The Adaptive Personalized Federated Meta-Learning for Anomaly Detection of Industrial Equipment

Federated meta-learning solves many challenges for industrial equipment anomaly detection, such as small-samples problems and significant heterogeneity in the equipment environment. However, this method is very sensitive and not easy to manually adjust an appropriate parameter for excellent results. To address this challenge, this paper proposes an adaptive personalized federated meta-learning framework. Specifically, we first design the hyperparameters online learning rate adaptation method to adjust hyperparameters dynamically. Then, we design a training mechanism with an adaptive learning rate based on model-agnostic meta-learning to update local meta-learning models adaptively. We validate our framework in a real anomaly dataset with a variable control approach. The experiments show that the value of the loss function hardly changes with differ learning rates
$\alpha $
, and it can guarantee the accuracy to fluctuate within 1%. Even when
$\alpha $
is 0, the model still has a good performance. Therefore, our framework has an excellent ability to adjust the hyperparameters. This framework can be used for anomaly detection of industrial equipment in different environments.