核心概念
CapsNets are effective noise stabilizers for time series data, outperforming CNNs in robustness against manual and adversarial attacks.
摘要
Standalone Note here
Abstract
CapsNets utilize capsules to learn position-equivariant features, enhancing robustness.
Investigated CapsNets' effectiveness on noisy time series sensor data, outperforming CNNs.
Introduction
CapsNet maintains object invariance in position-equivariant data.
Hypothesis: CapsNets resist noisy time-series sensor data and attacks from defective sensors.
Background
DR-CapsNets enhance fundamental temporal features while inhibiting noisy ones.
Experiments
Conducted experiments on ECG dataset comparing DR-CapsNets with CNN under various noise baselines.
Results
DR-CapsNets outperformed CNN in all noise baselines, achieving high accuracy and f1-score.
Conclusions
CapsNets effectively analyze noisy time series sensor data, functioning as noise stabilizers and improving robustness.
References
統計資料
Our method achieved an accuracy of 98.7%, 97.5%, 97.2%, and 94.8% for different types of noise attacks.
DR-Caps achieved an accuracy of 98.22% compared to CNN's 94.17%.