Fault Detection For Industrial Process Oriented To Multi-rate Data

With the rapid development of modem industry,the industrial automation systems are achieving remarkable improvement in complexity,informatization and intellectualization.The advent of Industry 4.0 and intelligentizing era particularly raises a higher demand on the production benefit and process safety.Process monitoring strategy has attracted more and more attention since they are powerful methods to ensure a stable and safe operartion status of the industrial processes,reduce power consumption and pollution,and improve productivity and product quality.However in modem industries,it is difficult to acquire traditional mechanism model because of the large and complex plant-wide production procedure.On the other hand,massive amounts of data's collection,storation and transmition becomes more convenient relying on the rapid development of data science and industrial sensor technology.As a result,industrial processes accumulate huge amounts of process data samples which contain rich process information.However,these data are generally collected from different manufacturing stages with different sa mpling time intervals,along with complex data characteristics such as auto-correlations,cross-correlations,constrained relationship between process measurements and quality-relevant variables,noises pollution etc.Therefore,how to utilize these complex and multi-rate data for process monitoring is urgently addressed in process monitoring area.In this paper,several novel methods are proposed for multi-rate processes modelling and monitoring from the perspective of data characteristics and the process categories.The following sections of this thesis are organized as follows(1)A multi-rate principal components analysis model and corre sponding online fault detection scheme is proposed to extract the cross-correlations between multi-rate process variables in continues and stable multi-rate processes.The proposed model projects the original multi-rate variables into a low-dimensional feather subspace and then performs fault detection strategy in the feather subspace.Compared to traditional multivariate statistical process monitoring methods,the proposed model are built without abandoning or fulfillment of the incomplete multi-rate data samples.The case study confirms that the proposed model is more adaptive in multi-rate process monitoring and more accurate in fault detection problems(2)A multi-rate partial least squares model and corresponding online fault detection scheme is proposed for quality-relevant multi-rate process monitoring issues Quality-relevant variables are frequently encountered in multi-rate processes due to the fact that the quality-relevant variables are always under a much slower sampling rate than process measurements.On the other hand,quality-relevant data samples contains rich production information which has a remarkable value in process monitoring problems.The proposed model can capture the cross-correlations between process measurements and the quality-relevant variables simultaneously.In the model training procedure,the proposed model build a new methods in model parameter solution to accommodate the multi-rate data characteristics.Based on the proposed model,the corresponding online monitoring scheme is proposed and plays a satisfying performance in quality-relevant multi-rate process monitoring.(3)A multi-rate mixture probabilistic principal component analysis model is proposed for multi-mode multi-rate process monitoring problem.The proposed model describe multi-mode process data samples through mixture sub-models for each mode The sub-model is set to be multi-rate probabilistic principal components model to handle the multi-rate data characteristics.In online process monitoring step,the online sample's posterior probabilities belonging to each mode is estimated first and then the statistics from each mode is built upon the estimated expectations of the latent vector's posterior probabilities under this mode.The global monitoring statistics are carried out by combining each mode's statistics and are applied for the online monitoring.The results of the case study show the superior in multi-mode multi-rate process monitoring(4)A multi-rate linear Gaussian state-space model is proposed to handle the dynamic multi-rate process monitoring scenario.The proposed model constrained the process feature vector in a low-dimensional subspace and employs a first-order Markov process to describe the process dynamics in the feature sub-space.Process variables under each sampling rate are generated by the feature vector and the emission matrix separately.As a result,the proposed model extracts the cros s-correlations between multi-rate process variables and auto-correlations between process samples.Corresponding online monitoring scheme is built based on the feature subspace and the construction error of process measurements.The proposed methods is effective in online fault detection and thus is more suitable for multi-rate dynamic process monitoring.