Hessian Regularization Semi-supervised Extreme Learning Machine for Superheat Identification in Aluminum Reduction Cell

摘要

Superheat degree identification of aluminum electrolysis cell condition has long been a challenging industrial issue due to complex and hard environment and limitation of physical devices. In addition, the degree of superheat is an influential index of the whole process, reflecting the distribution of the physical field of the electrolytic cell, the current efficiency in the electrolytic production, and the lifespan of the aluminum reduction cell. Traditional measuring method of SD is device-based which is inaccurate and with expensive cost. In this paper, we propose data-driven semi-supervised model, semi-supervised extreme learning machine (SS-ELM) for SD detection. Additionally, Hessian regularization was utilized as a regularization term to improve the model performances. The proposed HRSS-ELM was verified with industrial dataset in aluminum reduction cell. The experiment demonstrates HRSS-ELM has a maximum performance compared with expert rules and other methods, and the accuracy of test results is up to 85%.