Explainable AI techniques for quality improvement in automotive manufacturing processes

Artificial intelligence (AI) is widely applied to optimize manufacturing processes in several industrial sectors, including automotive industry. However, traditional black-box AI models often lack transparency, making it challenging for users to understand the rationale behind their output. Explainable AI techniques (XAI) can provide interpretability of AI models, making them more trustworthy and promoting higher user acceptance. In this work, we considered the problem of finding correlations between process parameters of a car painting process related with the aspect quality obtained on the bonnet of car bodies, from the accuracy and explainability balance perspectives. The results generated by decision trees, apriori and post-hoc explainability agnostic models (LIME and SHAP) were presented to and assessed by experts, who provided their feedback about their level of understandability. The experiments carried out allowed us to conclude that the technique that best fits the factory's needs is the apriori algorithm. The end users report that, since the manufacturing process can usually be modified and parametrized, apriori provides more interesting insights to guide the potential deployment of improvements from the obtained hints.

Palabras clave: Artificial intelligence (AI), Explainable Artificial Intelligence, Automotive industry, manufacturing