基于时序图像深度学习的电熔镁炉异常工况诊断

超高温电熔镁炉（Fused magnesium furnace, FMF）生产炉况监测困难,易发生欠烧异常工况,不仅造成产品质量下降,也直接危害生产安全与人员安全.现有的人工巡检方式实时性差,容易发生漏报和误报,甚至导致铁制炉壳烧透、烧漏.针对该问题,本文采用视频信号,利用电熔镁炉欠烧工况的时空特征,即在炉壳表面出现的局部不规则高亮区域的空间特征,以及该高亮区域随时间呈现出亮度增强、面积变大的时序特征,提出一种基于卷积循环神经网络（Convolutional recurrent neural network, CRNN）的电熔镁炉异常工况诊断新方法.该方法包括图像序列一致性变换和时序残差图像提取预处理、基于卷积神经网络（Convolutional neural network, CNN）的空间特征提取、基于循环神经网络（Recurrent neural network,RNN）的时序特征提取、基于加权中值滤波的工况自动标记.最后采用实际的电熔镁炉炉壳的视频信号,进行了所提方法与现有的两种深度学习网络模型的实验比较研究,结果说明了所提方法的优越性. 

Ultra-high temperature smelting process of fused magnesium furnace (FMF) is difficult to monitor and prone to the semi-molten abnormal condition, which exerts severe influences not only on product quality but also on security of production and workers. Traditionally, the practioners have to reach out for inspection by watching the furnace shell. This is difficult in real-time and can cause missed alarms, wrong alarms, even lead to melting through the iron shell of the furnace. To solve this problem, the paper uses the information from in-situ video and takes advantage of the spatial-temporal features of the semi-molten abnormal condition, namely, the spatial feature appearing as irregular highlighted regions on the local furnace shell and the temporal feature appearing as increasing brightness and increasing areas of those regions. Based on the spatial-temporal features, this paper proposes a novel convolutional recurrent neural network (CRNN)-based method for the abnormal condition diagnosis of FMF. The method is composed of three modules: a preprocessing of the original video, a convolutional neural network (CNN)-based spatial feature extraction and a recurrent neural network (RNN)-based temporal feature extraction and a weighted median filter-based automatic labeling algorithm. By applying videos from the furnace shell of a real FMF, the proposed method is tested and compared with other two deep learning-based baseline approaches, and the results demonstrate the superiority of the proposed method.