摘要
The operating conditions for industrial batch production often cover a wide range in order to produce different products. Inconsistent working conditions and recipes may change the data properties, but the generated batches may share similar mechanisms in terms of their qualitative and quantitative knowledge domains. In this study, we propose a transfer learning framework for both domains to improve the efficiency of monitoring in similar batch scenarios. First, a statistical pattern clustering strategy is developed for assessing and separating similar conditions. Based on this strategy, the phase-based generalized Procrustes analysis and ordinary Procrustes analysis are proposed to produce the nominal representations and also to transfer quantitative knowledge by accommodating batch-wise and recipe-wise discrepancies. Furthermore, a multiphase Bayesian network is constructed for qualitative knowledge transfer and statistical modeling with the nominal representations. Finally, a systematic monitoring flowchart is established for fault detection and isolation based on a just-in-time transfer strategy. Under this framework, the efforts required for similar process modeling can be reduced and the monitoring efficiency can be improved. The feasibility and effectiveness of the proposed diagram for industrial uses are validated on a fed-batch penicillin fermentation process.