A Feature Extraction Strategy of Fire Hole Video Based on VGG16 and Migration Learning

Abstract

The superheat degree (SD) is the comprehensive performance of cell condition. In the aluminum electrolysis process, identifying the state of SD plays an important role in improving the current efficiency. The existing methods are usually based on the artificial experience combining with video of the fire hole video, which including artificially extracting features and then classifying them via machine learning algorithms. However, because of dwindling and frequent change of experiential technicians, some features being difficult to extract manually. Therefore, the existing methods do encounter the bottleneck to ensure the classification accuracy of SD states. In this paper, a strategy that extracting video features of fire holes based on convolutional neural network is proposed. The proposed strategy effectively integrates VGG16 and migration leaning to selected useful features. Simulation results validate that the proposed method is able to dig out the deeper features and make a more accurate judgment on the state of SD.