Data-driven Dynamic Process Modeling And Monitoring

Data based models are used to mine and explain process behaviors.On the other hand,process knowledge can provide much meaningful information for constructing data based models.No matter how large the system is,what the degree of noise level and linearity are and whether the process is running under closed loop control,dynamics is the most fundamental attribute for process industries.Hence,it is important and promising to model process dynamics and monitor dynamic processes.Reviewing the researches related to dynamic process monitoring,the existing methods are mainly under the assumptions of known mode information,open loop processes,shallow or simple nonlinearities.In face of many applications,these assumptions do not live up to the real situations.Motivated by requirements from industrial applications and based on some existing approaches,the following issues are studies and solved in this dissertation.(1)For multivariate time series with miscellaneous patterns,also known as multimode data without any priors of mode changing points,dynamic mode segmentation,dynamic performance assessment and improved control performance diagnosis are developed.To segment different modes from the raw data,a new metric measuring two dynamic models is designed and interval halving algorithm,together with moving window is used to locate the mode changing points quickly and efficiently.Especially,the metric is based on the gap between two dynamic models,in place of some conventional indices like means or variance/covariance.The dynamic distance can better represent the whole process features so that the segmentation will be more accurate.Then,for the segmented modes,dynamic performance is concerned to discover the optimal mode.Finally,a new performance diagnosis algorithm based on sparse contributions is improved to figure out the root cause of performance change.Compared with the conventional contributions,the sparse contributions are able to suppress useless correlations causing some misdiagnosis.(2)The adverse effect of feedback control on data-driven process monitoring approaches is systematically analyzed,and some advanced fault diagnosis strategies are developed for closed loop systems.Specifically,taking dynamic principal component analysis for example,its performance of fault detection and fault reconstruction is researched.The performance degradation is proved when the closed loop data is directly applied to data-driven monitoring models.To solve these problems,a feedback invariant based sensor fault identification strategy is proposed when process delays are available.For general faults,output oversampling is proposed to recover the ability of fault reconstruction for closed loop systems.In addition,a special case of output oversampling,known as shifted sampling,is proved to be able to guarantee the identifiability for closed loop systems without external probing signals.Hence,the shifited sampling can also be applied to closed loop process monitoring.On the one hand,reconstructed-based contributions are improved under feedback control by the output oversampling scheme.On the other hand,contribution of variables can only provide the information related to variables,which is insufficient to figure out the root cause.Therefore,a fault location method for clarifying the abnormal devices is proposed to enrich the diagnosis information.(3)For complex nonlinear dynamic processes with uncertainty,a deep learning based method is proposed to learn stochastic nonlinear state space models,applied to fault detection.Both state transition and observation generation from states are implemented by deep neural networks,for fitting strong nonlinearities.A forward-backward recurrent neural network is used to generate the distributions of filtered states and smoothed states.Moreover,expectation maximization can be used to train these neural networks with gradient back propagation.This kind of deep learning model can integrate the process dynamics,nonlinearities and uncertainties.Compared with other shallow models,this method can deal with more complex situations.With the recurrent neural network,sequential sampling approaches with high computational costs for estimating states are avoided.And the corresponding monitoring procedures are designed with the learned model.(4)A complete monitoring procedure is designed for multimode and multiphase/multistage dynamic batch processes.Firstly,a state space model based phase division is proposed to extract phase models for each batch.Then,using the distance metric between two batches,different batch modes are clustered with K-means algorithm.For the same mode,phase re-identification technique based on expectation-maximization algorithm for multibatch data is developed.Moreover,in contrast with the conventional multivariate statistical methods,the proposed methods can directly deal with batches with uneven length because the state space model is a structural model described by several parameters.With the learned dynamic models in each phase for each mode,monitoring statistics can be designed with proper mode determination and phase switching strategies.