MACHINE LEARNING APPROACH FOR AUTONOMOUS ROOT CAUSE ANALYSIS

Abstract

Improving production yield and operation workflow are constant concerns in manufacturing processes. To achieve these goals, machine learning-based autonomous root cause analysis techniques are developed in this thesis. This thesis first proposes a feature selection method based on bootstrap resampling which can improve classification accuracy by 4.26% (with SVM) and 4.48% (with MLP) compared with the comparative methods. This thesis then proposes an interpretable and actionable model explainer algorithm that works with any machine learning classifiers. The proposed method can generate tuning solutions that convert 73.9% to 85.2% of the negative samples to positive samples and interpret the predictions. This thesis also proposes an uplift modelling method to estimate the causal effects change which reduces more than 30% of mean-squared error. This thesis finally proposes a pre-processing approach for imbalanced regression which outperforms the comparative methods on the imbalanced regression tasks and achieves 4.98% to 8.57% error reduction.