Research On Fault Isolation Method Based On Structured Sparsity Model

With the increasing complexity of modern industrial process,the safety and reliability of system has become more and more concerned.Once an abnormality occurs in the large-scale production process,it is necessary to determine the fault variables quickly and accurately to help the operator isolate the fault area and correct the faults.In recent years,with the continuous development of information technology and sensor technology,research on data-driven fault isolation method has received extensive attention.Aiming at the problem of fault isolation in industrial process,this paper proposes several fault isolation methods based on structured sparsity to make full use of expert knowledge in the process.Compared with traditional methods,the methods based on structured sparsity can use process information and obtain sparse fault isolation result,which is beneficial to improve the accuracy of fault isolation and the interpretability of result.The specific research work of this paper is as follows:(1)Fault isolation methods based on partially known sparse support and block sparsity are proposed.During production,operators usually have high confidence that some subsystems or variables are abnormal,which can be described by partially known sparse support.On the other hand,correlated variables in the system often have simultaneous exceptions,which can be described by block sparsity.By introducing partially known sparse support and block sparse regular term into the fault reconstruction objective function of multivariate statistical analysis,two new fault reconstruction methods are designed.Compared with traditional methods,the application results in simulation and coal-fired power plant show that the new fault reconstruction methods can effectively improve the fault isolation effect.(2)A fault isolation method based on tree-structured sparsity is proposed.In industrial process,some subsystems have obvious hierarchical relationship between upstream and downstream,which can be described by tree-structured sparsity.A new fault isolation model is established by introducing a tree structure sparse regular term into the fault reconstruction objective function.The application results in simulation and coal-fired power plant show that the tree structure sparse method can effectively identify fault variables and fault propagation path,which provides help for subsequent fault source diagnosis.(3)For dynamic multivariate processes,a fault isolation method based on Bayesian block sparse reconstruction is proposed.This method adopts a blockstructured single measurement vector model to consider the time correlation of variables,and then the Bernoulli-Gaussian probability model is used to promote the block sparsity of the index matrix to quantify the fault contribution of process variables,finally,the fault variables are determined according to whether the fault contribution is zero.The application in distillation unit shows that this model is more effective than the traditional method in dealing with fault isolation in dynamic process.