Comparative analysis of twelve transfer learning models for the prediction and crack detection in concrete dams, based on borehole images

Abstract

Disaster-resilient dams require accurate crack detection, but machine learning methods cannot capture dam structural reaction temporal patterns and dependencies. This research uses deep learning, convolutional neural networks, and transfer learning to improve dam crack detection. Twelve deep-learning models are trained on 192 crack images. This research aims to provide up-to-date detecting techniques to solve dam crack problems. The finding shows that the EfficientNetB0 model performed better than others in classifying borehole concrete crack surface tiles and normal (undamaged) surface tiles with 91% accuracy. The study’s pre-trained designs help to identify and to determine the specific locations of cracks.

Graphical abstract

Keywords

concrete dam / borehole closed-circuit television / deep learning models / crack detection / water resources management

1 Introduction

The visual presence of cracks is a manifestation of construction aging of infrastructure such as dams, roads, and buildings. Dams and their associated reservoirs are significant civil infrastructure in hydropower and water storage facilities, providing reliable sources of clean water and sustainable energy [1,2]. Design and construction of these structures must ensure their dependable functionality both during normal operating circumstances and when exposed to severe load combinations such as floods, rainstorms, earthquakes, and ice-related incidents. Notably, many dams are often located in topographically challenging areas such as mountainous terrains and valleys, exposing them to demanding geological and other natural circumstances [3,4].

The formation of surface cracks is a common concern in concrete structures, with concrete dams being particularly susceptible. These cracks may arise at different stages, including during construction, operation, overhaul, and reinforcement. The existence of cracks substantially affects the stress state of dam constructions, and undermines their structural soundness and material longevity. This ultimately results in a decrease in the strength and load-bearing capacity of the structures, as evidenced by Refs. [5,6]. In the absence of timely detection, minor cracks can expand and progress into invasive structural impairments, thereby constituting a peril to the integrity of the dam. The presence of cracks resulting from structural deformation, seismic activity, water flow impact, and other factors, poses a potential safety risk.

Generally, for a considerable period in the past, structural impairments have been identified through manual inspection and human field surveys [1,7]. However, specific challenges, such as limited reproducibility of findings and elevated labor expenses, significantly impede the precision and effectiveness of manual examination [8]. A hydraulic project usually incorporates a range of hydraulic structures, such as for retaining, discharging, conveying, remediation, and for other specialized purposes. Manual inspection, dependent on the expertise of engineers, is a process that is both costly and time-intensive, and subject to erroneous diagnoses. In addition, hydraulic structures that have been in service for an extended duration may manifest fissure-like imperfections due to extrinsic environmental corrosion and non-uniform illumination circumstances. Inspecting and detecting cracks in harsh environments, especially, poses significant logistical, labor-intensive, and costly challenges.

The research introduces a novel rigid-block discrete element method (RB-DEM) method for simulating mesoscale fractures in concrete [9,10]. This method is user-friendly and versatile, compatible with various finite-element mesh generators. The study investigates concrete behavior under uniaxial compression, exploring the influence of aggregate properties and Interface transition zone (ITZ) characteristics on stress-strain curves and crack propagation. Using finite element analysis, Ref. [11] studied microcrack behavior in self-healing concrete. The authors explored crack initiation and propagation influenced by fracture strength variations between mortar and capsules and their work provides guidance toward efficient self-healing conditions. Reference [12] focuses on modeling fractures in self-healing concrete, exploring capsule-material interactions and thickness variations, and providing detailed two-dimensional (2D) simulations that reveal fracture behavior using different material properties as well as insights relating to resilient concrete structures. Similarly, Ref. [13] explores self-healing concrete fracture modeling. It investigates capsule-matrix interaction and thickness variations, presenting 2D fracture simulations with diverse properties. The study informs self-healing concrete design for greater resilience. A new prediction model for Ultra-high performance concrete (UHPC) compressive strength has been proposed using artificial neural network (ANN) and support vector machine (SVM) [14]. Experimental validation and parameter sensitivity analysis were conducted in that work.

Regular inspection and detection of cracks regularly is of utmost importance [15] for the upkeep and functioning of dams already in existence, but local environments can be hazardous, particularly for human operators. Therefore, automatic crack detection has garnered significant attention from scholars and professionals in managing and maintaining dam infrastructure. In recent years, significant research endeavors have been directed toward developing crack detection techniques utilizing deep learning (DL). This is primarily due to the successful application of DL in the domains of computer vision (CV) and image analysis.

Transfer learning (TL) [16] is often superior to training a Convolutional Neural Network (CNN) from scratch due to its ability to utilize models that have been pre-trained on large data sets. When using TL, a pre-trained model, such as a well-established CNN architecture like ResNet or visual geometry group (VGG), is a feature extractor that has already learned meaningful features from extensive data. By fine-tuning this pre-trained model on a specific task or data set of interest, we can adapt these learned features to the new problem. This approach significantly reduces the need for large data sets and extensive training time, making it particularly advantageous when labeled data are scarce or computational resources are limited. In contrast, training a CNN from scratch on a small data set may lead to overfitting, as the network may need help to learn robust features. TL provides a robust starting point that includes complex knowledge learned from diverse domains and from large data sets, for boosting performance on various tasks.

