基于元学习的小样本铁路入侵检测研究

Research on Few-shot Railway Intrusion Detection Based on Meta-learning

基于监控视频的入侵检测是保障铁路安全运营,人工智能技术赋能智慧铁路的一项重要业务。然而,传统深度学习模型依赖海量训练数据,实际铁路场景中面临场景差异大、侵限样本稀缺等问题,使得传统深度学习模型在铁路场景中迁移难、过拟合,无法实现高识别准确率。本文提出一种基于MAML元学习算法的铁路入侵检测方案;设计一个轻量级的深度神经网络,并在国产飞桨深度学习平台实现。最终,在目标监控场景下,利用本文所提模型,仅经过少量标注数据快速微调训练,可达到82.4%的检测准确率。

Surveillance-video based intrusion detection empowered by artificial inteligence technologies is an important business to ensure the safe operation of smart railways.However,traditional deep learming models rely on massive amounts of training data.In actual railway scenes there exist the problems of high inter-scene dissimilarity and few shot intrusion samples,which make the traditional deep learming models overfit in the training phase and fail to achieve high detection accuracy when they are transferred to a new railway scene.This paper proposes a railway intrusion detection scheme based on model-agnostic meta-leaming and designs a lightweight deep neural network.By programming on the PaddlePaddle deep learning platform,the proposed method achieves 82.4%intrusion detection aceunacy after only a few fine-tuning steps with only small number of labeled data.