Robust operating performance assessment of flotation processes using convolutional neural networks and feature learning

Abstract

The use of computer vision, rather than manual observation, to assess flotation performance based on froth characteristics is crucial for optimizing and controlling the flotation process. Convolutional neural networks (CNNs) are widely employed for image recognition tasks related to evaluating flotation operating performance. However, previous studies have often overlooked the quality of feature learning within these networks, resulting in limited robustness, especially when industrial applications encounter image distortions that challenge network performance.

To address this issue, this paper proposes a CNN-based algorithm for robust assessment of flotation operating performance, focusing on learning features that accurately reflect froth characteristics. The network is guided through regression training to prioritize froth-specific features, while classification training enhances its ability to evaluate flotation performance. Iterative optimization is achieved by adjusting the regression training loss using feedback from classification results and expert knowledge, thereby refining the network’s performance.

Experimental results from industrial applications validate the effectiveness of the proposed algorithm, demonstrating its ability to learn key features of froth images and showing high robustness under various types and levels of image distortion.