基于多模型集成的结构健康状态趋势预测

Prediction of Structural Damage Trend Based on Multi-Model Integration

由于单一的预测模型不能充分地反映出结构振动数据的复杂规律和信息,为了提高结构健康状态趋势预测的精度,充分利用每个模型的优点,提出了基于深度置信网络(depth-confidence network,DBN)、长短时记忆神经网络(long-short-term memory neural network,LSTM)、小波神经网络(wavelet neural network,WNN)的多模型集成预测方法.首先,将工程结构振动信号经变分模态分解(variation mode decomposition,VMD)分和Hilbert变换得到瞬时频率;然后,将瞬时频率作为多模型集成的输入,通过加权平均法和投票法融合的方式对权值系数进行分配,分析不同的权值对预测精度的影响.实验结果表明,当权值ω=0.8时多模型集成方法的预测结果更接近实际值,相比于传统的算术平均模型和其他三种单一的预测模型,多模型集成方法结合了所有的预测模型所具有的全部优点,预测性能最优,预测精度最高.

Because a single prediction model cannot fully reflect the complex laws and structural vibration information,in order to improve the accuracy of structural health trend prediction and make full use of the advantages of each model,a multi-model integrated prediction method was proposed based on depth-confidence network(DBN),long-short-term memory neural network(LSTM)and wavelet neural network(WNN).Firstly,decomposing engineering structural vibration signal into instantaneous frequency with variation mode decomposition(VMD)and Hilbert transform,the system was arranged to take the instantaneous frequency as the input for the multi-model integration to decide the weight coefficients based on a combination of the weighted average method and the voting method.And then the influence of different weights on the prediction accuracy was analyzed.Finally,some verified experiments were carried out.The experimental results show that the prediction results of the multi-model integration method are closer to the actual values when the weight valueωequals to 0.8.Compared with the traditional arithmetic average model and other three single prediction models,the multi-model integration method can provide better prediction performance and higher prediction accuracy.