An On-line Extreme Learning Machine with Adaptive Architecture for Soft Sensor Design

Soft Sensors (SSs) have been widely investigated and employed as inferential sensing systems for providing on-line estimations of industrial processes' variables. However, industrial processes suffer from different complex characteristics (e.g. time-variance and non-linearity), being very difficult for the SS models to perform well over time. This paper proposes a SS model using an on-line Extreme Learning Machine (ELM) with Directional Forgetting Factor (DFF) which is able to provide online estimations of variables in industrial processes. The main contribution is that the proposed ELM model has the ability of adapting its architecture over time. For this purpose, it is used the Bordering Method and the Reverse Bordering Method. Experiments demonstrate the performance of the proposed method over the state-of-the-art methods.