摘要
城市轨道交通轨道结构的安全性、可靠性和稳定性是保障正常运营的关键要素。为有效应对轨道结构病害带来的安全挑战,开展高效、精准的轨道结构异常状态识别研究是行业的技术趋势。针对3种轨道结构形式,采用激振设备模拟列车荷载,开展实验室内病害模拟试验,模拟扣件失效、断轨、钢弹簧失效、剪力铰失效等4种轨道结构病害,获得了不同病害下钢轨和道床振动加速度响应;分析不同轨道结构病害表现的时域和频域特征。以加速度时域和频域数据作为输入参数,基于一维卷积神经网络算法,建立了轨道结构病害识别模型,基于ReduceLROnPlateau改进学习率调节策略,形成了可动态衰减和膨胀的学习率调节机制。研究结果表明:钢轨和道床振动加速度时域和频域特征均可反应轨道结构病害特征;病害识别模型推理时间短,鲁棒性高,能有效规避局部最优问题,可对4类轨道结构病害进行快速精准识别;频域数据集的识别精度优于时域数据集,在采用钢轨加速度、道床加速度、钢轨与道床加速度叠加情况下,识别精度分别为99.7%、100%、100%;在SNR=7 dB的噪声干扰下,频域模型依然能保证90%的预测准确率。研究显示了轨道结构病害识别模型的有效性,可为工程中轨道结构病害识别提供一种技术参考方法。
The safety,reliability,and stability of the track structures in urban rail transit are the key to ensure normal operation.In order to effectively address the safety challenges brought about by track structure diseases,carrying out research on efficient and accurate identification of abnormal states of track structures has become a technological trend in the industry.For three types of track structures,vibration-exciting equipment was used to simulate train loads,and laboratory-based disease simulation tests were carried out.Four types of track structure diseases,including fastener failure,rail breakage,steel spring failure,and shear hinge failure,were simulated.The vibration acceleration responses of the rail and the track bed under different diseases were obtained.The timedomain and frequency-domain characteristics of different track structure diseases were analyzed.Using the acceleration time-domain and frequency-domain data as input parameters,a track structure disease identification model was established based on the one-dimensional convolutional neural network algorithm.Based on the ReduceLROnPlateau improved learning rate adjustment strategy,a learning rate adjustment mechanism that can dynamically decay and expand was formed.The results are drawn as follows.Both the time-domain and frequency-domain characteristics of the vibration acceleration of the rail and the track bed can reflect the characteristics of track structure diseases.The disease identification model has a short inference time and high robustness.It can effectively avoid the local optimum problem and can quickly and accurately identify the four types of track structure diseases.The identification accuracy of the frequency-domain data set is better than that of the time-domain data set.When using rail acceleration,track bed acceleration,and the superposition of rail and track bed accelerations,the identification accuracies are 99.7%,100%,and 100%,respectively.Under the interference of noise with an SNR of 7 dB,the frequency-domain model can still ensure a prediction accuracy of 90%.This research demonstrates the effectiveness of the track structure disease identification model and can provide a technical reference method for track structure disease identification in engineering.
作者
刘潇
王少林
闫宇智
丁德云
刘敏
姜博龙
陈万里
LIU Xiao;WANG Shaolin;YAN Yuzhi;DING Deyun;LIU Min;JIANG Bolong;CHEN Wanli(Engineering Technology Center of Urban Rail Transit Vibration and Noise Control,Ministry of Ecology and Environment,Beijing 100071,China;Beijing Jiuzhouyigui Environmental Technology Co.,Ltd.,Beijing 100071,China;National Engineering Center for Digital Construction and Evaluation Technology of Urban Rail Transit,China Railway Design Corporation,Tianjin 300308,China;Beijing Metro Operation Administration Co.,Ltd.,Beijing 100044,China;Beijing Infrastructure Investment Co.,Ltd.,Beijing 100101,China)
出处
《铁道科学与工程学报》
北大核心
2025年第5期2333-2345,共13页
Journal of Railway Science and Engineering
基金
北京市基础设施投资有限公司科研项目(2023-GY-08)。
关键词
轨道结构病害
实验室试验
一维卷积神经网络
识别方法
学习率调节
track structure diseases
laboratory experiments
one-dimensional convolutional neural networks
identification methods
learning rate adjustment
