Enhanced Cyber-Attack Detection in Intelligent Motor Drives: A Transfer Learning Approach With Convolutional Neural Networks
Enhanced Cyber-Attack Detection in Intelligent Motor Drives: A Transfer Learning Approach With Convolutional Neural Networks
复制标题
DOI：
10.1109/jestie.2023.3346802
复制
发表时间：
2024-04
期刊：
影响因子：
--
通讯作者：
Bowen Yang;Shushan Wu;Kun Hu;Jin Ye;Wenzhan Song;Ping Ma;Jianjun Shi;Peng Liu
中科院分区：
文献类型：
--
作者：
Bowen Yang;Shushan Wu;Kun Hu;Jin Ye;Wenzhan Song;Ping Ma;Jianjun Shi;Peng Liu
随着网络化数字控制单元在智能电机驱动系统中越来越普遍，网络安全问题也越来越多，导致各种网络攻击检测方法的发展，以提高系统的可靠性。虽然数据驱动的方法比基于物理的方法具有优势，但对大量实验数据的要求提出了重大挑战。本文提出了一种基于卷积神经网络（CNN）的转移学习的新型电机驱动网络攻击检测方法。该方法首先使用大量模拟数据预训练CNN模型，然后使用有限实验数据的迁移学习对其进行微调，实现了99.5%准确率的出色检测性能，同时降低了开发成本，风险和时间。此外，该模型保持令人满意的检测精度超过96%，即使实验训练数据被限制到原始可用数据的10%。研究结果表明，与新训练的模型相比，迁移学习模型在有限的实验数据下表现出更快的收敛速度和更好的性能。所提出的方法大大减少了开发过程中对大量实验数据的依赖，降低了与网络攻击检测器开发相关的成本和风险，加强了仿真和实验之间的联系，并通过利用强大的仿真模型显着缩短了开发周期。
As networked digital control units become increasingly prevalent in intelligent motor drive systems, cybersecurity concerns have risen, leading to the development of various cyber-attack detection methods to improve system reliability. Although data-driven methods offer advantages over physics-based approaches, the requirement for extensive experimental data presents a significant challenge. This article proposes a novel cyber-attack detection approach for motor drives using Transfer learning based on convolutional neural networks (CNNs). The method initially pretrains a CNN model with substantial simulation data and fine-tunes it using transfer learning with limited experimental data, achieving outstanding detection performance with 99.5% accuracy while reducing development costs, risks, and time. In addition, the proposed model maintains satisfactory detection accuracy of over 96% even when experimental training data are limited to 10% of original available data. The findings indicate that transfer-learned models exhibit faster convergence and better performance when limited experimental data are available compared with newly-trained models. The proposed approach substantially reduces the reliance on large quantities of experimental data during the development process, lowers costs, and risks associated with cyber-attack detector development, strengthens the connections between simulations and experiments, and significantly shortens the development period by leveraging powerful simulation models.