Real-time vibration-based structural damage detection using one-dimensional convolutional neural networks
Introduction
Engineering structures have always been susceptible to various kinds of damage (deterioration, degradation, corrosion, fatigue, creep, shrinkage, etc.) during their service life due to environmental, operational and human-induced factors. With their relatively large size, damage inspection of civil infrastructure has been reported to be laborious and expensive. Yet, the civil structures need to be inspected regularly to remain operational, improve the lifecycle performance, avoid catastrophic failures and protect human lives.
When damaged, the material and geometric characteristics of a structural component change, affecting the stiffness and stability of the structure. Conventional damage assessment methods, which depend on periodic visual inspection of structures are not efficient especially for complex structures as they require highly-trained labor and easy access to the monitored structural members. Detecting, locating and quantifying the structural damage in civil infrastructure have remained a constant challenge for researchers and engineers. Therefore, a significant amount of research has been conducted to develop automated local and global structural health monitoring (SHM) techniques [1].
Global (i.e. vibration-based) damage detection methods are used to assess the overall performance of the monitored structure by translating its vibration response into meaningful indices reflecting the actual condition of the structure. The ultimate goal of vibration-based methods is to identify the presence, severity, and location of the damaged areas by processing signals measured by a network of accelerometers. Vibration-based techniques can be classified into parametric (model-based) and nonparametric (signal-based) approaches. In parametric methods, system identification algorithms are utilized to determine the modal parameters such as natural frequencies and mode shapes from the measured response. Changes in these parameters with respect to the parameters identified for the undamaged case are used to recognize the structural damage. On the other hand, nonparametric approaches employ statistical means to identify damage directly from the measured signals.
During the last two decades, machine learning algorithms have been extensively used by researchers to develop a wide range of parametric and nonparametric vibration-based structural damage detection techniques. The vast majority of machine learning based damage detection methods available in the literature involve two processes, feature extraction and feature classification.
Parametric machine learning based methods available in the literature use different techniques to extract the modal parameters from the measured vibration response for several undamaged and damaged cases. In other words, the features extracted in these methods are simply the dynamic characteristics of the structure such as natural frequencies and mode shapes [2], [3]. On the other hand, in nonparametric machine learning based damage detection methods, several signal processing and statistical analysis techniques have been utilized for feature extraction. Additionally, various classifiers have been used in both parametric and nonparametric methods to classify the extracted features. Section 2 presents a brief review on the feature extraction and classification techniques that have been used in both parametric and nonparametric structural damage detection methods.
Therefore, a classical signal-based structural damage detection approach typically consists of a continuous acquisition of signals by sensors, extraction of certain (hand-crafted) features and feature classification by a classifier. Accordingly, it is imperative to extract damage-sensitive features correlated with the severity of the damage in the monitored structure, and to have a well-configured and trained classifier that has the utmost ability to discriminate those features. This is why the choice of both features extracted and the classifier used usually depends on a trial-and-error process for a particular structural damage detection application and significantly varies in the literature. As discussed in Section 2, such fixed features/classifiers that are either manually selected or hand-crafted may not optimally characterize the acquired signal and thus cannot accomplish a reliable performance level for damage detection. In other words, which feature extraction is the optimal choice for a particular signal and classifier remains unanswered up to date. Furthermore, feature extraction usually turns out to be a computationally costly operation, which eventually may hinder the usage of such methods for a real-time SHM application.
This paper proposes a fast and highly accurate nonparametric vibration-based algorithm for structural damage detection based on adaptive 1-D Convolutional Neural Networks. The main objective is to identify and locate any structural damage in real-time by processing the raw vibration signals acquired by a network of accelerometers. With a proper adaptation over the traditional CNNs, the proposed approach can directly classify the accelerometer signal without requiring any feature extraction, pre- or post-processing. Consequently, this will lead to an efficient system in terms of speed, allowing a real-time application. Due to the CNNs ability to learn to extract the optimal features, with a proper training, the proposed system can achieve a superior damage detection and localization accuracy despite the noise-like and uncorrelated patterns of the accelerometer signal. Some samples of the latter are shown in Fig. 1. Furthermore, this study will demonstrate that simple CNN configurations can easily achieve a high detection performance compared to the complex ones commonly used for deep learning tasks over such complex and uncorrelated signals that can even defy a human expert inspector.
The proposed damage detection method is evaluated experimentally using Qatar University (QU) Grandstand Simulator. This structure is one of the largest lab structures used for testing machine learning based damage detection algorithms. Additionally, the structure is equipped with a large number of accelerometers compared to similar studies in the literature, which allowed the authors to test the CNN-based algorithm under several structural damage cases.
The rest of the paper is organized as follows. Section 2 includes a brief review of the recent applications of CNNs. Section 3 discusses the structural design and instrumentation of QU Grandstand Simulator. Overview, adaptation, and the back-propagation training of 1D CNNs are presented in Section 4. The proposed CNN-based damage detection algorithm is explained in Section 5. The experiments conducted to demonstrate the algorithm using QU Grandstand Simulator along with the results and performance evaluation are provided in Section 6. Finally, Section 7 concludes the paper and suggests potential topics for future research.
