Study On Data-driven Process Montioring Methods For Nonstationary Chemical Processes

Petrochemical industry plays an important role in promoting national economic and social development,and the chemical process safety is an important prerequisite for industrial production.The process monitoring technique,as a powerful step to ensure process safety and product quality,has attracted extensive attention from both academia and industrial community.In addition,at the era of informatization,a large number of advanced artificial intelligence and big data analytic methods are constantly proposed,which greatly promotes the prosperous development of the data-driven process monitoring technique.Meanwhile,the large-scale and highly integrated chemical equipment significantly increases the complexity of chemical process systems.The complex characteristics of the system,such as high dimensionality,non-Gaussianity,nonlinearity,temporality,and non-stationarity,also bring huge challenge to the industrial application of the process monitoring technique.Especially,most traditional process monitoring methods are established under the assumption that the process is stationary,while the nonstationary characteristic of chemical process data is general and with multiple sources.The effectiveness of the process monitoring model could be degraded or even invalid if the nonstationary data characteristic of chemical industrial processes is not considered.For this reason,the extraction of nonstationary data characteristics has become one of the most popular and difficult research focuses in the field of process monitoring.So far,relevant researches have not explored the nature of nonstationary characteristics in chemical processes in depth,nor have they provided comprehensive solutions for the extraction of nonstationary characteristics from various sources.Aiming at this issue,this work carries out research on the extraction of nonstationary characteristics resulted from long-term deterministic trends,local random trends,and the adjustment of operating conditions.A series of data-driven nonstationary process monitoring methods based on multivariate statistical analysis and deep learning are proposed and applied to continuous processes with single mode,continuous processes with multiple modes,and batch processes,including the Tennessee Eastman process,an industrial-scale penicillin fermentation process and several chemical industrial processes.The main contributions and innovations include:Firstly,although traditional multivariate statistical process monitoring methods have strong interpretability,it is difficult to consider all kinds of complex characteristics in chemical processes at the same time,resulting in low fault detection rate and high fault detection delay.To deal with this issue,an orthogonal projection based statistical feature extraction method is proposed in Chapter 2 for continuous process monitoring.The proposed method monitors the statistics of orthogonal transformed components,which not only considers the extraction of non-Gaussian and nonlinear features of data,but also greatly reduces the dimension of data.Regarding the selection of monitoring features,a new statistic called the slope is proposed to capture the nonstationary characteristics caused by long-term deterministic trend,and extremum statistics are also adopted to enhance the sensitivity of the fault detectability of the model.Then,the Mahalanobis distance statistic is used to perform fault detection,by which the early detection of various types of faults in continuous chemical processes is realized.For the frequent adjustment of operating conditions in chemical industrial processes,the statistical features of various variables show obvious differences between different operating modes and between-mode transitions.The statistical features selected for monitoring in Chapter 2 are no longer applicable to multimode processes.To address this issue,a novel statistical process monitoring method based on the dissimilarity of process variable correlation is proposed in Chapter 3 to extract the common features of data from multiple operating conditions.The proposed method is applied to monitor the correlation of process variables based on the idea that variable correlation should always conform to a certain process internal mechanism,no matter in which stable or transition mode.Mutual information is first employed to quantitate variable correlation with a moving-window approach.Cosine similarity between eigenvalues of mutual information matrices is selected as a dissimilarity index to evaluate the difference in variable correlation between two data sets and perform fault detection for multimode chemical processes.In addition,random disturbances during practical chemical operation or stage transition during batch operation also leads to local time-varying nonstationary characteristics of data.Under this circumstance,a differenceembedded recurrent neural network is proposed in Chapter 4 to extract the highly nonlinear and nonstationary characteristics of process data.A differential structure is embedded into the long short-term memory neural networks to capture the local nonstationary information,and then a memory cell is adopted to determine how much nonstationary information should be retained,by which both the long-term time dependency and short-term nonstationary characteristics can be extracted without loss of original data information.The proposed method has been verified through the quality prediction of an industrial-scale penicillin fermentation process and the fault detection of several continuous chemical processes.Moreover,considering that deep learning-based process monitoring methods are usually with poor generalization ability in practical applications and are prone to overfitting problems when the training samples are insufficient,a regularized Siamese autoencoder based fault detection method is proposed in Chapter 5.A multi-input Siamese network is established,in which data are input in pair to train the model.In this case,the number of training samples increases exponentially with the same amount of data,which makes it applicable to small sample datasets.And a contrastive loss is proposed as the loss function to regularize the latent space,which facilitates the feature extraction of the model and enhances the interpretability to a certain degree.Moreover,fault samples in historical data can also input into the model to further improves the extraction of useful information in historical data.Case studies on the numerical process and the Tennessee Eastman process show that the fault detection performance can be significantly improved under the condition of few model parameters and low structure complexity.Finally,this paper further carries out fault diagnosis research on the basis of the above fault detection research,in order to timely determine the root cause of the fault after it has been detected.With regard to the fault diagnosis based on causal reasoning,the real variable correlation could be buried by signal noises under normal operating conditions and there also could be multiple relationships among variables under the response of control systems.Therefore,a fault propagation path diagnosis method based on time-delayed mutual information is proposed in Chapter 6.The main idea of the proposed method is that variable correlation under normal conditions is mostly contributed by random noises and the real relationship between fault variables could be revealed when a process fault occurs.Therefore,the mutual information between variables under normal conditions is calculated to determine a threshold for each variable pair.Once a fault is detected,each pair of variables with mutual information beyond these thresholds are further investigated by time delayed mutual information analysis using current data,so as to determine the causal logic between them,which is represented as fault propagation paths.Case studies on the Tennessee Eastman process and a chemical industrial process show that the proposed method is able to identify multiple relationships between fault variables and provide correct fault diagnosis results.In summary,this paper provides a research framework for big data-based industrial process monitoring from fault detection to fault variable identification to fault propagation path diagnosis,and successfully realizes online application in a petrochemical production enterprise,which generates significant economic benefits for the enterprise after assessment in 2022.