| Fused magnesia is a kind of important refractories,widely used in aerospace,metallurgy,chemical industry and other fields.The production of fused magnesia mainly uses three-phase alternating current fused magnesia furnace to melt MgCO3 into magnesia crystal and CO2 under ultra-high temperature.Due to the complex composition of magnesite ore and the fluctuations of raw materials,the abnormal working condition represented by semimolten working condition is easy to occur,which directly affects the product quality,and even causes malignant accidents such as furnace wall penetration and molten liquid leakage.It threatens the safety of personnel and production,so timely diagnosis is necessary.However,due to the extremely high temperature of 2900℃ during melting,it is difficult to detect directly.At present,the diagnosis of semimolten situation in the production site is mainly judged by the regular inspection of human eyes combined with the production experience.However,due to the poor production environment and the untimely inspection,missed inspections are prone to occur.For this reason,the diagnosis of abnormal conditions has important practical research value.It is difficult to adopt the model-based fault diagnosis method in view of the complex melting mechanism of fused magnesia furnace,and it is difficult to adopt the data-based fault diagnosis method in view of the strong interference of melting process data.As a result,based on the existing fault diagnosis methods and the research status of abnormal working situation diagnosis of fused magnesia furnace,this paper proposed a new method,time-series imagedriven diagnostic method.The main results are as follows:1.According to the change of image brightness and the interference of water mist,a dynamic image preprocessing method combining gray normalization and moving median average filtering is proposed.First of all,aiming at the image brightness change caused by the furnace mouth flame beating in the smelting process,a sequential image preprocessing method based on gray-scale normalization is proposed;secondly,aiming at the influence of water mist interference brought by the cooling system in the smelting process,the image after gray-scale normalization is processed by the sliding median average filtering method.2.According to the characteristics of different temperature distribution in different melting areas of fused magnesia furnace and the change of semimolten area position with size and brightness,a diagnosis method of multilevel dynamic principal component analysis(MLDPCA)is proposed.The proposed method exploits the spatial and temporal characteristics of temperature fluctuation in fused magnesia furnace in normal situation as well as the partial glowing of the furnace wall and continuous expanding of the glowing area in abnormal conditions.In order to extract these spatial and temporal features,a new multi-level dynamic principal component analysis algorithm is developed for the dynamic images.A hierarchical monitoring method is then proposed to perform the abnormality diagnosis and locate the abnormality by using the MLDPCA based contribution plot.3.According to the difference between the image features of semimolten situation and normal working situation of fused magnesia furnace,a diagnosis method of abnormal working situation of fused magnesia furnace based on time and space feature extraction of principal component analysis network(PCANet)of dynamic image is proposed.First,the dynamic image is constructed from the preprocessed time sequence residual image;then,the time sequence spatial features of the dynamic image are extracted by PCANet;then,the time features are extracted by the method of dynamic correlation degree,and the time of semimolten situation is diagnosed according to the dynamic correlation degree.In this paper,two kinds of semimolten situation diagnosis methods proposed in this paper are compared and studied by using the real images collected in the smelting process of an electric fused magnesia plant.The results show that the two methods can diagnose the semimolten situation and have advantages in rapidity and situation location.Among them,the combination of PCANet and dynamic correlation has better rapidity and stability.The multilevel dynamic principal component analysis method can locate the position of semimolten situation according to the contribution of sub block pixels. |
