Investigation of Machine Learning Techniques for Fault Diagnosis in the Semiconductor Manufacturing Process

Industries are going through the fourth industrial revolution (Industry 4.0), where technologies like the Industrial Internet of Things (IIoT), Big Data Analytics and Machine Learning (ML) are being extensively employed for improving the productivity and efficiency of manufacturing systems. Recently, many researchers have demonstrated the ability of ML algorithms to meet various challenges presented by the next generation Smart Manufacturing Systems (SMSs). This work aims to investigate the applicability of several machine learning techniques for early fault diagnosis towards smart manufacturing process. Thus, in this thesis, we propose several fault diagnosis ML models for SMSs applications. A case study has been conducted on a dataset from a semiconductor manufacturing process. However, this dataset contains missing values, redundant and noisy features, and class imbalance problem. This imbalance problem makes it so difficult to accurately predict the minority class, due to the majority class size difference. Therefore, this work proposes and compares the effects of three synthetic data generation techniques to handle such class imbalance problem. To handle issues related to missing values and redundant features, we implemented and compared the performance of two missing values imputation techniques and two feature selection techniques using three adopted data synthetic generation techniques. We then developed and compared the performance of ten predictive machine learning models against the abovementioned proposed approaches. Experimental results across seven evaluation metrics of performance obtained from these models were significant. These results and a comparative analysis show the feasibility and validate the effectiveness of these proposed synthetic data generation techniques and the proposed methodologies. Some among the proposed methodologies could produce an accuracy in the range of 99.9% to 100%. Furthermore, a comparative analysis has been conducted with similar models proposed in the literature. Based on the results, our proposed models outpace those proposed in the literature.