An ensembled multilabel classification method for the short-circuit detection of electrolytic refining

Abstract

Short-circuits occurring in the electrolytic refining process of non-ferrous smelting are a main factor that consumes extra energy and affects the metal quality. This paper proposes an ensembled multilabel classification method for short-circuit detection based on infrared images and makes up for the defect of previous methods using object-detection neural networks being hard to directly apply in industrial sites. Different from the object-detection methods, the multilabel classification method does not output the imaging positions but directly obtains the realistic positions, i.e. plate numbers, of the faulty plates. By introducing a new convolutional neural network named FlatNet, no extra work is required to get the realistic positions of the faulty plates. To address the data imbalance inherent to multilabel classification, dynamic weights that pay more attention both to the minority class and difficult samples are presented, forming a bilateral constraint on the missed and the false detections. At the end of the method, we design a greedy ensemble approach driven by validation F1-scores for the promotion of detection performance and stability. Experiments conducted with real-world data verify the effectiveness of the proposed fault detection method.