Adaptive soft sensor for multi-phase bath process based on GMM and online sequence ELM

The traditional soft sensor is global model and may be ill-suitable for processes with multi-phase and time varying batch process. In this paper, we propose a novel online ensemble extreme learning machine (OEELM) method to predict quality variable. Firstly, the Gaussian mixture model (GMM) is utilized to spilt entire phases of batch process. Furthermore, the multiple initial local ELM models are developed to capture the nonlinear relationships within different phases. Meanwhile, for any test sample, the posterior probabilities belong to different operating phase is obtained by Bayesian inference. Then the most relevant local ELM model is selected to the final prediction by maximum likelihood estimation. In addition, during online operation, the moving window technique is employed to update corresponding local ELM with the largest posterior probability. Based on simulated penicillin process data, the experiment result demonstrate that our proposed OEELM method not only has better prediction performance than other methods, but also computational load is light.