Fault Diagnosis Of Polyester Fiber Polymerization Process Based On Graph Structure

The polymerization process is the first stage in the polyester fiber production process and a key stage that determines the quality of production.The polymerization process not only involves complex physical and chemical reactions,but also is sensitive to changes in external environment.Therefore,there is a lack of effective and comprehensive understanding of the internal mechanism of the polymerization process,which makes fault diagnosis difficult.The graph structure is an effective way to express the relationship among the variables in the polymerization process.The variables in the polymerization process are regarded as nodes in the graph,and the graph adjacency matrix is used to describe the relationship among the nodes.The adjacency matrix of the graph does not rely on expert knowledge in most cases,but is learned in a data-driven manner.Therefore,based on the graph structure,this paper studied and established the fault diagnosis model of polyester fiber polymerization process from three aspects: anomaly detection,the root cause analysis and fault propagation path identification,to assist the smooth operation of the polymerization process.The research work of this article includes the following aspects:(1)For the problem of polyester fiber polymerization process data containing many variables and high coupling between variables,based on the graph structure,a normalized flow model anomaly detection model with synergistic enhancement in time and space is proposed(TSANF).The model obtains historical time information through long and short-term memory networks,uses the adjacency matrix of directed graphs to obtain the dependencies between variables,and enhances the normalized flow by combining temporal and inter-neighborhood information together through graph convolution.The experimental results demonstrate that TSANF has better anomaly detection results compared with the baseline method.(2)For the polyester fiber polymerization process data nonlinearity,fault continuous occurrence and continuous propagation characteristics,a graph structure-based nonlinear continuous time-invariant network fault root cause ranking method is proposed(FRCR-NCTIN).The method designs a point-to-point bidirectional long and short term memory(PP-BLSTM)network model that combines smoothing error and dynamic thresholding to achieve tracking and continuity of nonlinear disappearance of constant relations.Meanwhile,the information of dynamic propagation of faults with industrial processes is extracted by calculating the transfer entropy,which makes the root cause ranking based on network diffusion and reconstruction of vanishing invariant relations more accurate.Experiments show that the accuracy of root cause analysis can be improved by comparing the proposed method with the baseline method.(3)To address the problem of closed-loop fault propagation path caused by the negative feedback regulation mechanism of polyester fiber polymerization process,a fault propagation path identification method is proposed based on K2 graph structure learning algorithm inspired by the negative feedback mechanism(FP-NFK2).The method obtains the significant fault propagation likelihood by generating a sequence of high and low alarms in the direction of their occurrence and by transmitting entropy and significance thresholds.Meanwhile,the causal confidence among variables is obtained in the fault-free data and combined with the important fault propagation possibility to obtain the final fault propagation possibility.Then the de-cycling is performed and the K2 preparation order is obtained,and finally the propagation paths of faults are identified by the K2 algorithm.Experiments show that FP-NFK2 has better performance in fault path identification.