Multi-task Learning Based Fault Detection Method For Fused Magnesia Industrial Process

With the rapid development of modern industry,the structure of industrial production equipment has become more complex,more automated,and more informative.In the industrial production process,a large number of real-time operational data of industrial production processes can be obtained through various types of sensors.In a large amount of data,there is a wide variety of information in industrial production.In the field of industrial safety production,it is possible to timely report the possible failures in the production process by modeling the massive data to achieve safe production of industrial processes.Therefore,data-driven fault diagnosis methods are widely used for fault warning and diagnosis of industrial processes.Based on the work of the predecessors,this paper develops a multi-task learning multi-view video data collaborative modeling monitoring method and multi-task learning heterogeneous data collaborative modeling monitoring method.The main research work is as follows:Considering the problem that the single-view video surveillance is not comprehensive in the industrial production process,the camera is arranged at multiple angles in the industrial site.In order to fully exploit the related information between the multi-view video data and realize the comprehensive monitoring in the industrial production process,this paper proposes based on multiple Task learning multi-view video data collaborative modeling monitoring method.For the high dimensionality of video data,this method combines data dimensionality reduction with dictionary learning,and implements sparse coding of video data,which greatly reduces computational complexity.Considering the relevance of multi-view video data,this method introduces the 2,1 norm to achieve the coupling feature selection of multi-view video data,and the shared subspace learning for sparsely encoded data.The method fully exploits the information between the multi-view video data and improves the accuracy of the industrial live video monitoring.Due to the particularity of the fused magnesium oxide production process,in the actual production process,there are not only image data but also various physical variable data such as current data and voltage data.These data also have important guiding effects on the production process.Only by modeling the image data can only reflect the surface condition of the fused magnesia production process,which ignores the influence of other data,lacks a deeper understanding of the smelting process,and makes the detection process become Not accurate enough.In view of the important guiding role of current data in the production process,it is necessary to model the image data and current data collected during the production process of fused magnesia to improve the detection accuracy.Considering the heterogeneity of video data and current data,this paper proposes a multitask learning heterogeneous data collaborative modeling monitoring method.Considering the contribution of different modal data,an empirically based dynamic combination coefficient is set,and heterogeneous data of different modalities are returned to the space composed of the same label,taking into account the deeper correlation between heterogeneous data,introduced a low rank constrained norm.The method can effectively combine image data and current data,and improves the accuracy of fault monitoring.Finally,the proposed algorithm is applied to the simulation of fused magnesium oxide process,and their fault detection performance is tested.The feasibility and effectiveness of the proposed method are verified.