Section snippets
Related work
There have been numerous parametric and non-parametric structural damage detection methods proposed in the literature. Among many parametric methods, the most common classifiers are multi-layer feedforward artificial neural networks (ANNs) [4], [5], [6], [7], [8], [9], [10]. Additionally, researchers have implemented other classifiers in parametric methods such as online sequential extreme learning machine (OS-ELM) algorithm [11], probabilistic neural networks (PNNs) [12], [13], and fuzzy
Qatar University grandstand simulator
Extensive analytical and experimental studies on structural health monitoring and vibration serviceability of stadia have been carried out at Qatar University. One of the main objectives is to develop efficient structural damage detection techniques that are suitable for monitoring of modern stadia. Before the application of the newly-developed techniques to real-life structures, it is pertinent to verify them experimentally in a controlled laboratory environment. Experiments performed using
Overview of CNNs
CNNs are biologically inspired feed-forward ANNs that present a simple model for the mammalian visual cortex. They are now widely used and have become the de-facto standard in many object and event recognition systems in an image or video. Fig. 6 illustrates a 2D CNN model with an input layer accepting 28×28 pixel images. Each convolution layer after the input layer alternates with a sub-sampling layer, which decimates the propagated 2D maps from the neurons of the previous layer. Unlike
The proposed damage detection and localization method
As discussed earlier, in this study, structural damage is simulated experimentally by loosening the bolts connecting the filler beams to the mean girders. The objective of the CNN-based algorithm is to detect the damage (if any) and identify the location of the damaged joint(s) accurately. The proposed algorithm requires designing and training a unique 1-D CNN for each one of the 30 joints instrumented with accelerometers. Each CNN is responsible for assessing the condition of one joint using
Experimental results
In this section, the efficiency of the proposed CNN-based damage detection algorithm is evaluated using the QU grandstand simulator. The experimental work was carried out in two phases. In the first phase, for simplicity, only a single girder on the steel frame ( joints) was monitored. In the second phase, the performance of the damage detection approach was tested utilizing the entire structure ( joints).
Conclusions and future work
This paper presented a novel damage detection approach with the adaptive implementation of 1D Convolutional Neural Network (CNNs). The proposed system was verified experimentally by monitoring the main steel frame of QU grandstand simulator. The results of the small and large-scale tests demonstrate a high performance level for real-time SHM and structural damage detection processes. Based on the experimental results presented in this study, the following conclusions can be drawn:
- •
This study
References (35)
- et al.
Fault diagnosis on beam-like structures from modal parameters using artificial neural networks
Measurement
(2015) - et al.
Damage detection of truss bridge joints using artificial neural networks
Expert Syst. Appl.
(2008) - et al.
Neural networks-based damage detection for bridges considering errors in baseline finite element models
J. Sound Vib.
(2005) - et al.
Structure damage diagnosis using neural network and feature fusion
Eng. Appl. Artif. Intell.
(2011) - et al.
Machine learning algorithms for damage detection: kernel-based approaches
J. Sound Vib.
(2016) - et al.
Reference-free damage detection using instantaneous baseline measurements
AIAA J.
(2009) - et al.
Iterated square root unscented Kalman filter for nonlinear states and parameters estimation: three DOF damped system
J. Civil. Struct. Health Monit.
(2015) - M.Mansouri, O.Avci, H.Nounou, M.Nounou, A comparative assessment of nonlinear state estimation methods for structural...
Application of Bayesian-designed artificial neural networks in Phase II structural health monitoring benchmark studies
Aust. J. Struct. Eng.
(2014)- et al.
Prediction of unmeasured mode shape using artificial neural network for damage detection
J. Teknol. (Sci. Eng.)
(2013)
Damage detection on a three-storey steel frame using artificial neural networks and genetic algorithms
Meccanica
Damage detection in beams using spatial fourier analysis and neural networks
J. Intell. Mater. Syst. Struct.
Online sequential extreme learning machine for vibration-based damage assessment using transmissibility data
J. Comput. Civil. Eng.
Damage localization of cable-supported bridges using modal frequency data and probabilistic neural network
Math. Probl. Eng.
Structural damage detection by integrating data fusion and probabilistic neural network
Adv. Struct. Eng.
Unsupervised fuzzy neural networks for damage detection of structures
Struct. Control Health Monit.
Bridge damage severity quantification using multipoint acceleration measurement and artificial neural networks
Shock Vib.
Cited by (966)
Prediction of California bearing ratio and modified proctor parameters using deep neural networks and multiple linear regression: A case study of granular soils
2024, Case Studies in Construction MaterialsUtilizing nanotechnology to boost the reliability and determine the vertical load capacity of pile assemblies
2024, Environmental ResearchOperational transfer path analysis based on neural network
2024, Journal of Sound and VibrationBridge damage localization and quantification using deep learning and FEM static simulation
2024, Mechanical Systems and Signal Processing