Di-CNN: Domain-Knowledge-Informed Convolutional Neural Network for Manufacturing Quality Prediction.

Area of Science:

• Manufacturing Engineering
• Artificial Intelligence
• Materials Science

Background:

• Convolutional Neural Networks (CNNs) are prevalent in manufacturing for process monitoring and quality prediction using image sensor data.
• Purely data-driven CNNs lack integration of physical measures and domain knowledge, limiting prediction accuracy and practical interpretability.
• Integrating manufacturing domain knowledge is crucial for enhancing CNN performance in industrial applications.

Purpose of the Study:

• To develop a novel CNN model, Di-CNN, that leverages manufacturing domain knowledge to improve accuracy and interpretability in quality prediction.
• To adaptively weigh design-stage information and real-time sensor data during model training.
• To demonstrate the superior performance of the proposed Di-CNN model in a real-world manufacturing case study.

Main Methods:

• A novel Di-CNN model was developed, incorporating both design-stage information (e.g., working conditions) and real-time sensor data.
• The Di-CNN model adaptively weighs these data sources during training, guided by manufacturing domain knowledge.
• Performance was evaluated on resistance spot welding quality prediction using mean squared error (MSE) via sixfold cross-validation.

Main Results:

• The Di-CNN with adaptive weights achieved a mean MSE of 6.8866 and a median MSE of 6.1916.
• A Di-CNN without adaptive weights showed significantly higher errors (mean MSE: 13.6171, median MSE: 13.1343).
• Conventional CNNs exhibited the highest errors (mean MSE: 27.2935, median MSE: 25.6117), confirming the proposed model's superiority.

Conclusions:

• The proposed Di-CNN model, integrating manufacturing domain knowledge and adaptively weighing data sources, significantly outperforms conventional CNNs in quality prediction.
• Leveraging domain knowledge enhances both the accuracy and interpretability of CNN models in manufacturing settings.
• The Di-CNN approach offers a promising direction for data-driven quality prediction in complex industrial processes like resistance spot welding.