Encyclopedia
Please note this is an old version of this entry, which may differ significantly from the current revision.
Subjects:
Engineering, Civil
For the purpose of maintaining and prolonging the service life of civil constructions, structural damage must be closely monitored. Monitoring the incidence, formation, and spread of damage is crucial to ensure a structure’s ongoing performance.
- structural health monitoring
- deep learning
- vibrational damage detection
1. Introduction
Large-scale infrastructures have seen accelerated aging and deterioration of functionality in the current era of urbanization and climate change. The structures’ operational states have frequently been disrupted as a result of rising population and traffic, unanticipated natural disasters, and human-caused damage, in addition to countless instances of catastrophic breakdowns. In the context of civil engineering, damage can be defined as a change in a system’s geometric or material properties that has a negative impact on its performance, safety, dependability, and operational life [1]. This definition states that damage does not always translates into a system total failure, but rather a relative decline in system functionality leading to subpar performance [2]. Furthermore, the damage may build up until it reaches the failure state if no corrective action is done. Depending on the type of damage, systems may fail suddenly or gradually [3].
Different system damage detection (SDD) techniques are available depending on their data analysis approach. Visual inspections are the foundation of traditional methods for diagnosing damage to civil constructions. However, a number of issues prevent the practical adoption of these approaches [4]. The size of civil structures is generally quite vast, making routine inspection tedious, time-consuming, and costly. Moreover, because traditional methods rely on human judgment, experienced and highly trained labor is unquestionably needed. Finally, some structural defects are difficult to detect, even for human experts, resulting in unnoticed damages until the structure becomes impaired. In order to overcome these issues, a promising approach is to rely on vibration-based SDDs [5]. These systems ultimately aim to overcome the issues associated with the conventional SDD approaches by providing a methodical, practical, and consistent way of identifying the presence, as well as locating and quantifying the severity, of the structural damage based on the vibration response of the monitored structure. Moreover, vibration-based SDDs can be divided into time domain and frequency domain methods. The former relies only on a system’s temporal properties, usually by taking as input the raw or filtered data coming from physical sensors distributed on the structure. These approaches are conceptually simpler to use, but, because they are more susceptible to noise contamination and environmental influences, they show significant performance issues for large buildings [6]. In fact, phase shifts between the vibration responses of the undamaged and damaged structures occur in the time domain as a result of changes in natural frequencies brought on by damage. In order to obtain this information, response functions of both damaged and undamaged structures are subtracted from one another, which results in a beating phenomenon in the presence of a frequency mismatch. Even so, meaningful findings can only be obtained if the experimental hardware is kept constant during the whole study. This recommends the usage of smart structures that can both actuate and monitor the vibrations of the host structure using piezoceramic patches that are surface mounted on, or implanted in, the structure [7]. On the other hand, frequency domain techniques exploit a system’s frequency properties. Although discrete variations in the natural oscillation frequency might not be enough to uniquely identify the damage location, changes in structural frequencies may be a reliable indicator of the existence of damage [8]. In order to correlate analytical data with experimental data, frequency-domain-related damage detection algorithms often retrieve damage characteristics by modifying an analytical reference model. Unfortunately, computational inaccuracies are often introduced because the analytical model can only approximate the behavior of a real mechanical structure [9]. Some of the existing research, such as [10,11], shows that, in the field of frequency-based damage identification, either a small subset of the first few model frequencies or all frequencies recorded during measurement are used for damage detection. Nevertheless, the field is plagued by a lack of systematic investigations for choosing modal frequencies, which could be a better tool for identifying damage in practical applications, where errors or mismatches between experimental and numerical analyses cannot be disregarded since they are sensibly affected by data noise. Finally, time–frequency (TF) algorithms [12] are a promising way to have the best of both worlds because, in contrast to pure time or frequency domain approaches, they can recognize both the frequency components of the signal and their time-variant features, resulting in efficient and valuable tools to extract structural health information. Short-time Fourier transform (STFT) [13] and Wigner–Ville distribution [14] are two common TF techniques. The wavelet transform (WT), which divides a signal into a number of local basis functions, is the most popular TF technique [15].
