Revolutionizing Flotation Process Working Condition Identification Based on Froth Audio

Abstract

Effective and timely identification of working conditions is critical for ensuring safe and efficient long-term operations in industrial flotation processes. Previous studies on the flotation process have mainly focused on the froth images while neglecting the potential value of the audio derived from the froth. The audio generated during froth falling from the scraper to the bottom of the flotation cell can reflect the changes in flotation working conditions. This study innovatively uses froth audio to identify the working conditions of the flotation process, which is a new attempt and prospect in this field. First, to tackle the problem of difficult classification of froth audio categories, this study proposes a dual-channel Mel spectrogram (DMS) method to enhance audio recognition capabilities. It utilizes the attention mechanism to efficiently find and amplify the critical frequency bands while mitigating the impact of irrelevant frequency bands. Further, this study proposes an iterative domain adaptation (IDA) model to address the dataset shift issue arising from the temporal characteristics of industrial processes. By gradually assigning pseudo-labels to the target domain data and then fitting marginal and conditional distributions on both source and target domains, the improved generalization performance of the model is achieved. Extensive experiments on a real-world flotation process dataset demonstrate the effectiveness of the proposed method.