Structure damage diagnosis using neural network and feature fusion

https://doi.org/10.1016/j.engappai.2010.08.011Get rights and content

Abstract

A structure damage diagnosis method combining the wavelet packet decomposition, multi-sensor feature fusion theory and neural network pattern classification was presented. Firstly, vibration signals gathered from sensors were decomposed using orthogonal wavelet. Secondly, the relative energy of decomposed frequency band was calculated. Thirdly, the input feature vectors of neural network classifier were built by fusing wavelet packet relative energy distribution of these sensors. Finally, with the trained classifier, damage diagnosis and assessment was realized. The result indicates that, a much more precise and reliable diagnosis information is obtained and the diagnosis accuracy is improved as well.

Introduction

Various damages like crack or delamination in structures are unavoidable during service due to the impact or continual load, chemical corrosion and aging, change of ambient conditions, etc. It has been theoretically and practically proved that damage in a structure will cause the change of structural stiffness, natural frequency and damping, leading to the variation of dynamic response of the whole structure. Detection of related acceleration and vibration frequency response signal of non-stationary components could reflect the wealth of information injury. However the traditional signal processing methods cannot properly reflect the characteristics of non-stationary signals, making it difficult to get satisfying results (Tseng and Naidu, 2002).

Most of vibration-based damage assessment methods require the model properties, which can be obtained from the measured signals through the system identification techniques such as the Fourier transform (FT). The Fourier analysis transforms the signal from a time-based or space-based domain to a frequency-based one. Unfortunately, the time or space information may be lost during performing the transform and it is sometimes impossible to determine when or where a particular event took place (Han et al., 2005). The wavelet transform (WT) overcomes the problems that other signal processing techniques exhibit. The main advantage of using wavelets is the capacity to perform local analysis of a signal, i.e., to zoom on any interval of time or space. Wavelet analysis is thus capable of revealing some hidden aspects of the data that other signal analysis techniques fail to detect. One possible drawback of WT is that the frequency resolution is quite poor in the higher frequency region. The wavelet packet transform (WPT) is an extension of the WT, which provides a complete level-by-level decomposition of signal (Mallat, 1989). The wavelet packets are alternative bases formed by the linear combinations of the usual wavelet functions (Coifman and Wickerhauser, 1992). Therefore, the WPT enables the extraction of features from the signals that contain both the stationary and non-stationary components with an arbitrary time-frequency resolution. Moreover, each wavelet packet decomposition frequency band has component energies which are more sensitive to damage and thus can better describe structure damage feature.

The artificial neural network (ANN) model is robust, adaptive and fault tolerant (Kao and Hung, 2003). ANN can also effectively deal with quality and uncertainty, making it highly promising for detecting structural damage. The feasibility of applying ANN and WPT to detect structural damage has received considerable attention (Sun and Chang, 2002, Yam et al., 2003, Yuen and Lam, 2006, Castro et al., 2007). However, these researches are based on information from a single sensor. Since a single sensor is generally subject to its efficiency, performance and environment noise, only limited partial signals about the structures can be collected and those signals might be incomplete, inconsistent or even imprecise. Signals from different sensors may provide complementary data in addition to the redundant information content. Merging of redundant data can help improve the imprecision; and data fusion of complementary data can create a more consistent recognition of land cover patterns, in which the associated uncertainty is reduced and the classification accuracy is improved by combining and analyzing the multi-sensor data to take advantage of their characteristics and improve the information extraction process (Smyth and Wu, 2007, Gros, 1999, Guo, 2006).

In this study, dynamic signals measured from different sensors are firstly decomposed into wavelet packet components; component energies are then calculated and fused as feature vector which are used as inputs into ANN models for damage assessment. Various levels of damage detection for this structure including the occurrence, location and severity of the damage are studied.

