基于CNN与DCGAN的结构振动监测传感器故障诊断及监测数据恢复

Fault self-diagnosis of structural vibration monitoring sensor and monitoring data recovery based on CNN and DCGAN

传感器是结构健康监测系统的关键组成部分,其在服役期间可能发生性能退化甚至故障,故障传感器采集的错误信号会影响结构状态评估结果的准确性。为此,提出基于卷积神经网络(convolutional neural networks, CNN)和深度卷积生成对抗网络(deep convolutional generative adversarial networks, DCGAN)的结构加速度传感器故障自诊断及故障信号自恢复算法。以加速度时程数据为输入,建立基于CNN的传感器故障诊断模型,判断传感器故障类型和故障位置。根据传感器故障诊断结果,对数据集进行相应的处理。将故障传感器信号作为判别器的输入,利用剩余健康传感器信号潜在特征与故障传感器信号之间的相关性,训练基于DCGAN的信号恢复模型,对故障传感器信号进行恢复。采用Benchmark模型和实桥测试结果验证所提方法的可行性及可靠性,并探讨不同噪声水平对信号恢复结果的影响。研究结果表明:基于CNN传感器故障诊断模型具有较好的抗噪性能,传感器的故障诊断准确率在90%以上。恢复信号在时域、频域与真实信号匹配良好。重构误差随着信噪比的降低和故障传感器数量在总传感器数量中占比的增加而增大,但重构信号与真实信号的R~2均在0.8以上。

Sensors are the key components of structural health monitoring systems, which may degrade or even fail during service. The error signal collected by the fault sensor will affect the accuracy of the structure health status evaluation. Therefore, a fault self-diagnosis and fault signal self-recovery algorithm of structural acceleration sensors was proposed, according to the convolutional neural networks(CNNs) and deep convolutional generative adversarial networks(DCGANs). First, the sensor fault diagnosis model given by the CNN was trained with the acceleration time history curve as input. The fault type and fault location of the sensor were determined. According to the sensor fault diagnosis results, the dataset was processed accordingly, and the fault sensor signal was used as the input of the discriminator. The signal recovery model based on DCGAN was trained by using the correlation between the latent features of the remaining healthy sensor signal to recover the fault sensor signal. The benchmark model and real bridge test results are used to verify the feasibility and reliability of the proposed method. The influence of different noise levels on signal recovery results is discussed.The results show that the sensor fault diagnosis model based on CNN has better anti-noise performance. The fault diagnosis accuracy of the sensor is above 90%. The recovered signal matches well with the real signal in the time domain and frequency domain. The reconstruction error increases with the decrease in SNR and the increase in the number of faulty sensors, but the R2 of the reconstructed signal and the real signal is above 0.8.