基于深度学习的无人驾驶汽车导航传感器异常诊断方法

近年来，无人驾驶汽车（Unmanned autonomous vehicle，UAV）作为未来智能交通系统（Intelligent transportation system，ITS）的重点发展方向已成为业内研究的热点。作为一种新兴智能交通工具，无人汽车的行驶完全依赖于导航传感器提供的精确位置和路径数据。GPS传感器因计算机黑客的恶意网络攻击或物理故障而造成导航位置数据异常或篡改给无人车的安全行驶带来了巨大的威胁和挑战。针对无人汽车GPS传感器易受攻击的现实问题，提出一种基于深度学习的无人车GPS传感器异常检测新方法。该方法通过改进传统一维卷积神经网络（1D Convolutional neural network，1D-CNN）的拓扑结构，设计出1DGAP-CNN算法框架，使其满足快速实时性的诊断要求。首先，原始的多传感器位置数据直接输入到提出的算法中进行数据融合和预处理；其次，提出的算法自动的进行特征提取、降维减参和模式辨识；最后，模型直接输出诊断结果，整个诊断过程由模型自主完成。结果表明，提出的方法相比于主流的智能诊断算法具有更高的诊断准确率和更快的检测速度。

Anomaly Diagnosis for Navigation Sensors of Unmanned Autonomous Vehicles Based on Deep Learning

As a key technology of future intelligent transportation systems(ITS), unmanned autonomous vehicles(UAVs) have become a research hotspot in recent years. As an emerging intelligent transportation tool, unmanned vehicles rely on the precise location observations provided by navigation sensors. Once compromised, navigation sensors may generate abnormal observations deviating away from the ground truth, which as a result will cause severe consequences or even fatal accidents. To enhance the security of UAVs, a new deep learning based anomaly diagnosis method is proposed in this paper for the detection and identification of sensor anomalies in UAVs. The proposed method improves the topological structure of the traditional 1D Convolutional neural network (1D-CNN) and designs a 1DGAP-CNN algorithm framework to achieve a real-time rapid diagnosis. First, the original pose measurements from multiple sensors are directly input into the proposed algorithm for data fusion and preprocessing. Secondly, the proposed algorithm automatically performs feature extraction, dimension transformation, parameter reduction, and anomaly identification. Finally, the diagnosis results are automatically generated. Evaluation results show that the proposed method has higher diagnostic accuracy and faster detection speed than the state-of-the-art intelligent diagnostic algorithms.