Data driven root cause analysis to assist human experts in quality control

Manual assembly lines are susceptible to quality defects, necessitating the need for early defect detection methods to minimize the negative consequences. However, manual assembly line data is primarily limited to assembly execution times, which poses limitations on the defect analysis. In this study, we analyze the assembly step’s execution times and end-of-the-line quality issues to find insights into quality defect prediction. Several classification algorithms are implemented.  The random forest algorithm slightly outperforms the others. We categorize the quality issues into linearly correlated, non-linearly correlated, and low/not correlated to the processing times categories based on our observations. However, execution times alone may not be sufficient for accurate quality issue prediction, emphasizing the need for additional factors that influence product quality. To gather more comprehensive data, the integration of internet-connected or smart tools in the assembly line is necessary. A future direction is to develop a Bayesian network to analyze the data. Bayesian networks can handle mixed data types, including both categorical and continuous variables.  Another factor that could lead to potential false positives is the inline fixing of errors, which might result in extended processing times. Counterfactual reasoning and confounding concepts should be used to analyze the hidden variables that influence the inputs and outputs.