Mode Identification And Process Monitoring For Complex Multimode Processes Based On Data-driven Methods

With the fast development of global economy and science,the modern industrial scale is becoming larger and more complex,and the efficient production process of multiple products has become emphasis in many industries.It is a main challenge in the modern industry to guarantee the safety of production,reduce the production cost effectively,and improve product quality,as well as enhance the economy benefit.Effective process monitoring can be employed to solve the problem.Data-based multivariate statistical techniques show particular advantages to deal with the high-dimension and coupling data,which makes them specially and increasingly attractive.The traditional multivariate statistical methods often contain some assumptions such as Gaussian and unimodal processes.However,due to different manufacturing strategies or varying feedstock,the process often contains multiple operating conditions or multiple operating phases.Different conditions or phases have different process characteristics,which are collectively called multiple modes.Multimode has been widespread in complex industrial processes.To deal with the practical monitoring challenges of multimode processes,some improvements of the conventional monitoring methods have been developed,and a series of novel strategies including mode identification,statistical modeling,and data characteristic analysis,etc.have been proposed in the thesis,which are summarized as follows:1.To handle the problem that offline mode identification of multimode processes cannot be realized automatically,a two-step rough partition and precise identification based strategy is proposed for offline mode identification.First,the data is divided into a series of data segments by a large cutting window,and the segments are assigned into different clusters.Finally,a small moving window is used to determine the accurate boundary between different modes."Rough partition-precise identification" based identification strategy can reduce computing complexity effectively and is robust to the random disturbance such as noise and singular points.Besides,all kinds of unsatisfactory clustering results have been considered and analyzed to give the reasonable and effective identification result,which can realize the mode identification of multimode processes automatically.2.For statistical modeling and online monitoring of multimode continuous processes,a novel monitoring strategy based on multiple Gaussian mixture models is proposed for multimode process monitoring,which combines the advantages of multiple modeling strategies and Gaussian mixture model(GMM).A single Gaussian model or GMM is established for stable mode according to Gaussianity test result of the data.GMM is established for each transition mode due to its complex characteristics.Thus,the number of models is less and the model parameters are more simple than the traditional multiple methods.Different from GMM,the novel strategy can efficiently capture the local features of transitional modes as well as stable modes.When the mode can be decided for sure,the specific model is utilized for monitoring,which can avoid the bad effects of those irrelevant models.All possible models are probabilistically united for monitoring when the mode cannot be identified for sure.This can avoid the severe error which results from wrong identification of multiple mode strategy.3.Uneven length in batch processes may bring many problems in data normalization,phase division and online identification,etc.To handle these issues,a variable moving window-k nearest neighbor(VMW-kNN)based strategy is proposed in this thesis.A variable moving window is used to determine the range of searching the relevant data samples,which can reduce the algorithm complexity and make local model more reliable.The k nearest neighbor of each sample is the same as the time-slice matrix in the even batch process,which can be used to normalize the sample and extract the load matrix.The novel strategy can solve the uneven-length problem without equalizing the batch data and it also applies well to the slow batch processes with limited batches.4.For nonlinear problem in batch processes,two similarity indexes of kernel load matrixes in high feature space have been identified and a novel strategy of phase division and monitoring based on these two similarity indexes have been proposed.The current algorithms only consider the linear characteristic while ignore the effect of the nonlinear characteristic in phase divisions.The nonlinear characteristic is extracted and used by the novel strategy,which makes phase division more reliable.And the nonlinear monitoring strategy based on phase division is more accurate.5.To solve the model selection problems for practical processes,a novel strategy based on data characteristics test has been proposed in this thesis.Traditional MSPC methods often contain some assumptions.If one process monitoring method is chosen without considering its constraints,it may get wrong conclusions and increase the rate of leaking and false alarm.But prior knowledge is hard to obtain in most industrial processes.To cope with the problem,a novel method has been proposed to test the correlations between variables and the data distribution characteristic.Then the appropriate monitoring method is selected automatically according to the test result,which can avoid wrong conclusions and high false alarm rate.The proposed approaches provide new ideas and solutions to handle the practical problems of multimode processes.The proposed methods are applied to the multimode processes such as Tennessee Eastman process and injection process and are compared with several monitoring methods to demonstrate the effectiveness of these new approaches.Finally,based on the conclusion of the thesis,some future research directions are discussed.