Technology And Software For Identifying Underfeeding Condition In Magnesium Furnace Based On Semi-supervised Learning

In China,high-grade fused magnesia is mainly obtained by melting and recrystallization in arc furnace.In the smelting process of electric magnesium furnace,it is necessary to feed raw material at an appropriate time,but the fixed feeding cycle does not guarantee that it's the best time to feed.Therefore,it often requires workers to observe the flame of furnace for a period of time to find out the underfeeding situation.At this time,through manual intervention,workers can feed magnesia in advance.However,considering the frequent missed inspection of underfeeding condition and the labor intensity is high.Enterprises urgently need the technology of automatic identification of underfeeding condition.In the smelting process of fused magnesium furnace,the color and shape of the flames under different working conditions are significantly different,which makes it possible to automatically identify the underfeeding condition through images.Most of the existing image-based working condition discrimination methods are based on single-frame image,or use static image features as input,and then identify the underfeeding working condition.The problem is that the above methods doesn't take into account the time series information,and it is not convincing to use several seconds of image features as all the features of the current moment to discriminate the working condition.And all the above studies are carried out under the condition that there are a large number of tagged videos by default,but tagged videos are not easy to obtain.In summary,the main challenges of intelligent identification of underfeeding condition are as follows:1.The flame of the smelting process is dynamic,the normal melting and underfeeding conditions can't be accurately distinguished only by the brightness and morphological characteristics of the flame in a single frame image.This requires simultaneous analysis of image and time dimensions.2.Although a large amount of video can be collected at the scene,the working conditions of the video need to be marked by experienced workers,which is cumbersome and time-consuming.So usually only some of the videos are marked.Therefore,it is necessary to recognize the underfeeding condition use a small number of labeled samples.To solve the problems of identifying the underfeeding condition of magnesium smelting furnace,this paper proposes a semi-supervised learning method to identify the underfeeding condition of magnesium smelting furnace from dynamic flames.The main work is as follows:1)In image preprocessing process,an adaptive segmentation algorithm of multi-dimensional time series is adopted to realize automatic segmentation of surveillance video,and the obtained segments are used as basic data units for training the classifiers,which solves the problem that the classifier's input data should based on time series.2)In the aspect of classifier design,under the self-training semi-supervised learning framework,an iterative classifier training method based on LSTM and sample label propagation is proposed,namely LSTM-KNN method,which solves the problem of training classifiers with less labeled data.A comparative study of algorithm performance with LCLC and KNN-SVM shows the effectiveness of the proposed method.3)A demonstration software for identifying the underfeeding condition of magnesium smelting furnace is designed and implemented.which realizes the model structure's selection,parameter's learning and result's verification in the algorithm training phase and online identification of the underfeeding condition of the fused magnesium furnace.