| Summary: | Just-in-time learning (JITL) has been used to construct soft sensor models online for its ability of handling strong nonlinearity and changes in processes. The most key procedure in JITL modelling is selecting relevant samples similar to a query sample. However, the common similarity criterions used to select relevant samples do not always function well for only considering the similarity of input data. Large noise or outliers in output data may result in inappropriate predictions of JITL based soft sensors. In this work, a combination of similarity measures, the conventional similarity of input and a novel similarity of output, is proposed for comprehensively understanding and selecting relevant samples. The effectiveness of the proposed soft sensor is demonstrated through an industrial fed-batch Erythromycin fermentation process.
|
|---|