As a way to enhance the identification of time-varying parameters of structures, synchrosqueezing transform (SST), a promising and flexible WT, is investigated. SST, as opposed to standard TF techniques such as STFT or continuous wavelet transform (CWT) that do not take advantage of the signal sparsity, adds a nonlinear post-processing mapping to a conventional STFT or CWT representation. The mapping results in a condensed and sparse TF representation of the signal, pushing the energy content into the STFT’s most noticeable frequencies [16]. Any structure’s sensitivity to change with any modification in the structural attributes determines its damage characteristics in relation to its modal frequency. In this context, machine learning (ML) approaches can be successfully used to discriminate between damaged and undamaged structures, thanks to recent advances in both the developed algorithms and computational power. These techniques include, but are not limited to, convolutional neural networks (CNNs), support vector machines, and self-organizing maps, among others. In general, those methods can be applied with success to both time domain and frequency domain data. However, these approaches rely on the availability of structured and consistent data, often performing poorly if the input data are noisy or redundant, or if the training set is not big enough to generalize.
2. Vibration Damage Detection Methods
The traditional method of structural monitoring (visual inspection) entails hiring a qualified structural inspector to examine the building, spot problems, and put proper maintenance plans in place. Although arduous, subjective, and prone to error, this kind of manual structural inspection uses up a significant portion of the annualized maintenance budget. In order to overcome the issues related to manual visual inspection, structural health monitoring (SHM) offers a sensor-driven real-time inspection method [17]. Based on the collected data, vibration-based SHM approaches provide realistic solutions for tracking the time-varying behaviors of aging buildings. Instrumentation and data collection, condition evaluation, damage detection, and damage prognosis are the four main components of vibration-based SHM. The structures are initially equipped with a variety of sensors to gather useful measurements (such as acceleration and displacement). Using various system identification and damage detection techniques, the gathered data are then evaluated to determine the structure’s state and spot any changes. Then, SHM exploits a variety of prognosis approaches, maintenance, and retrofitting solutions to determine the remaining structure’s usable life and the actions required to enhance its structural condition. Most vibration data are typically gathered via a dense network of wired sensors placed throughout the structure. However, because installing cables requires much effort, they are not a practical or affordable solution for towering buildings or long-span bridges. For this reason, smart wireless sensors have been proposed as a solution to the drawbacks of wired sensors [18]. Another possible solution is to rely on test vehicles equipped with sensors to ease the process of data acquisition, with promising results [19].
Recently, SHM has been accomplished by the analysis of images and videos taken by modern sensors such as cameras, robotic sensors, telephones, and drones [20]. Regardless of the type of sensor, the accuracy of the existing damage diagnosis and localization algorithms heavily depends on the accessibility of several high-quality sensors and datasets. This is a significant barrier that often prevents the application of SHM on big structures such as buildings and bridges. In order to identify changes in the vibrational qualities of modal parameters (such as frequency, damping, and mode shapes) or physical parameters (such as stiffness, damping, or mass), the essential idea underlying vibration-based damage detection (VDD) systems is to use pattern recognition algorithms [21]. Any modification of these characteristics through time can be potentially correlated to structural damages. These techniques are simpler to use, but they pose problems for large construction projects because of noise pollution and environmental issues [22]. Moreover, although discrete variations in natural frequencies can in principle be used to pinpoint the precise position of structural damage, in practice this may not always be the case because a crack in two separate places could have the same frequency variation regardless of where it is. Time–frequency methods [23] can recognize the signal frequency components and comprehend their time-variant properties, in contrast to time-only and frequency-only domain techniques. The hidden information in the data that is missed by stand-alone temporal or frequency domain approaches can be found and tracked using TF methods applied on VDD [24]. Short-time Fourier transform is a popular version of Fourier transform that enables the investigation of nonstationary signals in the TF domain (STFT) [25]. The Fourier transform of a fixed windowed signal serves as the basis for STFT. Only a small portion of the signal is examined by this windowing technique at each time step t, after which a 2D signal depending on time and frequency components is obtained. Moreover, the time–frequency resolution of the STFT technique is inversely correlated with window length. While lengthening the window improves frequency resolution, it also hinders the representation’s ability to monitor frequencies. Due to the fact that the chosen window size is identical for all frequencies, one of the greatest disadvantages of STFT is that a high resolution in both time and frequency cannot be simultaneously achieved [26]. By generalizing the connection between a nonstationary, time-variant process power spectrum and autocorrelation function, the Wigner–Ville distribution can be obtained [27], as a signal can be represented in a high-resolution TF space using the Wigner–Ville distribution.
