Improved conditional Gaussian regression soft sensor based on just-in-time learning

Data-driven online soft sensing is an important research direction in current industrial intelligent sensing. In the practical use of algorithms, process mode switching and data drift might reduce the performance of soft sensors. Traditional adaptive approaches confront limitations, such as a limited variety of models and a tendency to forget previously acquired modes. A sample temporal and spatial weighted conditional gaussian regression (STWCGR) soft sensor algorithm based on just-in-time learning is proposed to overcome these issues. This algorithm uses probability density and conditional probability for soft sensing modeling and prediction. First, a sample spatiotemporal mixed-weight technique is used to pick local modeling data in accordance with the just-in-time learning principle. Then, the local Gaussian probability density models are accumulated to fit the data distribution by incorporating the concept of Gaussian mixture regression. Finally, momentum updates and mode updates are introduced to enhance prediction stability and endow the model with adaptability to new working conditions. The efficacy of the suggested algorithm is confirmed by simulation studies with respect to forecast precision, stability, and flexibility to accommodate new modes. © 2024 Materials China. All rights reserved.