Artificial Intelligence (AI) has proven effective in assisting manufacturing companies to achieve Zero Defect Manufacturing. However, certain products may have quality characteristics that are challenging to verify in a manufacturing facility. This could be due to several factors, including the product’s complexity, a lack of available data or information, or the need for specialized testing or analysis. Prior research on using AI for challenging quality detection is limited. Therefore, the purpose of this article is to identify the enablers that contributed to the development of an AI-based defect detection approach in an industrial setting. A case study was conducted at a transmission axle assembly factory where an end-of-line defect detection test was being developed with the help of vibration sensors. This study demonstrates that it was possible to rapidly acquire domain expertise by experimenting, which contributed to the identification of important features to characterize defects. A regression model simulating the normal vibration behavior of transmission axles was created and could be used to detect anomalies by evaluating the deviation of new products compared to the model. The approach could be validated by creating an axle with a built-in defect. Five enablers were considered key to this development.

The Zero Defect Manufacturing (ZDM) concept combined with Artificial Intelligence (AI), a key enabling technology, opens up new opportunities for improved quality management and advanced problem-solving. However, there is a lack of applied research in industrial plants that would allow for the widespread deployment of this framework. Thus, the purpose of this article was to apply AI in an industrial application in order to develop application insights and identify the necessary prerequisites for achieving ZDM. A case study was done at a Swedish manufacturing plant to evaluate the implementation of a defect-detection strategy on products prone to misclassification and on an imbalanced data set with very few defects. A semi-supervised learning approach was used to learn which vibration properties differentiate confirmed defects from approved products. This method enabled the calculation of a defect similarity ratio that was used to predict how similar newly manufactured products are to defective products. This study identified four prerequisites and four insights critical for the development of an AI solution supporting ZDM. The key finding demonstrates how well traditional and innovative quality methods complement one another. The results highlight the importance of starting data science projects quickly to ensure data quality and allow a ZDM detection strategy to build knowledge to allow for the development of more proactive strategies, such as the prediction and prevention of defects.