Abstract

In this work we focused on the development of an adaptive Soft Sensor which may be deployed in a real-life environment, for example as inferential control support. To be able to do this, the Soft Sensor must fulfil certain constraints like being able to deal with data impurities or to adapt itself with changing data. The task is approached by training a set of models with limited validity in the data space and by proposing a statistically-based technique for the combination of the local models. The combination weights are related to the estimated performance of the local models in the neighbourhood of the processed data sample. The performance and other benefits of the proposed Soft Sensor are demonstrated in terms of a case study where the model deals with raw industrial data.