Nonlinear Process Monitoring Based On Auto-encoder Model

Recently,modern industrial processes are pushed towards complicate,automatic and digital direction.However,process safety is still the core target of process control systems,and process monitoring is the key part of process control.Due to the expansion of process complexity and scales,it is difficult for process monitoring to obtain the mechanism model.Meanwhile,with the development of distributed control system and sensor network,the process control system collects huge number of process data.Therefore,data driven process monitoring has become a popular research subject in both academic research and industrial applications.Over recent years,deep learning has become state-of-the-art area in a wide variety of fields,including image processing,speech recognition and natural language processing.It demonstrates the ability of processing nonlinear data correlation for deep learning methods.At the same time,how to mine the inherent knowledge for chemical process lies at the core problem of process monitoring.It highlights the importance of feature learn methods such as deep learning methods.As a typical deep learning method,the Auto-encoder model is introduced to process monitoring to sovle the nonlinear problem.Furthermore,the process to be monitored can be dynamic,multimode and so on.And fault detection,fault identification and fault classification are also part of the process monitoring.The research topic of this thesis is focused on developing Auto-encoder model for process monitoring based on different requirement.The main research area include these four parts:(1)Due to the shortcoming of conventional Auto-encoder model in fault detection,a Robust Auto-encoder model is proposed.The target of Robust Auto-encoder model is making the model output robust to the change in input during training procedure.Theoretical analysis shows that the Robust Auto-encoder model has better performance in fault detection area.The case study also demonstrates the advantage of proposed method.(2)The Back-propagation based contribution methods are proposed based on the fact that the conventional contribution plot and reconstruction based contribution methods cannot be used in the nonlinear model.The proposed fault identification methods focus on how to back-propagate the fault data to input variables.So the contribution of variable can be determined.The case study shows the efficiency of Back-propagation based contribution methods in fault identification area.(3)In order to solve the fault classification problem in the dynamic semi-supervised learning background,the Dynamic Sparse Stacked Auto-encoders model is proposed based on faulty data and normal data.The training procedure includes unsupervised pre-training and supervised tuning.The fault classification accuracy is significantly improved in TE process fault classification.(4)For those industrial processes which have multi-operating conditions,the 5-layer Stacked Auto-encoders model is introduced to do feature learnig.Then he extracted feature can be used for classify the multiple operational conditions.