DATA-DRIVEN FAILURE PREDICTION FOR MANUFACTURING PROCESSES

Abstract

Companies are increasingly using data-centric approaches for process improvement and early-stage failure prediction under current Industry 4.0 paradigms. However, limited availability of failure data often constraints the use of complex machine learning models. The first contribution in this thesis combines feature extraction capabilities of convolutional neural networks (CNN) with the domain knowledge characteristics of fuzzy systems. This framework addresses data imbalance using a combination of class weighted CNN models and biased fuzzy inference rules. The individually biased models can be combined using ensemble techniques for achieving balanced predictions. The second contribution tackles failure prediction in high-volume, low-cost production environments, where quality control is carried out at production lot level. Layer-wise L-CNN model ensures process data from each production run is treated independently within the model architecture, while lot-loss L-CNN utilizes a custom loss function for providing run-level predictions. The proposed models are validated on real-world semiconductor industry datasets.