This review report analyses and investigates the progress of DL research in the field of crack detection and categorizes the relevant works [17–30]. We present a comparisons of DL models in the methodology and use pre-train 12 models for crack detection in borehole subsurface image data set in concrete dam in China. Additionally, we compare the extent to which the reviewed research has comprehensively presented prominent research methods, data sets, future research directions, and challenges.

1.1 Background

The formation of cracks in dam structures is a common occurrence that can be attributed to the interaction of various unfavorable factors, such as hydraulic fracturing, freeze–thaw cycles, and chemical corrosion [31–33]. Cracks can negatively impact the structural soundness, imperviousness, and stability of concrete dam structures, thereby exacerbating the rate of material deterioration [34,35]. The conventional approach to detecting cracks in dams involves manual inspection [36–39], for which reservoir water is typically drained in the initial stages of crack inspection. However, flaws in dam structures cause water to drain in the form of leakage, which causes the reservoir to fail catastrophically, whereas manual inspection has limitations in adequately monitoring the crack and leakage phenomena.

The task of image classification is a fundamental aspect of the field of CV, serving as a basis for various other CV tasks such as object detection, localization, and semantic segmentation. The principal aim of image classification is to assign images to predetermined categories. Automated systems encounter significant difficulties in image classification despite being a seemingly effortless task for human perception. The utilization of machine learning (ML) methodologies, such as ANNs, SVMs, and random forests, has become increasingly prevalent in crack detection [40–44].

Yamaguchi and Hashimoto. [45] utilized percolation-based image processing methodologies to effectively identify cracks in extensive concrete surface images. Chun et al. [46], developed an automated methodology that utilizes a two-step light gradient boosting machine to detect cracks on concrete surfaces. The model [47] emphasizes how ML techniques are limited when predefined image processing and feature extraction techniques are used prior to model training. Such methodological approaches have limitations when extracting profound features from the source images, rendering them inadequate for handling complex embedded information within the images [15].

Traditional ML methods necessitate manual feature extraction, which can be laborious and error-prone, and the techniques have depended on manually designed features extracted from images through feature extractors. These features were subsequently employed as inputs to trainable classifiers. The effectiveness of the classification task is significantly contingent upon the feature extraction process design, which has been demonstrated to present a challenging issue [48]. Crack monitoring has experienced noteworthy progressions in recent years, primarily attributable to rapid advancements in artificial intelligence. Several techniques based on vision have been devised and executed for this objective.

Researchers have been drawn to developing a more convenient and cost-effective approach to dam inspection through visual recognition-based crack detection methods utilizing cameras. DL methodologies and CNN present a more efficient and effective approach for identifying cracks end-to-end. The present work investigates the potential of DL models, specifically CNNs, in tackling the difficulties of identifying of cracks [49–52]. CNNs are a popular form of DL algorithms that have found extensive application in the field of scene recognition [53–55]. A Faster Region-based CNN (R-CNN) has been utilized to detect damage in concrete structures [56].

The authors of Ref. [57] employed a sensitivity detection network with a Faster R-CNN to identify sealed and unsealed pavement cracks. Study [58] employed the You Only Look Once (YOLO) v3 model to detect and categorize seven distinct forms of pavement distress such as cracking; distortion; disintegration; skidding hazards; and surface treatment distresses., incorporating cracking. The utilization of subsurface closed-circuit television (CCTV) borehole inspection with integration of DL algorithms enhances the automation in equipment industry as well as it minimized the time and cost of concrete dam inspection and improve the technique of crack detection in dams. Liang et al. [59] developed an active monitoring mechanism that employs a cement-based piezoelectric intelligent module array to identify and quantify underwater fractures. Chen et al. [60] has previously devised methodologies for identifying cracks in underwater concrete structures using a sensing-heating mechanism inside a permeable enclosure.

A method proposed by Si et al. [61] involves using piezoelectric and water energy-relief blasting techniques to monitor the height and depth of concrete surface damage in water engineering. Only a few of these approaches are relevant for real-world crack detection, and most research efforts in dam crack identification have been undertaken in controlled laboratory conditions. Due to the size and surface area involved, installing and burying sensors within dam structures presents a significant financial expense. It takes much work to determine the geometry and physical proportions of cracks correctly. Furthermore, these sensors not only have a limited level of durability but also have limits when it comes to precisely locating cracks in submerged dam structures.

DL algorithms can precisely identify and classify different kinds of cracks [47,62–64], including hairline cracks, by training deep neural networks on a sizable data set of annotated photographs. The application of DL to crack detection increases inspection speed and efficiency while also increasing inspection accuracy. Automated crack identification is highly effective when DL algorithms are employed [65–67], allowing engineers and inspectors to conduct thorough analyses, spot potential risks, and take preventative action to preserve the security and durability of dams. The field of dam inspections is likely to undergo significant change due to ongoing developments in DL techniques, becoming more effective, dependable, and affordable in the years to come.

