Proposal of VAE-Based Deep Learning Anomaly Detection Model for Industrial Products

Abstract

AbstractThe anomaly detection of images using deep learning has a great demand. When building a deep learning model to detect defects in products during the production process in a factory, it is difficult to procure anomaly products for training since most of the products do not have defects. Therefore, it is desirable to use unsupervised learning rather than supervised learning. Hence, generative models that can be learned unsupervised such as Variational AutoEncoder (VAE) or Generative Adversarial Network (GAN) are often used for anomaly detection. In this research, the authors built an anomaly detection model named Adaptive Weighted Loss (AWL) VAE, based on VAE. AWL VAE is focused on the two terms of the loss function in the general VAE and changes each weight multiplied by each term depending on certain conditions as training progresses. The results of several experiments show that AWL VAE can detect with better accuracy than conventional VAE for an image dataset of industrial products.KeywordsAnomaly detectionIndustrial productVariational autoencoderDeep learningGenerative modelLoss functionUnsupervised learning