce the distribution difference between them. The historical data through distribution adaptation is used to establish a regression model to predict the real-time data, by which the unsupervised batch process soft sensor modeling is realized. The application of predicting the concentration of penicillin between different batches during the fermentation of penicillin demonstrated that the prediction accuracy of the model can be improved more effectively than the traditional soft sensor method.