To address the parameter, count, and computational complexity issue in crack detection models, lightweight models predominantly simplify the backbone, resulting in a decreased feature extraction capacity compared to general models. Furthermore, it should be noted that the geometric configuration of a crack resembles a line, and the quantity of pixels that comprise a crack represents a relatively small fraction of the overall image. As a result, the range of informative characteristics that can be utilized to analyze cracks is restricted, so that lightweight models need help to efficiently discern crucial crack information and they exhibit modest precision.

It is important to note that implementing DL models to detect cracks in dams may necessitate the use of significant computational resources. Consequently, it is imperative to possess appropriate hardware or utilize cloud-based resources for training and inference. This review proposes detecting structural cracks in dams using 12 pre-trained models inspired by the ImageNet Challenge. The methodology utilizes a data set of pillow dam borehole CCTV images. It involves categorizing the data, pre-processing the images into tiles, and organizing the data set into training, validation, and testing subsets. The pre-trained models are then applied to detect cracks in the dam. This approach aims to improve the accuracy and efficiency of crack detection and enhance dam safety.

Guo et al. [68] introduce a novel Deep Collocation Method (DCM) that utilizes computational graphs and backpropagation algorithms, diverging from traditional mechanical problem-solving approaches. Their approach involves generating batches of randomly distributed collocation points to achieve C1 continuity for Kirchhoff plate bending problems with various geometries. The author of Ref. [69] explores the application of DNNs for approximating solutions to PDEs in Computational Mechanics, employing the energy of mechanical systems as a novel loss function to tackle mechanical problems. This approach offers an innovative perspective, focusing on PDEs in an energetic format, and contributes to solving PDEs through DNN approximations, yielding valuable insights and generalized problem discussions. The paper, Ref. [70], introduces a deep autoencoder-based energy method (DAEM) for analyzing Kirchhoff plates, combining deep autoencoder features with unsupervised learning. The DAEM, implemented in PyTorch with TL, efficiently extracts patterns in energy systems, enhancing computational efficiency and accuracy.

The authors of Ref. [71] developed a stochastic DCM using neural architecture search (NAS) and TL for analyzing porous media. The approach involves sensitivity analysis, NAS, and TL for efficient groundwater flow simulation in heterogeneous aquifers, outperforming finite difference methods in accuracy and computational cost. Guo et al. [72] developed a physics-informed DL model for transient heat transfer analysis in functionally graded materials (FGMs) using Runge−Kutta integration, introducing the DCM. Our approach incorporates prior physics knowledge via temperature variables and boundary conditions, features fitted activation functions, and validates well against analytical solutions and other methods, suggesting potential applications in FGM transient dynamic analysis. Guo et al. [73] developed an adaptive DCM with enhanced computational efficiency and accuracy using fitted NAS and model-based TL. Our approach addresses the challenges of borehole inspection in subsurface of concrete dam during subsurface imaging and present and ideology to automize the manual inspection to replace with digital deep TL based algorithms for speedup concrete dam inspection and minimized the error during manual interpretation as well as automized the inspection equipment for control the project cost and acquisition time during inspection.

2 Materials and methods

2.1 Study area and data set

The Pillow Dam is a vital hydropower project located in the Jinkou District of Leshan City, situated on the south-western edge of the Sichuan basin, approximately 120 km from Chengdu in Sichuan Province, China. As shown in Fig.1, it is constructed on the Dadu River. One of the critical functions of the Pillow Dam is its significant role in flood management within the Dadu River basin, which is highly susceptible to inundation during the monsoon season. By effectively controlling the water flow, the dam helps mitigate the devastating impact of floods, ensuring the safety of lives and infrastructure in the region. Moreover, the dam facilitates irrigation activities by providing a controlled and reliable water supply to support agricultural practices in the basin.

2.2 Data set

The subsurface geological and geophysical information was obtained in real-time through borehole CCTV equipment, as shown in Fig.2, with approximately 192 image data sets from the Pillow Dam investigation site. The data summary is discussed below, in Tab.1. Each image was 3m long, providing information on sub-surface discontinuities like cracks, fractures, leakage paths, and cavities. Fig.3, illustrates examples of such images, and Fig.1 presents an overview of the image data collection site for geological discontinuity investigation in the pillow dam.

2.3 Preprocessing

This study focuses on identifying borehole subsurface cracks in concrete structure of dam which is classify into two categories, i.e., crack concrete and normal (undamaged) concrete class through subsurface images using Digital Discontinuity Analysis through Deep Dense TL (DDTL). The data set used for this investigation comprises 192 CCTV images of the Pillow Dam. The data set is initially divided into tiles measuring 400 × 400 pixels. These tiles are then subjected to pre-processing steps before further research. The data set includes 5290 normal images and 1763 images depicting cracks. These images (Fig.3) are processed using edge enhancement and masking techniques. Further details about the data set can be found in Tab.2.

