A new Self-Organizing Extreme Learning Machine soft sensor model and its applications in complicated chemical processes

The control of product quality of complex chemical processes strictly depends on the measure of the key process variables. However, the online measure device is extremely expensive, and these devices are hard to protect. Meanwhile, there is a delay for these online measure devices. Therefore, the soft sensor technology plays a vital role in measuring the key process variables. Extreme Learning Machine (ELM) is an efficient and simple single layer feed-forward neural networks (SLFNs) to building an exact soft sensor model. However, unsuitable selected hidden nodes and random parameters will greatly affect the performance of the ELM. Therefore, this paper proposes a novel Self-Organizing Extreme Learning Machine (SOELM) algorithm constructed by the biological neuron-glia interaction principle to solve the issue of the ELM. Firstly, the weights between input layer nodes and the CNS are tuned iteratively by the Hebbian learning rule. Then the network structure is adjusted self-organizing by Mutual Information (MI) among different structures of networks. Secondly, the weights between the CNS and output layer nodes are obtained by the ELM. The experimental results based on different UCI data sets prove that the SOELM has a better generalization capability and stability than that of the ELM. Moreover, our proposed method is developed as a soft sensor model for accurately predicting the key variables of the Purified Terephthalic Acid (PTA) process.