Deep learning based self-adaptive modeling of multimode continuous manufacturing processes and its application to rotary drying process

Real-time prediction of future process outputs is critical for the model predictive control of continuous manufacturing processes. It helps identify when and how to adjust the process variables under the disturbances. A lot of recurrent neural network-based predictive models have been developed and validated on the simulated processes. Based on that, some works further consider the existence of multiple operating conditions that can be unforeseen and transitive in real-world manufacturing. However, their designed online learning mechanisms mostly focus on fast local tracking without preserving important old knowledge. Besides, the proactive input data adaptation is largely unexplored. To bridge this gap, a novel self-adaptation mechanism is proposed in this paper. This mechanism can be easily integrated into different choices of predictive model to improve the stability of performance towards various changes in a long period of manufacturing. In the proposed mechanism, the components of adaptive sequence filtering and adaptive input normalization first extract the compact and properly scaled features from the growing multivariate time series subject to delayed output response and non-stationarity. Based on the encoder-decoder network as an exemplary predictive model, the component of adaptive model update consists of a non-Euclidean loss for evaluating sequential predictions and a task-free knowledge consolidation strategy for continual learning-based regularization. The application to an industrial rotary drying process is demonstrated, where data streams are collected from four production lines over 14 months. Extensive comparative study shows the superior performance of proposed mechanism and ablation study further verifies the effectiveness of each individual component.