Research On Fault Detection Method Of Chemical Process Based On Network Analysis

With the trend of large-scale structure and complex in modern chemical process,it's becomes more and more difficult to describe the system operating status and fault conditions through the traditional way.In addition,it sets obstacles for fault detection research on the chemical process system due to the characteristics of the equipment hardware in chemical process itself,data collection and handling,internal strong correlation in chemical systems and the factors of transmitted,random,cascade in process failure.At the same time,complex network have been applied in many areas and achieved fruitful results because of its advantages in the study of complex network structure and complexity.The paper carries the research of fault detection based on the complex network theory.It considers both the local attributes of the system and overall structural features,what is the traditional fault detection method can't be achieved,and has contributes the ideas in the field of fault detection based on complex network.The main research contents include:(1)Carry out research on the complex network theory and the visible graph method,explore its feasibility in the analysis of data,build data network model through the corresponding way,then research fault detection method in a new perspective.(2)Proposed a complex network fault detection method based on cosine similarity.View the cosine similarity as the correlation,and then the network model is constructed by adjacency matrix.Finally,detect fault according to the difference of fault status and no fault for the network topology and measure of the system.And validate the effectiveness and superiority by examples for TE process.(3)Because the method in(2)involves in the threshold setting,and has no consideration of the timing characteristics of chemical system data in detail,it proposed a fault detection method based on horizontal graph analysis.The data of each variable in the system is regarded as a time series.Each time series is modeled as a network by horizontal view method,then treat each single-layer network as a node,and the correlation between the mololayer network is used to characterize the correlation between the corresponding nodes.According to the variance of the obtained correlation matrix from the fault state and the normal state,we can confirm the faulty node.It avoids the interference of threshold factors and analyzes the timing characteristics of data,and validates the effectiveness and feasibility by the TE process.