Chemical Process Monitoring Based On Data Driven Methods

With the expansion of industrial scale and the increasing degree of integration,the relationship between various process variables becomes more complex,and some small faults may cause serious consequences.Therefore,timely detection and accurate diagnosis of the root cause,propagation path and trend of faults is crucial for the chemical industry.The current distributed control system provides a large amount of data,and the development of data information science provides advanced methods for process monitoring,supporting data-driven process monitoring.Based on this,this thesis conducts research on data-driven process monitoring methods in the chemical industry.To fully utilize historical fault data,a discriminant global and local preserving projection feature extraction algorithm is proposed,which introduces historical fault data to extract discriminant information based on the traditional multivariate statistical method of global and local preserving projection.After data dimensionality reduction,to deal with the characteristics of process data with Gaussian and non-Gaussian mixture distributions,a fault detection statistic is constructed using the support vector data description algorithm.A new process monitoring method,named DGLPP-SVDD,is proposed by combining the two algorithms.The DGLPP-SVDD algorithm is applied to the Tennessee Eastman process simulation.The results show that the DGLPP-SVDD algorithm has lower fault detection delay and higher fault detection rate,indicating better process monitoring capability.The cost of generating and labeling fault data in chemical production is very expensive.Therefore,virtual fault data are generated to simulate real faults for training.In this thesis,a new process monitoring method,named Adversarial Contractive Autoencoder(ACAE)based chemical process monitoring method,is proposed.This method combines the advantages of contractive autoencoder and adversarial training to extract more robust and discriminative features.Firstly,the contractive autoencoder is used to learn the denoising of normal operation data,which can obtain the manifold structure of data in low-dimensional space and then extract features with stronger robustness to noise.Then,virtual fault data is generated on this basis to simulate real faults for training,in order to optimize the contractive autoencoder parameters and enhance its discriminative ability to fault data.After training,t-distributed stochastic neighbor embedding is used for two-dimensional visualization of the dimensionalityreduced data to qualitatively analyze the feature extraction capability of the proposed algorithm.The process monitoring effect is quantitatively compared by using the Tennessee Eastman process simulation platform.The results show that the proposed method has better feature extraction capability and process monitoring effect.On the basis of the ACAE method proposed in this thesis,the low-dimensional representations of normal operation data and virtual fault data are extracted.After obtaining the low-dimensional data,a monitoring statistic is constructed using the negative class constrained support vector data description(NSVDD)algorithm with negative class constraint,instead of reconstruction error,for process monitoring.The NSVDD algorithm optimizes the hyper-sphere parameters by introducing virtual fault data,which can further enhance the process monitoring effect.Combining these two algorithms,a process monitoring method named ACAE-NSVDD is proposed.The effectiveness and superiority of the proposed method are validated in the Tennessee Eastman process simulation experiment.