2.4 Proposed methodology

The proposed approach relies on utilizing 12 pre-trained models (Fig.4), and it draws inspiration from the ImageNet Challenge and leverages advancements in CV. This proposal’s foundation lies in a Pillow Dam borehole CCTV image data set. This data set consists of images that encompass not only structural cracks but also stone tiles. By utilizing this data set, the methodology aims to effectively detect cracks, and other types of discontinuities in the dam. The acquired data from various geophysical tools is input for the 12 pre-trained models. These models have been pre-trained on diverse data and are well-suited for analyzing images. The proposed methodology enhances the accuracy and efficiency of crack detection in the dam by leveraging the knowledge embedded within these models.

The first step in the methodology involves organizing the data into two categories: crack and normal class. This categorization is based on geological and geophysical indicators such as undamaged geological rock strata and specific zones where cracks are likely found within the concrete wall in the dam’s borehole. The Pillow Dam borehole CCTV image data set undergoes a preprocessing stage. This step aims to improve the quality of the images and prepare them for analysis. The images are enhanced and transformed into slides measuring 400 × 400 tiles. This division into smaller tiles allows for a more detailed and focused examination of the dam’s structural elements. Once the data set has been preprocessed and divided into tiles based on normal (undamaged) and crack classes during the image segmentation technique, it is further organized into three subsets: training, validation, and testing.

The pre-trained models are then applied to the data set, leveraging their capabilities to detect cracks and other structural irregularities. The integration of 12 pre-trained models, coupled with the utilization of the pillow dam borehole CCTV image data set, provides a robust framework for accurate crack detection and contributes to ensuring the safety and integrity of dams. To enhance the accuracy of crack and normal (undamaged) concrete classification, each pre-trained model underwent a fine-tuning process by adding two extra layers. The first layer, the Gap layer, was employed to regulate the learning capacity of the model, aiding prevention of overfitting and optimizing the model’s performance. The second layer is a dense layer specifically designed for binary classification, enabling the model to distinguish between cracks and stones more precisely. By incorporating these additional layers, the pre-trained models were refined to improve the accuracy of crack and stone classification.

2.5 Pre-trained models

Pre-trained models, ML models trained on extensive data sets utilizing neural network architectures, are readily accessible and can be implemented in diverse applications. The models are subjected to pre-training procedures tailored to particular tasks, such as image classification, language modeling, or speech recognition. These models are intentionally optimized for specific use cases, such as ResNet, a pre-trained model frequently employed in CV applications for tasks such as image recognition, object classification, and detection. Use of pre-trained models can substantially decrease the time and resources needed to construct machine-learning models from the ground up while concurrently enhancing the precision and efficacy of the models.

2.5.1 EfficientNetB

The EfficientNet is a collection of CNNs specifically engineered to attain top-notch performance with reduced computational resources and parameters in contrast to pre-existing models. The EfficientNet model was initially presented by Tan and Le [25]. Numerous iterations of EfficientNet have been introduced since its inception, each characterized by distinct parameter magnitudes and trade-offs in terms of performance. The primary focus of this comparative study is the utilization of the primary variations of EfficientNet.

2.5.2 EfficientNetB0

The EfficientNetB0 [29] investigation focuses on the smallest iteration of EfficientNet, encompassing roughly 5 million parameters. This specific model attains the highest accuracy currently available on the ImageNet data set. Significantly, the current model is 8.4 times less voluminous and 6.1 times more expeditious than its predecessor.

2.5.3 EfficientNetB1

This investigation centers on a particular version of EfficientNet that encompasses roughly 7 million parameters. The current iteration of the model demonstrates a reduction in size by a factor of 7.8 and an increase in speed by a factor of 5.7 compared to the previous state-of-the-art model on the ImageNet data set. The approach utilized in this study exhibits a higher degree of precision than EfficientNetB0 [30] while having a notable efficiency.

2.5.4 Residual network (ResNet) 101

The ResNet [20] deep neural network features unique skip connections and heavy use of normalization. The issue of the vanishing gradient, which hinders the practical training of deep neural networks, is tackled by this approach. The ResNet architecture utilizes residual connections to facilitate seamless information propagation throughout the network. The connections above establish a contiguous route for incorporating input into the output of a designated layer, thereby enabling the smooth propagation of gradients and augmenting the learning process.

The ResNet architecture is distinguished by incorporating residual blocks consisting of several convolutional layers, batch normalization, and a shortcut connection. Residual blocks have the potential to be arranged in a manner that enables the construction of deep neural networks. ResNet has exhibited efficacy across diverse domains, encompassing image recognition, object detection, and natural language processing. Additionally, ResNet has been utilized as the basis for creating alternative architectures.

2.5.5 Mobile NetV2

The architecture of MobileNetV2 [19] is founded on the principle of separable convolutions, which may take the form of either depth-wise or point-wise convolutions. Depth-wise convolutions involve the application of a filter to each input channel, followed by using a convolutional layer to merge the results of the preceding 191 convolutions. The initial set of results is derived via a singular methodology that encompasses convolutional filtration and pooling. The procedure is bifurcated into two distinct phases: filtration and amalgamation. The factorization process yields a substantial decrease in computational requirements.