Another way to frame VDD is as a pattern recognition task. In fact, pattern recognition-based VDD algorithms’ main goal is to identify patterns in the features of damaged and undamaged structures under the same operational and ambient conditions. Different time-series modeling and machine learning methods have been applied to pinpoint the crucial VDD properties [28]. Time-series modeling is one of the most commonly used techniques for identifying structural degradation. The fundamental steps of this approach are the creation of a time-based model, the assessment of model coefficients, and the computation of residual errors; any divergence in the coefficients or residual errors might be interpreted as a structural damage. Another popular VDD technique is based on auto-regressive (AR) models. The AR models make the assumption that the observations contain noise and thus sample their modeling error from a Gaussian distribution. Several variations of the AR models, such as the auto-regressive 15 moving-average (ARMA) [29], the AR-integrated moving-average (ARIMA) [30], and the AR model with exogenous input (ARX) [31] are widely employed in SHM as well as for damage detection. Moreover, the Mahalanobis distance, when applied to time-series pattern recognition, has demonstrated encouraging results, for instance, when [32] tested several distance metrics for damage identification. Finally, out of the several available ML methods, other popular approaches such as artificial neural networks (ANN) [33], support vector machines [34], random forests [35], and clustering techniques [36] have been extensively employed in the VDD field, with reviews that deepen their usage for civil structure health monitoring [37].
Traditional VDD approaches are based on the stationarity assumption of the vibration response and selection of modal orders. However, these approaches have trouble identifying the nonstationary component of vibration response that comes from natural hazards. A structure’s intrinsic flaws also contribute to the frequency-dependent nonstationarity of the response, along with excitation-induced nonstationarity. In fact, when there is an amplitude- and frequency-dependent nonstationary response, it becomes much more difficult to identify the damage [38]. This problem can be handled by employing sophisticated TF techniques. WT is an enhancement over-fixed window-based STFT and offers the fundamentals of the conventional Fourier transform with flexible window placement and size [39]. It can be divided into discrete and continuous wavelet transform (CWT) [40], and provides the flexibility of combining high time and frequency resolutions with a suitable basis function. Many VDD applications, including signal noise filtering, data compression, and pattern recognition, use the CWT signal processing technique. To identify sudden changes in a time-variant system, a wavelet-based frequency response function [41] has been adopted. A suitable approach is to first process the signal with a CWT, and then employ the generalized discrete Teager–Kaiser energy operator to locate and magnify the modes of a damaged structure [42]. More in detail, in the previous work the operational deflection forms of the structure were also obtained using a state-of-the-art method that applied joint approximation diagonalization of the power spectral density matrices. However, CWT and CCWT require a significant increase in frequency resolution to detect minute damage because of sensitivity issues when dealing with minute frequency changes in the structures.
Another possible way to perform damage detection is to rely on deep learning algorithms trained in a supervised setting. In fact, for the establishment of a statistical model during the training phase, supervised learning methods need labeled data for both undamaged and damaged categories. The identification, classification, and quantification of the damage are usually carried out using DL techniques such as CNNs, fully convolutional networks (FCNs), or recurrent neural networks (RNNs) in VDD-based literature. For effective bridge SHM, a sparse coding-based CNN method with wireless sensors was investigated [43]. In order to extract high-level features from acceleration data, sparse coding was applied as an unsupervised layer for unlabeled data. For a three-span bridge that was instrumented using wireless sensors, several levels of damage situations were taken into account. The proposed method outperformed previous techniques such as logistic regression and decision trees, with a final accuracy equal to 98%98%. Another possibility is to identify structural degradation directly using CNNs [44]. To find the best CNN settings, 50 network topologies with different hyper-parameters were examined. Recently, the use of autoencoder data compression and a one-dimensional (1D) CNN for anomaly detection in a lengthy suspension bridge has been proposed [45]. Moreover, a CNN-based VDD technique for damage identification in compressed data has also been proposed [46]. In a following study [47], a 1D CNN technique was employed on three structural assemblages: an iron girder, a short steel beam bridge, and a long steel viaduct bridge to detect changes in material properties. Finally, an alternative approach [48] has been developed, in which the time history of a vibration signal can be given as input directly into CNNs, requiring only basic array operations and a shallow architecture with fewer hidden layers. All these methods are characterized by the use of a 1D CNN trained from scratch. In order to exploit the expressive power of pre-trained CNNs, the vibration data need to be transformed into images by using time–frequency transformations; the generated colormap can be then used as input for a pre-trained 2D CNN [49]. In order to provide the necessary labeled data, a possible solution is to simulate a scaled-down bridge model [50], while exploiting a 2D-CNN architecture that reaches up to 97% accuracy while being able to distinguish damage from structurally symmetrical locations. A recent work [51] successfully applied a TF transformation on signals coming from a real railway bridge that underwent a retrofitting procedure to perform an anomaly detection task. Other approaches with their pros and cons are reviewed in a survey paper [52].
This entry is adapted from the peer-reviewed paper 10.3390/s23136152
This entry is offline, you can click here to edit this entry!
