结构健康监测异常数据诊断的自动优化方法

Automated Optimization for Anomaly Diagnosis of Structural Health Monitoring

本文针对深度神经网络超参数人工调优成本高昂和性能受人工经验限制的问题,将大型桥梁结构健康监测系统中异常数据诊断建模为图像分类任务。使用卷积神经网络进行训练,同时基于贝叶斯理论,以模型的验证集误差作为目标函数,以高斯过程为概率模型,以期望改进(expected improvement,EI)方法为采集函数,对网络训练过程中的L2正则化系数、初始学习率以及动量进行超参数自动优化。结果表明:人工经验设定超参数的卷积神经网络模型诊断准确率为89.3%,经贝叶斯超参数自动优化后准确率提高到93.2%。通过在最优取值临近空间对超参数进行独立优化发现模型准确率对L2正则化系数取值并不敏感,初始学习率和动量分别宜在[1×10^(-4),1×10^(-2)]和[0.8,0.9]范围取值,为深度神经网络的超参数调优提供实用参考依据。

To address the problems of costly manual tuning of deep neural network hyperparameters and performance limited by manual experience,the diagnosis of abnormal data in large bridge structural health monitoring system is modeled as an image classification task.A convolutional neural network is used for training,while a Bayesian theory-based automated deep learning method is developed.The validation set error of the neural network is set as the objective function,the Gaussian process is utilized as the probability model,and the Expected Improvement(EI)method is considered as the acquisition function.Three hyperparameters,L2 regularization coefficients,initial learning rate,and momentum of the convolutional neural network during training are considered for automated optimization.The results show that the diagnostic accuracy of the convolutional neural network model with manually empirically set hyperparameters is 89.3%,and the accuracy is improved to 93.2%after automatic Bayesian hyperparameter optimization.The model accuracy was found to be insensitive to the value of L2 regularization coefficients by independent optimization of hyperparameters in the optimal taking proximity space,and the initial learning rate and momentum are suitable to be taken in the range of[1×10^(-4),1×10^(-2)]and[0.8,0.9],respectively,which provides a practical reference basis for hyperparameter tuning of deep neural networks.