Domain-adaptation with knowledge accumulation through parallel stacked autoencoders: methodology and application to sulfur recovery

Data-driven predictive models for end-point quality variables are important tools in industrial process modeling. However, establishing an effective predictive model with limited labeled data remains challenging. Transfer learning (TL) offers a solution by leveraging knowledge from similar but different tasks. This paper introduces a novel TL-based predictive model, domain-adaptation parallel stacked auto encoders (DA-PSAE), which can extract and accumulate knowledge from multiple similar processes. First, a parallel model structure is designed for the simultaneous extraction of static and plastic latent features. Besides, an effective TL-based training strategy is proposed, which utilizes data from multiple similar processes. The proposed model is applied to a sulfur recovery unit composed of four parallel sub-units. Experimental results verify the effectiveness of the proposed model.