Data-driven Process Monitoring And Fault Diagnosis For Nonlinear Process Industries

Modern process industries are becoming increasingly large-scale,complex and integrated.The long-term safe,stable and efficient operation of industrial processes is the key to the realization of economic benefits for enterprises.The majority of controlled processes in process industries have nonlinear characteristics.Traditional linear methods usually assume that the processes work near a certain steady operating point,and the correlations among variables are approximately linear in a narrow window.The research of nonlinear process monitoring and fault diagnosis for process industry is a highly important subject in the field of automatic control.In the past two decades,nonlinear process monitoring has mainly been based on kernel methods,local models,and neural networks,etc.However,for process industries with large-scale systems,numerous measurement variables,multiple operating procedures,and complex correlations among variables,it is difficult to obtain complete monitoring and diagnosis results with traditional methods.The main research contents of this paper include the following aspects:(1)To address the problem of complex nonlinear process modeling and fault variables identification in the process industries,an integrated solution of process monitoring and fault isolation based on a variational autoencoder(VAE)model is proposed.Firstly,a variational autoencoder is used to construct a probability model for nonlinear process,which maps the process data with complicated distribution to the latent variable space to make it obey the Gaussian distribution.Secondly,a monitoring statistic is constructed based on the marginal joint probability density using the VAE model to monitor the nonlinear process.Then,the idea of missing value reconstruction is used to achieve the nonlinear fault isolation based on reconstruction contribution analysis.To avoid the smearing effect caused by the influence of multiple fault variables,the branch-and-bound technology is adopted to realize the fast search of multiple fault variable set and improve the fault isolation performance.Finally,the monitoring and diagnosis performance of the proposed method is verified by the data of the TE simulation process and a hot strip rolling process.(2)To solve the nonlinear problem caused by multi-operating conditions in process industries,a multi-operating process monitoring method based on Gaussian mixture variational autoencoder(GMVAE)is proposed.Firstly,the Gaussian mixture model is introduced into the variational autoencoder,and a GMVAE model is constructed to map the nonlinear process data under multi-operating conditions into the latent variable space,so that the latent variable projections of each operating condition approximately obeys Gaussian distribution;Then,the degree of deviation of the latent variable projection in each Gaussian component and the reconstructed probability distribution provided by the model are used to construct monitoring statistics to realize the non-linear process monitoring with the multi-operating conditions.Finally,the effectiveness of the proposed method is verified by the TE simulation case and a hot strip rolling case.(3)To address the problem of quality monitoring in non-linear process industry,a quality monitoring method combining deep variational information bottleneck(VIB)and VAE is proposed.Firstly,a joint model of deep VIB and VAE is constructed to decompose the quality-related and irrelevant information in latent variable space.The theoretical analysis is adopted to verify the decomposition capacity.Then,monitoring statistics are established by the latent variable distribution and the reconstructed probability distribution provide by the joint model to realize quality-related fault detection.Finally,the effectiveness of proposed method is verified by a numerical simulation case and a hot strip rolling case.(4)To solve the problem of nonlinear modeling and process monitoring in plant-wide processes,a plant-wide process monitoring method based on conditional variational autoencoder(CVAE)is proposed.Firstly,the sequential plant-wide processes,which is the common processes in practical process industries,are chosen as the research objects.A distributed conditional variational recurrent autoencoder(CVRAE)model is built to extraction the spatio-temporal features within and between the sub-processes.Secondly,the deep support vector data description model is used to extract the global features by fusing local spatio-temporal features.Then,the local statistics are built by the distributed model,and a global monitoring statistic is built by the fused features.By this means,a global-local process monitoring is realized for plant-wide processes.Finally,the proposed algorithm is verified by a hot strip rolling case.(5)To solve difficulty on the integrated solution of process monitoring,fault isolation and fault identification for nonlinear processes,a deep causality graph(DCG)modeling method and a process monitoring and fault diagnosis framework based on this modeling method is proposed.Firstly,a deep causality graph modeling method is proposed to construct a directed graph structure,which can describe the relationship among variables.Then,the conditional probabilities of process variables provided by DCG model are used to build the monitoring statistic and variable contribution indicators to achieve the fault detection and isolation.The directed graph structure is used to identify the root cause and propagation path of the faults.Finally,the proposed method is verified by the TE simulation process case.