Section snippets

WPT

The WPT of a time domain signal f(t) can be calculated using a recursive filter-decimation operation (Coifman and Wickerhauser, 1992). After j-levels of decomposition, the original signal f(t) can be expressed asf(t)=i=12jfji(t)fji(t)=i=12jcji(t)ψj,ki(t)

Here, the component signal fji(t) can be expressed by a linear combination of wavelet functions ψj,ki(t), integers i, j and k are the modulation, scale and translation parameters, respectively; cji(t) and ψj,ki(t) are defined as the wavelet

Damage diagnosis procedure

A data fusion technique can combine data from several information sources as well as information from relative databases, to achieve a higher accuracy and more specific inferences than that could be achieved by a single source alone (Telmoudi and Chakhar, 2004). Feature fusion is one kind of data fusion; it integrates information from different sensors and obtains feature vectors (Chen and Jen, 2000).Since the neural network is very suitable to feature fusion detection, a damage diagnosis

Damage diagnosis example

The structure shown in Fig. 2(a) is a four-story, two-bay by two-bay, 12 degree of freedom (DOF), steel-frame quarter-scale symmetrical model structure in the Earthquake Engineering Research Laboratory at the University of British Columbia (UBC) (Johnson et al., 2004). It has a 2.5 m×2.5 m plan and is 3.6 m tall. The members are hot rolled grade 300 W steel with a nominal yield stress 300 MPa (42.6 kpsi). The excitation is low-level ambient wind loading at each floor in the y direction. To take the

Conclusions

This paper studies the use of the wavelet packet transform, the multi-sensor feature fusion and the neural network model for damage assessment of civil engineering structures. It is assumed that reliable structural models for healthy and damage conditions are available. Measured acceleration signals are first decomposed into component signals using the WPT, then selected component energies are fused and used as inputs to the ANN models for various levels of damage assessment.

A numerical study

Reference (19)

There are more references available in the full text version of this article.

Cited by (74)

  • Seismic fragility analysis of steel moment frames using machine learning models

    2023, Engineering Applications of Artificial Intelligence
  • Information fusion for automated post-disaster building damage evaluation using deep neural network

    2022, Sustainable Cities and Society
    Citation Excerpt :

    Based upon the nature of the feedback mechanism, information can be passed through the connections between neurons in each layer, which can drive the input data into a more creative and analytical component (Wu et al., 2021). Due to the robustness, adaptiveness, and strong ability in handling uncertainty and large data, the wide applications of DNNs have been found in civil engineering, such as to forecast mechanical properties of construction material (Bui, Nguyen, Chou, Nguyen-Xuan & Ngo, 2018; Nguyen, Kashani, Ngo & Bordas, 2019), to detect and diagnosis structural damage for early warning (Abdeljaber et al., 2018; Liu, Ju, Duan & Zhao, 2011), to predict participants’ behavior in the BIM-based project (Pan & Zhang, 2020a), and others. Apart from the merits of DNNs (i.e., the robustness, adaptiveness, and uncertainty handling ability), it is noteworthy that DNNs are also outstanding in their great generalization and transferability to learn feature combinations at multiple levels of abstraction directly from raw data (Bengio, 2012).

  • A deep learning-based approach for condition assessment of semi-rigid joint of steel frame

    2021, Journal of Building Engineering
    Citation Excerpt :

    In contrast, some signal processing techniques and statistical methods are used for the feature extraction under the non-parametric approaches. Hilbert-Huang transform (HHT) [3], wavelet transform (WT) [17–19], principle component analysis (PCA) [20–22], autoregressive modelling [23,24] are often utilized as feature extraction tools in signal-based damage identification approaches. Numerous classifiers such as artificial neural networks (ANN) [20,25–28], probabilistic neural networks (PNN) [29,30], fuzzy neural network (FNN) [31], singular value decomposition (SVD) [32], and support vector mechanism (SVM) [33] are used to learn the correlation between extracted features and corresponding damage states.

View all citing articles on Scopus

Foundation item: Project (2005E205) supported by the Natural Science Foundation of Shaanxi Province

View full text