**Sheng Li 1,\*, Xiang Zuo 2, Zhengying Li <sup>2</sup> and Honghai Wang <sup>1</sup>**


Received: 4 January 2020; Accepted: 6 February 2020; Published: 8 February 2020

**Abstract:** Improving the accuracy and efficiency of bridge structure damage detection is one of the main challenges in engineering practice. This paper aims to address this issue by monitoring the continuous bridge deflection based on the fiber optic gyroscope and applying the deep-learning algorithm to perform structural damage detection. With a scale-down bridge model, three types of damage scenarios and an intact benchmark were simulated. A supervised learning model based on the deep convolutional neural networks was proposed. After the training process under ten-fold cross-validation, the model accuracy can reach 96.9% and significantly outperform that of other four traditional machine learning methods (random forest, support vector machine, k-nearest neighbor, and decision tree) used for comparison. Further, the proposed model illustrated its decent ability in distinguishing damage from structurally symmetrical locations.

**Keywords:** bridge damage detection; fiber optic gyroscope; deep learning; convolutional neural network

#### **1. Introduction**

Dynamic modal analysis has been the most commonly used approach for structural damage detection in civil engineering [1–3]. The use of wavelet, Hilbert–Huang transform, and other signal processing methods are also the conventional choices for structural damage detection that directly analyze the perturbation of vibration signals [4]. Various structural non-destructive testing approaches [5–7] are also significant means for detecting structural damage. Over the last decade, machine-learning algorithms have been used to address a wide range of vibration-based damage detection problems [8,9]. Although most of these techniques are based on vibration responses and such approaches still dominate the diagnosis and prognosis of structural health monitoring [10], feature extraction processes heavily relying on handcrafted intervention prior to damage classification [11] have often become major challenges that limit the effectiveness of various methods.

With the ability of automatic feature extraction and classification, deep convolutional neural networks (CNN) have been explored to address the range of difficulties in such following areas as computer vision [12,13], speech recognition [14], natural language processing [15], medical image processing [16], pathological signal classification [17,18], mechanical fault diagnosis [19–21], impact evaluation of natural disasters on infrastructure systems [22], and structural damage detection [23–25].

Most of the research efforts on deep CNN-based structural damage detection are essentially associated with the supervised learning processes. In this emerging area, Cha et al. [23] pioneered the deep CNN study of damage detection for cracks in concrete structures, and subsequently, Cha et al. [26]

further expanded the detection objectives of structural damage based on the faster Region-CNN (R-CNN). The recent study based on R-CNN to quantify the identified concrete spalling damage in terms of volume was reported in [27]. Xue and Li [28] established a fully convolutional neural networks model to classify the concrete tunnel lining defects. Other image-based researches on structural damage detection using deep learning were reported in [29–31]. In addition to two-dimensional convolution operations on structural images, one-dimensional convolution operations which usually spend a considerably cheaper computational cost than that of recurrent neural networks [32], are employed by researchers to perform the signal-based structural damage detection. The structural vibration signal as a typical type of one-dimensional time series [33] data is used to perform deep CNN-based structural damage detection. For instance, Abdeljaber et al. [18] proposed a method for detecting structural damage using one-dimensional CNN for multi-nodal vibration testing of steel frames. Lin et al. [25] simulated the vibration response of simply supported beams under various damage scenarios and proposed a procedure for detecting the categories of structural damage using the one-dimensional CNN. Huang et al. [34] analyzed the mechanical operation process through vibration signal by constructing a one-dimensional CNN.

Although deep CNN technology to some extent relieves the heavy pre-processing on the raw data or feature crafting for the damage detection when using vibration signals, the analysis of bridge damage detection is more complex comparing to that of simple structures, which needs more support of structural responses. In other words, compared with the amount of the degree of structural freedom, the scale of available vibration sensors used for bridge structural damage detection are often finite or even insufficient. To obtain as much structural dynamic information as possible in the case of limited measurement points, sensor optimization layout [35] is generally considered, which results in a decrease in damage detectability of complex structures. Therefore, a novel type of structural response which can easily cover the whole and local test requirement and provide enough structural information for the analysis of damage detection by using deep CNN should be attempted and explored. In this paper, we aimed at the multi-dimensional type of signal and chose a test technique for continuous curve mode of deformation based on fiber optic gyroscope (FOG) to produce continuous deflection of bridge. Detailed fundamental principles of the FOG-based testing technique were reported in [36–38]. A corresponding sample set based on supervised learning techniques was established, and a specific one-dimensional CNN model was proposed to automate feature extraction and classification. Specifically, the scheme procedure for the production of structural damage scenarios based on deformation responses was elaborated. The deformation responses and corresponding output labels were established through data augmentation and pre-processing. Furthermore, architectures and algorithms of the proposed one-dimensional CNN, and the partition rules of the dataset used for method verification were discussed. Finally, the performance of the proposed approach was compared with the results of other pattern recognition methods, all of which were conducted under the ten-fold cross-validation [39].

