Explainable AI Framework with Multi-Source Data-Driven Anomaly Detection for Injection Molding Machines

This paper proposes an explainable framework for a deep learning-based anomaly detection (AD) model with visual analysis of multi-source data from injection molding machines (IMMs). In the proposed framework, a reconstruction model is trained to identify data as normal or faulty with two different conservative levels of thresholds. Beyond the per instance reconstruction error information and corresponding AD results, a visual analysis of the underlying causes of anomaly is provided, fostering further insights for users. Furthermore, a possible classification of detailed fault modes is suggested by clustering the latent embedding vectors for detected anomalies. The proposed framework is demonstrated on a real-world IMM dataset, suggesting how the information analyzed could enhance the management of IMMs.