基于FESN的结构健康状态智能预测研究

发布时间：2018-12-30 20:50

【摘要】：在实际生活中,大型建筑结构和设备在服役过程中会或多或少的出现损伤问题,如果没有被及时发现和处理,往往会造成人财两尽的严重后果,所以结构的健康监测、诊断、评估和预测显得尤为重要。本文在结构健康监测的前提下,以结构健康状态趋势预测为目的,进行结构损伤特征的分析与提取,研究结构的损伤趋势。具体的研究内容如下:研究了基于经验小波变换(Empirical Wavelet Transform,简称EWT)的信号预处理方法,利用EWT对采集的结构损伤加速度振动信号的频谱进行自适应分割,构造出合适且正交的小波带通滤波器组,得到具有紧支撑傅里叶频谱的调幅-调频(Amplitude Modulated-Frequency Modulated,简称AM-FM)分量,再提取出包含丰富损伤信息的分量,对其进行Hilbert变换,计算出瞬时频率和瞬时幅值。实验结果表明:瞬时频率能够反映出结构发生损伤前后的刚度变化形式,而且检测节点位置不同或者损伤工况不同其瞬时频率都会有明显的差异,故将其作为结构健康状态的预测指标,可以很好地反映结构健康状态的变化趋势,为进一步的损伤趋势预测奠定了基础。研究了模糊理论和回声状态网络相结合的非线性时间序列预测方法,对其推理算法、训练过程和网络的关键参数进行了详细地研究说明,并对该网络算法的稳定性能进行严格的定义。实验结果表明:选取合适的参数对预测精度有一定影响,采用双曲正切(tanh)型神经元激活函数的预测精度比泄露(leaky)型更高,相比于传统的回声状态网络,模糊回声状态网络(Fuzzy Echo State Network,简称FESN)的非线性逼近能力强,预测精度高且能够处理较大的样本数据,训练效率也有一定的提高,为实际工程结构的健康状态预测提供了理论依据。研究了基于FESN的结构健康状态趋势预测方法,应用EWT方法提取出结构内部具有损伤信息的AM-FM分量,并进行Hlibert变换,得到瞬时频率,再将其作为预测模型的输入。应用FESN分别对单自由度结构和多自由度结构模型进行工程仿真预测,并应用于实际工程数据的预测。实验结果表明:FESN预测模型更加逼近真实值,预测精度更高。
[Abstract]:In real life, large building structures and equipment will be damaged more or less in the course of service. If they are not detected and dealt with in time, they will often result in serious consequences for both human and financial resources. Therefore, structural health monitoring and diagnosis, Evaluation and prediction are particularly important. On the premise of structural health monitoring, this paper analyzes and extracts the structural damage characteristics and studies the damage trend of the structure in order to predict the trend of structural health state. The specific research contents are as follows: the signal preprocessing method based on empirical wavelet transform (Empirical Wavelet Transform,) is studied, and the spectrum of structural damage acceleration vibration signal collected by EWT is segmented adaptively by EWT. An appropriate and orthogonal wavelet bandpass filter bank is constructed to obtain the amplitude modulation-frequency modulation (AM-FM) component with compact support Fourier spectrum, and then extract the component which contains abundant damage information, and then carry on the Hilbert transform to it. The instantaneous frequency and amplitude are calculated. The experimental results show that the instantaneous frequency can reflect the stiffness change of the structure before and after the damage, and the instantaneous frequency will be obviously different with the different location of the detection node or the different damage condition. Therefore, taking it as a predictor of structural health state can well reflect the changing trend of structural health state and lay a foundation for further prediction of damage trend. The nonlinear time series prediction method based on fuzzy theory and echo state network is studied. The reasoning algorithm, the training process and the key parameters of the network are studied in detail. And the stability of the network algorithm is strictly defined. The experimental results show that choosing appropriate parameters has certain influence on the prediction accuracy. The prediction accuracy of hyperbolic tangent (tanh) neuron activation function is higher than that of leaking (leaky) type, compared with the traditional echo state network. Fuzzy echo state network (Fuzzy Echo State Network,) has strong nonlinear approximation ability, high prediction accuracy and ability to deal with large sample data, and the training efficiency is also improved to a certain extent. It provides a theoretical basis for the prediction of the health state of practical engineering structures. In this paper, the structure health trend prediction method based on FESN is studied. The AM-FM component with damage information is extracted by EWT method, and the instantaneous frequency is obtained by Hlibert transform, which is used as the input of the prediction model. The single degree of freedom structure and multi-degree-of-freedom structure model are simulated by FESN and applied to the prediction of practical engineering data. The experimental results show that the FESN prediction model is more close to the real value and the prediction accuracy is higher.
【学位授予单位】：长安大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TU317

【参考文献】