#### **2. Design and Implementation of Structural Damage Scenarios**

#### *2.1. Experimental Platform and Instrumentation*

A scale-down model of cable-stayed bridge was used as the experimental platform to represent the responses due to the simulation damage. The model shown in Figure 1 with the main span of 9.7 m, tower height of 3.46 m, deck width of 0.55 m, and 56 stay cables, was manufactured at a scale of 1:40. Figure 2 further illustrated a structure diagram and a physical structure of the device dedicated to measuring continuous deflection of the bridge model, which was triggered by a remote controller to perform mobile measurement. When the test device [36] moved along the surface of the bridge model, data of continuous deflection can be collected at sampling rate 150 Hz. Since the measurement period based on the motion carrier was relatively short, the continuous deflection obtained chronologically at each time was regarded as a multi-dimensional variable acquired at the same moment, and the deflection of main span was chosen as the structural deformation response used for subsequent analysis.

**Figure 1.** Experimental platform for testing continuous deflection.

**Figure 2.** Measuring device integrated in a motion carrier.

### *2.2. Experimental Design and Procedures*

The change in structural geometry can reflect a certain degree of transformation of interior mechanical properties. Further, structural damage is one reason for the change of interior mechanical properties of structure. Therefore, the different damage scenarios of the structure theoretically have corresponding structural deformation states. Continuous deflection can provide the dense deformation information, which can present more abundant structural response information than other finite point-based geometry measurement methods [40–42]. In the context of an experiment based on supervised learning, it was assumed that the change in the continuous deflection of bridge was only due to the result of structural damage. A metal pad (42.8L × 12.8W × 0.2H cm) with slope at both ends was used to simulate structural deformation caused by damage rather than physically destroying the structure [43]. The pad as an obstacle was placed on the movement path of the measuring device to simulate the deformation caused by structural damage. Compared with the situation without the pad, the measuring device can capture responses of the continuous deflection of bridge under the disturbance of the pad. This localized continuous deflection caused by the influence of the pad was clearly the most important of the continuous deflection of the entire bridge. Using such local responses instead of the global deflections can undoubtedly simplify the training process of the following supervised learning algorithm.

By this way, as shown in Figure 3, when the pad was not placed, the corresponding continuous deflection of the bridge was defined as **U**0. For each of the three damage scenarios, one pad was placed at a position each time, and therefore, **U**1, **U**2, and **U**<sup>3</sup> can be obtained. Here, **U**0, **U**1, **U**2, and **U**<sup>3</sup> as raw data of continuous deflection represented four types of simulated structure states, respectively. To improve the training efficiency and save the computational overhead of the supervised learning, **U**1, **U**2, and **U**<sup>3</sup> were truncated to **u**1, **u**2, and **u**3. Such truncated selection in the areas affected by the pad can be estimated through both the original testing curves and the dimension of pad in the context of experiment based on supervised learning. In the intact and three damage scenarios, the actual

benchmarks of **u**1, **u**2, and **u**<sup>3</sup> were **u**01, **u**02, and **u**03, respectively. A common baseline for the three damage scenarios was defined to facilitate analysis. The weights of **u**01, **u**02, and **u**<sup>03</sup> were regarded as equal and their average **u**<sup>0</sup> was designated as the nominal benchmark of **u**1, **u**<sup>2</sup> and **u**3. The following work utilized **u**0, **u**1, **u**2, and **u**<sup>3</sup> to conduct the damage detection based on deep CNN algorithm.

**Figure 3.** Locations of the metal pad used to assist in simulated damage scenarios.
