Research And Application Of Monitoring Method Based On Online Updat

As the most important technique of statistical process control,control charts can be applied to various user scenarios such as industrial production,disease prevention,and geological monitoring.Limited by factors such as the sample size of historical observations,theory foundation and computational complexity,traditional control charts are generally based on a constant sample size of observations and constant parameter estimates.With the follow-up of industrial big data as well as the rapid development of data collection and analysis technology,data streams to be monitored significantly presents patterns of "complicated forms"and "dynamic features".The "constant" or so called "static" design of control charts can no longer fully meet the practical needs of modern quality control and management.Therefore,it is urgent to break the "static" stereotype and construct novel control charts,which flexibly adapt to new data forms and application scenarios.This thesis mainly studies online monitoring based on update strategies,namely sample update and parameter update.Regarding the former,a new monitoring perspective motivated by transfer learning is proposed.We concentrate on self-starting monitoring of continuous data process and also Bayesian monitoring of Poisson process.The latter part designs a new recursive update procedure of parameter estimation and proposes the corresponding control chart design.Previous simulation works show that existing self-starting control schemes are not efficient enough when monitoring processes with shifts at the start-up stage.Inspired by the landslide monitoring scenario,the first part of this thesis focuses on improving the performance of monitoring continuous target processes based on a available pre-observed data set.By modifying the self-clustering algorithm from unsupervised transfer learning,a sample update strategy based on transfer learning is innovatively proposed.With the novel update strategy,referenced data set is selected from the pre-observed data set and afterwards,a self-starting control chart based on transfer learning is accordingly designed.Numerical simulation results show that with the "assistance" of the pre-observed data set,the proposed self-starting control chart performs better than the non-transferring control chart as a whole,especially when monitoring a process that experiences a large shift at the early stage.However,when the distribution difference between the preobserved data set and the target in-control process increases,the advantage of our proposed scheme shrinks to a certain extent.We introduce a step-by-step illustration for the proposed algorithm and illustrate the superiorities of our design using the landslide monitoring example,in which the novel chart alarms faster and performs more robust.The second part of the thesis is also developed from the new perspective of transfer learning.In order to monitor Poisson counting process,strategy of information transfer is considered under the Bayesian statistics framework.Under the background of modern advanced manufacturing,especially in industries with significant "customization" characteristics,information transfer may effectively broaden operators’ thoughts to mine key information from the similar data process and thus improve the performance.The study uses the binary Gamma distribution to build a "bridge" between the source domain and the target domain,and simultaneously measure the "similarity" of the two domains.Through Bayesian inference,the in-control posterior probability density function,posterior expectation as well as posterior variance of the target process based on the historical data of both domains is derived.Using the theoretical inference results of Poisson Bayesian,a control chart based on the cumulative probability density function and an EWMA-type control chart are constructed respectively.The former one is demonstrated to be suitable for monitoring shifts of large magnitudes and the latter one performs better for small-scale shifts.Through numerical simulation and case analysis,the superiority of the proposed Poisson Bayesian transferring and corresponding control chart is verified.The third part studies the parameter update strategy of self-starting control scheme to online monitor binary logistic regression profile data.Online monitoring not only pursues the sensitivity and robustness,but also requires consideration of operation efficiency.Previous works on self-starting monitoring generally adopt the method of "using all historical observations to estimate the in-control state",which is not economical enough when monitoring complex processes like profiles.Motivated by the idea of aggregated estimation equation estimation for massive data set,this part of thesis proposes a novel recursive parameter update strategy for in-control parameters.Under the premise of not increasing the computational cost,the coefficients and covariance matrix of the in-control logistic regression model are recursively updated.Simulation results verify that the proposed method performs more accurate than the previous one.Based on the aggregated recursive parameter update strategy,we also propose a self-starting control scheme of Hotelling’s T2 type and proves the asymptotic property of the monitoring statistics.Finally,the strength of the novel control chart is verified by numerical simulation and real data example.