Research On Process Monitoring Method Based On Multi-source Heterogeneous Data

With the progress of intelligence,heterogeneous data such as video images relative to traditional physical variables also presents great application value for process monitoring.Therefore,through modeling learning methods such as multivariate statistics and machine learning,a model with stable operation and excellent performance can be learned collaboratively based on amount of historical data,which can be used to monitor industrial sites and help ensure the operational safety of industrial systems.This makes data-driven fault diagnosis methods have a bright future in security warnings in the era of big data.In this paper,the fault detection and diagnosis of magnesia smelting production process are studied.Aiming at how fusion of massive heterogeneous data collected in the smelting process is used for process monitoring,heterogeneous data collaborative modeling and semi-supervised classification learning of massive data for fault detection and identification are proposed.The contributions in this study are as follows.(1)Taking into account the manifold distribution of physical data while extracting image features,a collaborative feature regression model for heterogeneous data is constructed to ensure that the state information reflected by the sample data is unified,and then data information of physical variables and image features can be shared collaboratively for fault diagnosis.And,independent component analysis(ICA)is used to extract non-Gaussian feature information from video image features.Next,the remaining image residual information and physical variable samples are together analyzed and monitored by principal component analysis(PCA)to realize information shared collaborative modeling of heterogeneous data.The constructed process monitoring method with shared information based on collaborative feature regression has been verified the effectiveness in simulation experiments for monitoring smelting process of magnesia production.(2)Kernel independent component analysis(KICA)method is sensitive to traditional physical variables and the edge distribution characteristics of image data,but the detection accuracy is greatly reduced when applied to massive data with a lower proportion of labeled states.Due to the massive characteristics of multi-source heterogeneous data and the huge cost of marking data,a semi-supervised kernel independent component analysis method for fault identification is proposed by combining semi-supervised learning and KICA.In addition,ICA can remove strong correlations between variables and transform them into independent subspace,which helps to improve the separability and accuracy of classification algorithms.The semi-supervised KICA method performs a brief fault diagnosis based on the subordination probability of samples in each category,and then builds an operating state library to achieve accurate fault identification.Finally,the proposed semi-supervised KICA method is applied to the simulation experiment for monitoring the smelting process of magnesia production,which can effectively improve the identification degree and accuracy of smelting operating states.