2.5.6 DenseNet

Huang et al. [18] proposed DenseNet, a series of DL models with a distinctive connectivity structure, setting it apart from other CNN models such as VGG, ResNet, and Inception. DenseNet employs a dense feed-forward connection whereby each layer is connected to all other layers. The establishment of connectivity is accomplished using condensed interconnections, whereby the output of each stratum is combined with the input of the following strata. The observed connectivity pattern differs from the conventional sequential layer linkage observed in typical feed-forward neural networks.

The DenseNet architecture’s dense connectivity pattern presents various benefits compared to traditional CNN architectures. The optimization of information transmission within a network and the facilitation of efficient feature recycling are conducive to improved training efficiency and superior generalization performance. Moreover, DenseNet assists in mitigating the challenge of the vanishing gradient problem encountered in deep neural networks, by promoting seamless gradient propagation across the network. In addition, DenseNet can diminish the number of parameters essential for network training because each layer can proficiently exploit characteristics from antecedent layers.

DenseNets have exhibited remarkable efficacy in diverse CV applications, such as image categorization, object identification, and semantic partitioning, thereby solidifying their position as the present pinnacle of achievement. Additionally, these techniques have demonstrated usefulness in various fields, including but not limited to natural language processing and speech recognition.

The DenseNet architecture comprises various versions, namely DenseNet-121 [26], DenseNet-169 [27], and DenseNet-201 [28], which exhibit differences in the number of layers and filters per layer. Frequently, these models are pre-trained on extensive data sets such as ImageNet and can be subsequently adapted for particular tasks through fine-tuning.

2.5.7 Visual geometry group

The VGG lineage comprises CNN architectures extensively employed in CV domains, such as image classification and object detection. The VGG16 [22] model was first presented and demonstrated remarkable proficiency on the ImageNet data set, a vast compilation of annotated images. The VGG16 architecture comprises 16 layers, 13 convolutional layers, and three fully connected layers.

The VGG16 [22] architecture employs convolutional layers with a 3 × 3 filter with a stride of 1. These layers are then followed by a rectified linear unit (ReLU) activation function and a 2 × 2 max pooling layer with a stride of 2. In addition to VGG16, the VGG lineage encompasses alternative models, including VGG19 [74], which comprises 19 layers, and VGG-M, which exhibits a diminished quantity of layers compared to VGG16. The structural frameworks of these models exhibit minor differences yet adhere to the same fundamental design principles.

The VGG models are renowned for their uncomplicated and uniform architecture, rendering them easily comprehensible and deployable. However, these images classification tasks necessitate substantial memory and processing capabilities and are computationally intensive. Notwithstanding these difficulties, the VGG lineage remains a noteworthy benchmark in DL. and it has exerted a significant impact on the advancement of various other CNN structures.

2.5.8 InceptionV3

The network utilized in this research is founded on inception modules and encompasses 48 layers [23]. The network’s inception blocks integrate convolutions utilizing diverse kernel sizes to extract features and consolidate the outcomes. The input image is subjected to a series of operations comprising convolution, batch normalization, and ReLU activation, succeeded by pooling and several inception layers for feature extraction. In the classification stage, overfitting is mitigated by implementing a dropout layer. In contrast, the output layers employ the SoftMax function in conjunction with cross-entropy for classification.

2.5.9 InceptionResNetV2

The present investigation pertains to a network comprising 164 layers systematically arranged into deep and inception blocks featuring residual connections. The network [24] has been specifically designed to handle image inputs with dimensions of 299 × 299 × 9. The architectural design comprises a foremost stem module succeeded by a sequence of Inception ResNet-A, Reduction-A, Inception ResNet-B, Reduction-B, and Inception ResNet-C modules.

The stem component comprises a series of 3 × 3 × 3 convolutions, 3 × 3 × 3 max pooling, and filter concatenation. This collective operation yields the input to the inception ResNet-A block. The utilization of the ReLU activation function is observed in each inception block, whereas the reduction block integrates a pooling layer and convolutions of 1 × 1 × 1 and 3 × 3 × 3 dimensions.

Batch normalization is selectively employed solely on the uppermost conventional layers in implementing of the Inception-ResNet model, with exclusion from summations.

2.5.10 Xception

The Xception framework, introduced in 2016, represents a DL architecture that builds upon the Inception model, offering a unique approach to balance computational efficiency and model accuracy. At its core, Xception incorporates depth-wise separable convolutions as a key innovation. This technique is designed to reduce the computational burden on the network while preserving its precision. In the Xception architecture, traditional Inception modules are replaced with blocks of “depth-wise separable convolutions”. These blocks consist of two critical components: a depth-wise convolution layer, which applies individual convolution filters to each input channel, followed by a pointwise convolution layer that performs 1x1 convolutions on the output from the depth-wise convolution. This approach enhances the network’s ability to capture meaningful features while significantly reducing the number of parameters involved. One of the standout advantages of Xception is its ability to achieve state-of-the-art accuracy across a wide range of CV tasks while utilizing fewer parameters and computational resources compared to other DL architectures. This characteristic makes Xception particularly well-suited for implementation on resource-constrained devices, such as embedded systems and mobile phones. The Xception framework [21], proposed in 2016, is a DL architecture that functions as a variant of the Inception model. The utilization of depth-wise separable convolutions is implemented to mitigate the computational load of the network while simultaneously maintaining its precision. The Xception architecture replaces every Inception module with a block of “depth-wise separable convolutions”. The present block comprises a layer of depth-wise convolution, which executes a distinct convolution filter on each input channel, and subsequently, a layer of pointwise convolution, which applies a 1 × 1 convolution to the output of the depth-wise convolution. The utilization of this technique allows for the neural network to achieve more effective feature representations while minimizing the number of parameters. Xception possesses a noteworthy benefit in that it can attain cutting-edge precision in diverse CV assignments, all the while utilizing fewer parameters and computational resources in comparison to alternative DL architectures. This characteristic renders it highly suitable for implementing devices with limited resources, such as embedded systems and mobile phones. The Xception model has exhibited versatility and efficacy across various domains, such as image classification, object detection, and semantic segmentation, since its inception.

