A Multi-element Anomaly Detection Approach for Complex Industrial Images

In industrial manufacturing process, inevitable failures due to anomalies can significantly impact yield and reliability. While the traditional automatic optical inspection (AOI) system is capable of catching the anomalies, it falls short in identifying specific anomaly types and analyzing potential causes in complex industrial production lines. This work presents a multi-element anomaly detection (MulEAD) model, specifically designed for complex industrial images. MulEAD extends common detection approaches to identify multiple elements with different classes and shapes with a single network. To test its efficacy, the Wire Bonding Anomaly Dataset (WBAD), which comprises 13,000 complex anomaly instances across 7 categories, is collected from real industrial facilities. Following a five-month evaluation in a manufacturing environment, our model outperforms the state-of-the-art industrial detection methods with an accuracy of 97.2%, and our work enabled the factory to save 2 hours of labor cost per day. In addition, the multiple elements from our model assist engineers in diagnosing the root causes of anomalies and enable them to make data-driven adjustments in manufacturing.