Online Adaptive Modeling Framework for Deep Belief Network-Based Quality Prediction in Industrial Processes

Soft sensors have played increasingly important roles in modern industries for process monitoring, control, and optimization. Deep learning has shown great capacity for complex nonlinear modeling in soft sensors. However, it suffers from performance degradation due to the process time-varying problem. To deal with this problem, an adaptive updating strategy is proposed for deep networks in this paper, which is based on an online adaptive fine-tuning of deep belief network (OAFDBN). In OAFDBN, an initial DBN model is first trained by offline pre-training with raw input data and fine-tuning with labeled data in the historical data set. For the online prediction, the DBN model is fine-tuned adaptively upon each query sample. To predict the output of each query sample, the most relevant samples are then selected from the historical labeled data set, which is dynamically augmented by adding newly available labeled data into it. After that, the DBN model is fine-tuned online by a few iterations with the most relevant samples, which can keep timely and accurate tracking of data patterns. To verify the performance of the OAFDBN-based modeling framework, it is applied to an industrial hydrocracking process and a penicillin fermentation process.