3 Experiments details and results

The significance of Data-Driven DDTL and its applications for analyzing cracks in subsurface borehole images for problems with concrete dams are emphasized in this paper. The sub-surface concrete structure is examined using a unique data set, which is obtained by CCTV equipment deployed via a borehole in the dam. Cracks were primarily targeted. The classification of normal and discontinued crack images along the borehole was done using DDTL.

3.1 Experimental setup

The suggested pre-trained models are trained with the Python programming language and the Keras framework. To conduct all the experiments for this paper, we use the most well-liked Google Colab laptop environment with the GPU runtime type environment. Google grants free access to Google Colab tools to researchers who can run experiments. We train and test the suggested model using an NVIDIA GeForce MX350 Tesla GPU with 8 MB RAM.

3.2 Performance evaluation matrix

In this study, we are evaluating the performance of 12 pre-trained models on the Pillow Dam Borehole CCTV image data set. In this study, we evaluate the performance of 12 pre-trained models on the Pillow Dam Borehole CCTV image data set. We evaluate their accuracy, precision [75], recall [75], F1 score [75], and support metrics, which are defined as follows based on the confusion matrix obtained during the classification phase [75,76].

(1)\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}=\frac{TP+TN}{TP+FP+FN+TN}\times 100,

(2)\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}=\frac{TP}{TP+FP}\times 100,

(3)\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}=\frac{TP}{TP+FN}\times 100,

(4)\mathrm{F}{1}_{\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}}=2\times \frac{Precision\times Recall}{Precision+Recall}\times 100,

where True Positive (TP) and True Negative (TN) values represent the number of normal (undamage) and discontinued cracks regions in the Pillow Dam Borehole CCTV image data set that are correctly classified as Normal rock (non-fracture) and discontinues in cracks, respectively. False positive (FP) and False Negative (FN) describe incorrectly classified instances. We use the terms “n(T)” and “n(C)” to represent the number of actual cases and classified cases, respectively.

3.3 Comparative analysis

We focus on analyses of normal (undamaged) borehole wall surface prediction and concrete crack surface prediction. Normal prediction seeks to detect non-defective or intact rock surfaces in dam images. A DL model must be trained to detect and differentiate between normal or undamaged regions and other anomalies, such as cracks or structural deformities. Engineers and analysts can evaluate the overall structural soundness of the dam and pinpoint any potential problem areas by precisely identifying normal surfaces. On the other hand, crack prediction focuses exclusively on locating and identifying cracks within dam images, and a DL model must be trained to recognize the presence, position, and size of cracks. Both tasks play vital roles in dam inspection and servicing, permitting early detection of possible weaknesses and assisting in ensuring the safety and integrity of the dam’s structure.

The confusion matrix is a valuable tool for assessing a crack detection model’s performance. It enables us to examine the accuracy, precision, recall, and other performance measures by providing a thorough summary of the model’s predictions and the actual ground truth labels. The confusion matrix and visual crack detection of crack prediction generated through DenseNet169, DenseNet121, DenseNet201, EfficientNetB0, EfficientNetB2, InceptionV3, InceptionRestV2, MobileNet, ResNet101, VGG16, and VGG19 state-of-the-art DL models are shown in Fig.5 and Fig.6. The models’ training average accuracy provides information on how well they perform on the training data, demonstrating their capacity to recognize patterns and features for crack detection.

The comparison results based on their precision, recall, F1 scores, and accuracy are shown in Fig.7.

We analyze the average training accuracy of the 12 DL models mentioned above in the assessment shown in Fig.8.

Understanding how well the models learn and classify cracks in the training data are possible to training accuracy. We analyze these 12 DL models designed explicitly for crack prediction on their training accuracy, as shown in Fig.9.

We are examining the training loss shown in Fig.10 and the average loss in Fig.11 of 12 DL models created exclusively for crack prediction assessment. Training loss quantifies the disparity between predicted and actual crack labels during the training process. The average loss, on the other hand, provides a comprehensive assessment of the model’s performance across multiple training epochs. This evaluation aims to assess both the training loss and the average loss, shedding light on their implications for crack prediction. The difference between the expected and accurate crack labels during training is measured as training loss. The average loss displays the total training effectiveness over several epochs. This evaluation attempts to measure the models’ training loss and average loss and to discuss how those results may affect crack prediction.

