Few-shot pump anomaly detection via Diff-WRN-based model-agnostic meta-learning strategy

Abstract

As a critical component in agriculture, industry, and the military, pump anomaly detection has recently aroused wide attention, which requires deep and abundant development and application. Researchers emphasize deeper networks that are long vast computational resources despite insufficient training samples prepared. To break through this obstacle above, we propose a few-shot model-agnostic meta-learning strategy (MAMLS) model to mitigate the data scarcity problem. Inspired by the diffusive ordinary differential equations (ODEs) and Wide-Resnet (WRN), we made great strides by connecting diffusion (Diff) mechanism and self-adaptive Lr with MAMLS. We generate two classical synthetic datasets (circle and spiral) to clarify the diffusion algorithm’s capability to enhance the relationships and weaken the noise. The experimental results under synthetic data confirm that accuracy quickly reached 99% after several iterations. In an actual case anomaly detection study on the pumps simulation platform, the proposed Diff-WRN-MAMLS brings substantial advantages in saving hardware resources. Compared to current models, our model achieves 98% accuracy in 9-way 25-shot tasks. In the operating efficiency experiment, our algorithm only consumed 14.37 quality factors. The final experiment with four state-of-the-art model-agnostic meta-learning (MAML)-enhanced methods demonstrates the highest reliable test accuracy in different cases, reaching 98.5, 97.8, and 98.4%, respectively. Results showed that the proposed method will generalize surprisingly well in anomaly detection in future research.