Few-Shot Early Anomaly Detection in Industrial Manufacturing Using Explainable Siamese Networks

Abstract

Early defect detection is essential to ensure product quality and reduce waste in industrial manufacturing. However, traditional defect detection methods rely on large labelled datasets to train models or manual inspection, both of can be time-consuming as well as prone to errors. The challenge lies in developing an automated system for early anomaly detection that requires minimal labelled data, making it adaptable to various industrial environments.

To address this challenge, a Siamese network, a few-shot learning technique, was utilized. The network was designed to detect defects in images of products with only a few labelled examples. A custom lightweight Convolutional Neural Network (CNN) was developed for the embedding phase of the Siamese network to reduce inference time while maintaining model performance. This architecture, coupled with Explainable AI (XAI), enabled the model to provide transparent and interpretable results, which is crucial for industrial applications where quick decision-making and understanding of model behaviour are vital.

The model was tested on multiple datasets, including the MVTec leather dataset, and achieved promising results with an inference time of approximately 0.03 seconds per image. It demonstrated an accuracy of 96.58% and an F1-score of 0.97, showing its effectiveness in early defect detection with minimal labelled data and high precision. These results underscore the potential for this approach to improve defect detection in industrial manufacturing.