Operating Performance Assessment Based on Stacked Sparse Denoising Auto-encoder for Gold Hydrometallurgy Processes

Abstract

The process operating performance assessment (POPA) is critical for industrial processes to pursue optimal comprehensive economic benefit. In this article, a two-stage deep unsupervised feature learning approach for the industrial POPA is proposed. In the first stage, we utilize stacked sparse denoising auto-encoder (SSDA) to extract deep features from raw input data with noise of different performance grades, which can overcome drawbacks of traditional approach and automatically extract features from process variable correlation characteristics. In the second stage, softmax regression is employed to train a neural network classifier for the features of different performance grades. The presented method is illustrated by a gold hydrometallurgy process. The simulation results show that the proposed SS-DA method obtains fairly high assessment accuracy and strong robustness than the total projection to latent structures (T-PLS) method even under strong noise interference environment.