3.4 Discussion

In this research, we tackle the challenge of data set imbalance by harnessing advanced image augmentation strategies. Our approach involves the incorporation of Image Data augmentation techniques to enrich both the diversity and scale of our data set. This augmentation process includes the application of three distinct techniques applied to the existing data samples: rotation (with a range of 90°), flips (both horizontally and vertically), and zoom (both in and out). These augmentation techniques have been carefully selected to harmonize with the inherent characteristics of our data set and meet the precise requirements of our task. These augmentation methods are thoughtfully chosen to align with the intrinsic characteristics of our data set and the specific demands of our task. We recognize the key role played by the selection of appropriate data augmentation techniques in enhancing model performance. To validate the efficacy of our chosen augmentation configurations, we conduct a series of extensive evaluations, exploring various augmentation parameter values. Following rigorous experimentation, we approach the augmentation settings with the aim of fine-tuning them to enhance the overall performance of the model. we arrive at the augmentation settings poised to optimize the model’s overall performance.

The models considered are DenseNet169, DenseNet121, DenseNet201, EfficientNetB0, EfficientNetB2, InceptionV3, InceptionRestV2, MobileNet, ResNet101, VGG16, and VGG19. The DenseNet169 model achieves high precision and recall for the normal borehole surface of wall and crack images, with F1 scores above 0.80. This indicates that the model performs well in correctly identifying instances of both classes. The overall accuracy of 0.84 further emphasizes its ability to classify images correctly. DenseNet121 model demonstrates good precision, recall, and F1 scores for both the normal rock images and crack images. With an accuracy of 0.86, the model performs accurately in classifying images from both categories. DenseNet201 model shows a higher recall of 0.96 for undamaged surface images, indicating its effectiveness in capturing most instances of normal surface images. However, the precision for this class is relatively lower at 0.80. For the crack image class, it achieves a high precision of 0.94 but a lower recall of 0.75. The overall accuracy of the model is 0.86. Moving on to the EfficientNetB0 model, it demonstrates high precision, recall, and F1 scores for both the normal rock and crack images. With an accuracy of 0.91, this model exhibits reliable overall performance in classifying images from both categories. Similarly, the Efficient NetB2 model shows consistent precision, recall, and F1 scores for both the normal rock and crack images. It achieves an accuracy of 0.91, indicating reliable performance in accurately classifying images from both classes. The InceptionV3 model, however, achieves relatively low precision, recall, and F1 scores for the normal surface class than is achieved by the other models. For the crack class, it exhibits a high recall of 0.86 but a lower precision of 0.65. The model’s accuracy is 0.71, suggesting room for improvement in its classification performance, especially for normal surface images. Moving forward, the InceptionRestV2 model demonstrates a low precision of 0.51 for the normal surface class, indicating a high rate of false positives. However, it achieves a high recall of 0.99, meaning it can effectively identify most instances of normal surface. The F1 score of 0.67 suggests a reasonable balance between precision and recall. In terms of crack, precision, and recall are 0.00, indicating poor performance in correctly identifying instances of crack images. The model’s accuracy is 0.51, which aligns with its low precision and recall. Shifting to the MobileNet model, it demonstrates a moderate precision of 0.76 for the normal surface class, suggesting a relatively low false positive rate. The recall of 0.89 indicates its ability to capture a significant portion of the instances of normal surface. The F1 score of 0.82 suggests a reasonable balance between precision and recall. For the crack class, the precision of 0.86 indicates a relatively low false positive rate, while the recall of 0.70 suggests room for improvement in identifying instances of crack images. The model achieves an accuracy of 0.80, indicating overall reliable performance. The ResNet101 model exhibits a high precision of 0.93 for the normal surface class, indicating a low false positive rate. With a recall of 0.88, it captures a considerable portion of the instances of normal surface images. The F1 score of 0.91 suggests a good balance between precision and recall. For the crack class, the precision of 0.88 and recall of 0.93 indicate strong performance in correctly identifying instances of crack images. The model achieves an accuracy of 0.91, indicating reliable performance in classifying normal surface and crack images. Moving to the VGG16 model, it demonstrates a high precision of 0.87 for the normal rock class, indicating a low false positive rate. The recall of 0.92 suggests its ability to capture most instances of normal rock images. The F1 score of 0.89 indicates a balance between precision and recall. For the crack class, precision and recall are 0.92, indicating effective identification of instances of crack class. The model achieves an accuracy of 0.89, suggesting reliable overall performance. The VGG19 model exhibits a high precision of 0.89 for the normal surface class, indicating a low false positive rate. With a recall of 0.92, it captures a significant portion of the instances of normal rock images. The F1 score of 0.90 suggests a balance between precision and recall. For the crack class, the precision of 0.92 and recall of 0.88 indicate effective identification of instances of crack images. The model achieves an accuracy of 0.90, indicating reliable performance in classifying normal surface and crack images. Finally, based on the provided information, the Xception model achieves an accuracy of 0.67. It shows a precision of 0.61 for the normal rock images, indicating a relatively high false positive rate. However, it achieves a high recall of 0.97, capturing a significant portion of instances of normal rock image. The F1 score of 0.75 suggests a reasonable balance between precision and recall for this class.

