SNN Based Anomaly Detection Using ESVAE

Namji Kim, Jongpil Jeong, Chaegyu Lee, Dokyung Lee, Jaewon Seo, Jaewon Jung

Abstract: In the manufacturing industry, Anomaly Detection (AD) technology is emerging as an essential element for equipment anomaly detection and quality improvement. In this paper, we implemented an SNN-based Efficient Spiking Variational Autoencoder (ESVAE) based anomaly detection model to improve the efficiency of anomaly detection in manufacturing and provide technological innovation, and conducted comparative experiments with ANNs-based VAE (ANN-VAE) and FSVAE models using MVTec AD dataset. ESVAE improved the average image-level receiver operating characteristic curve (ROCAUC) by 5.7% over Fully Spiking Variational Autoencoder (FSVAE) and the average pixel-level ROCAUC by 5.3% over ANN-VAE, while reducing the training time by about 8.1%. This demonstrates that the ESVAE model is both energy efficient and temporally robust. The proposed model is expected to make a significant contribution to real-time anomaly detection and quality inspection systems.