Recipe Based Anomaly Detection with Adaptable Learning: Implications on Sustainable Smart Manufacturing

TL;DR

This paper introduces a new AI framework for detecting defects in injection molding by adapting to different manufacturing settings, improving quality control and productivity.

Contribution

The novel contribution is a recipe-based anomaly detection framework with adaptable learning for injection molding processes.

Findings

Recipe-Based Learning with K-Means clustering and Kruskal-Wallis tests improved defect detection compared to traditional methods.

Adaptable Learning using KL-Divergence outperformed integrated and additional training models in predictive accuracy.

The proposed framework detected 61 defects versus 41 existing defects, demonstrating enhanced quality inspection.

Abstract

The advent of Industry 4.0 has significantly transformed the manufacturing sector, bringing advancements in quality control efficiency, environmental sustainability, and production development. These changes have led to the development of intelligent technologies such as artificial intelligence (AI). However, implementing AI solutions in manufacturing processes still presents challenges in many aspects, particularly in handling irregular datasets influenced by diverse manufacturing settings. In the field of injection molding, quality inspection often occurs at the batch level rather than at the individual level, providing only the overall defect ratio of batch production instead of labeling each individual product. These issues limit the general application of AI and data-driven decision-making. To address these limitations and enhance product efficiency, this study proposes a novel anomaly detection framework for a specific manufacturing process. In Recipe-Based Learning, we first apply K-Means clustering to account for the flexible manufacturing process, which relies on diverse settings. The injection molding data are classified into setting-specific recipes to ensure data normality and uniqueness. The Kruskal-Wallis test is conducted to provide statistical evidence of differences in data based on varying settings, further justifying the necessity of Recipe-Based Learning. Then, Autoencoders for anomaly detection are trained with normal data from each recipe. With this data-driven AI approach, 61 defective products are predicted, compared to the existing 41 defects. Meanwhile, the integrated model, which does not consider variations in settings, only predicted 2 defects, indicating poor and distorted quality inspection. For Adaptable Learning, which focuses on new inputs with unseen settings, we apply KL-Divergence to identify the closest trained recipe data and its corresponding model. This approach outperformed both the integrated and additionally trained models in predictive power. As a result, continuous prediction is achieved without the need for further training, successfully enhancing process optimization. In the context of smart factories in the injection molding industry, such improvements in process management can significantly enhance overall productivity and decision-making, primarily through a data-driven AI approach.

Figures