基于集成神经网络的超高层结构异常监测数据诊断

Abnormal Monitoring Data Diagnosis for Super High-Rise Structures Based on Ensemble Neural Network

摘要:

针对结构健康监测中多种异常数据人工识别效率低下的难题，提出一种基于集成式神经网络模型的超高层结构异常监测数据诊断方法。采用短时傅里叶变换，压缩并提取其中包含主要结构振动模态的时频域信息，实现了数据的高保真压缩；采用Bagging集成策略，通过自主采样方法生成多个子训练集，并行训练若干个子神经网络模型，通过融合子模型的预测结果，提高异常数据类型辨识的准确率。基于上海中心大厦健康监测系统的实测数据，对所提出方法的有效性和可靠性进行了验证。结果表明，基于集成式模型的异常监测数据诊断方法的准确率高达98.8%，实现了精准度高、鲁棒性强的异常监测数据诊断。

Abstract:

To address the inefficiency in manual identification of diverse anomalies in structural health monitoring （SHM） data， a method based on ensemble learning model is proposed to detect the abnormal data in super high-rise building SHM systems. By employing short-time Fourier transform， the time-frequency domain information containing the main structural vibrational modes is extracted and compressed， thus the feature extraction and high-fidelity compression of original data attained. Moreover， a Bagging ensemble strategy is introduced， and multiple training subsets are generated through bootstrap sampling， based on which each individual neural network model is trained independently. By aggregating the prediction results of multiple well-trained models， the precision of anomaly detection is enhanced. Furthermore， the proposed method is applied into the Shanghai Tower SHM system to validate the feasibility and reliability. The results indicate that the diagnosis accuracy reaches 98.8% by the proposed ensemble model-based abnormal data detection method， and high precision and strong robustness of the anomaly SHM data diagnosis are confirmed.