Machine Learning-based Control Charts with Variable Parameters for GLM Profiles

Sammanfattning

This thesis investigates machine learning-based control charts with Variable Pa- rameters(VP)formonitoringGeneralizedLinearModel(GLM)profiles. Previous studies have examined machine learning control charts under Fixed Parameters (FP) schemes for several GLM distributions, while VP schemes have mainly been studied for Binomial profiles with logit link. This thesis extends the VP frame- work to two additional GLM settings: Poisson profiles with log link and Gamma profiles with log link. Three machine learning methods are considered: Neural Networks, XGBoost, and Support Vector Regression. For each method, FP and VP control charts are calibrated using simulated in-control profiles, and their performance is evaluated under intercept shifts, slope shifts, and simultaneous shifts. FP performance is measured using Average Run Length, while VP performance is measured using Average Time to Signal. The results show that VP schemes often reduce detection time compared with FP schemes, but the size of the improvement depends on the distribution, model, and shift size. For Poisson profiles, VP provides clear improvements for Neural Networks and Support Vector Regression, especially at moderate shifts. For Gamma profiles, the main VP advantage appears at very small shifts, since larger shifts are usually detected almost immediately. Neural Networks and Sup- portVectorRegressionshowstableperformanceinmostsettings, althoughNeural Networks under the first training method show large-shift instability for Gamma profiles. XGBoost is less competitive overall and appears more sensitive to cal- ibration and tied prediction values. Overall, the thesis demonstrates that VP schemes can improve machine learning-based GLM profile monitoring beyond the previously studied Binomial case, particularly when the shift is not already detected almost immediately by the FP chart.