Hidden Markov Model-based Industrial Process Monitoring

With the increase in the complexity of industrial processes,effective monitoring and diagnosis play an important role in ensuring the safe operation of chemical equipment,maintaining product quality,optimizing product profits,and improving the sustainable development of the environment.Due to the wide use and rapid development of sensors,more and more industrial data have been collected,which has prompted the rapid development of data-based industrial monitoring technologies over the past decade.However,data-driven industrial monitoring technology still has many deficiencies.Therefore,in this paper,for the dynamic characteristics,uncertainties,nonlinearity and other complex characteristics of data in industrial processes,the hidden Markov model under dynamic Bayesian network is used as the main research method.The main content of this thesis includes the following aspects:1.In view of the complexity of modern industrial data characteristics,a single feature extraction method may not achieve good results.An industrial fault diagnosis method based on adaptive feature extraction is proposed.Before feature extraction,the characteristics of the data are analyzed first,and different feature extraction methods are selected for different data characteristics.In order to achieve the purpose of retaining the most effective information,the accuracy of fault diagnosis is improved.2.Taking into account that not all collected data information is conducive to fault identification,this paper presents an effective variable selection and moving window hidden Markov model-based approach for process monitoring.First of all,a variable selection method based on coefficient of variation is proposed to select the variables which are more sensitive to the fault.Second,a single sample is more sensitive to noise and the information it contained could be insufficient,which are harmful for fault identification.So moving window is introduced to utilize the dependency of samples for improving the accuracy of on-line fault identification.Finally,considering that in addition to known faults,the unconsidered faults in industrial processes are likely to show up and require more attention.A threshold statistic based on MVHMM is defined to identify unknown faults.3.Aiming at the complicated industrial process that traditional fault detection methods are not suitable due to the existence of multiple working modes,a multimode process monitoring approach based on MVHMM is proposed.First,a moving window is introduced to utilize the independence of samples for improving the accuracy of online mode identification instead of just considering the posterior probability of one single sample.Second,an MVHMM-based threshold statistic is defined for identification of the unknown mode.Then,the Viterbi algorithm is employed to identify the known modes,which include stable modes and transitions.At last,a new index monitoring scheme which combines two kinds of probability information is utilized to complete the monitoring task for known modes after mode identification.