Based on the provided information, the models that generally perform well in terms of precision, recall, F1 scores, accuracy, training accuracy, average training accuracy, training loss, and training average loss are DenseNet169, DenseNet121, EfficientNetB0, EfficientNetB2, ResNet101, VGG16, VGG19.

These models consistently achieve high precision, recall, and F1 scores for both classes, indicating their effectiveness in classifying normal surface and crack t images. Their accuracy values range from 0.84 to 0.91, demonstrating reliable overall performance.

The InceptionV3 and InceptionRestV2 models show limitations in precision and recall, particularly for the crack class. These findings provide valuable insights into the classification performance of the various models, enabling informed decision-making in selecting the most appropriate model for specific applications.

3.5 Future direction

Improved performance metrics. In addition to precision, recall, and F1score, other performance metrics can offer an improved evaluation of crack detection models, including specificity, area under the receiver operating characteristic curve , and area under the precision-recall curve. These measures provide more information about the model’s general discriminating capability, actual negative rate, and trade-off between precision and recall.

TL and pretrained models. Crack detection may benefit from TL that uses pre-trained models. Models can discover helpful features and patterns that can generalize well to crack detection tasks by using existing DL models trained on massive data sets, like ImageNet, as a starting point. These pre-trained models can be improved and trained more quickly using data sets relating specifically to cracks.

Data augmentation and balancing. By using data augmentation methods like rotation, scaling, and flipping, crack data sets can be more diverse and more extensive. Enhanced data can increase the model’s robustness and capacity to generalize to various fracture variations. Furthermore, resolving class imbalance through oversampling or under-sampling methods can reduce bias toward the dominant class and enhance model performance.

Hybrid approaches. For better crack identification, DL models can be combined with other methods, such as conventional CV algorithms or signal processing techniques. By gathering supplementary data from many modalities, hybrid techniques can make use of the advantages of each technique and improve overall performance.

Real-time crack detection. The proactive repair and monitoring of dam infrastructure can be made possible by developing real-time crack detection systems utilizing DL. Real-time deployment can be facilitated by accelerating models for hardware, researching lightweight architectures, and optimizing models for speed and efficiency.

4 Conclusions

DL models for crack prediction show promise for dam infrastructure maintenance and monitoring. They can attain high accuracy. Promising future options include enhancing performance measures, adding TL, data augmentation, and hybrid techniques, as well as providing real-time crack detection. We may contribute to a safer and more resilient infrastructure by tackling these issues, which will help DL models’ crack detection accuracy, efficiency, and practical application.

References

[1] Li Y, Bao T, Xu B, Shu X, Zhou Y, Du Y, Wang R, Zhang K. A deep residual neural network framework with transfer learning for concrete dams patch-level crack classification and weakly-supervised localization. Measurement, 2022, 188: 110641

[2] Mulligan M, van Soesbergen A, Sáenz L. GOODD, a global dataset of more than 38,000 georeferenced dams. Scientific Data, 2020, 7(1): 31

[3] EmbankmentsTechnical CommitteeDamsSlopes. Remote Sensing for Monitoring Embankments, Dams, and Slopes: Recent Advances. Reston, VA: American Society of Civil Engineers, 2021

[4] Li Y, Bao T, Shu X, Chen Z, Gao Z, Zhang K. A hybrid model integrating principal component analysis, fuzzy C-means, and Gaussian process regression for dam deformation prediction. Arabian Journal for Science and Engineering, 2021, 46(5): 4293–4306

[5] Feng D, Feng M Q. Computer vision for SHM of civil infrastructure: From dynamic response measurement to damage detection—A review. Engineering Structures, 2018, 156: 105–117

[6] Kim H, Ahn E, Shin M, Sim S H. Crack and noncrack classification from concrete surface images using machine learning. Structural Health Monitoring, 2019, 18(3): 725–738

[7] Chen B, Zhang H, Wang G, Huo J, Li Y, Li L. Automatic concrete infrastructure crack semantic segmentation using deep learning. Automation in Construction, 2023, 152: 104950

[8] Dong C Z, Catbas F N. A review of computer vision-based structural health monitoring at local and global levels. Structural Health Monitoring, 2021, 20(2): 692–743

[9] Meng Q, Xue H, Song H, Zhuang X, Rabczuk T. Rigid-block DEM modeling of mesoscale fracture behavior of concrete with random aggregates. Journal of Engineering Mechanics, 2023, 149(2): 04022114

[10] Ishfaque M, Dai Q, Haq N, Jadoon K, Shahzad S M, Janjuhah H T. Use of recurrent neural network with long short-term memory for seepage prediction at Tarbela Dam, KP, Pakistan. Energies, 2022, 15(9): 3123

[11] Mauludin L M, Budiman B A, Santosa S P, Zhuang X, Rabczuk T. Numerical modeling of microcrack behavior in encapsulation-based self-healing concrete under uniaxial tension. Journal of Mechanical Science and Technology, 2020, 34(5): 1847–1853

[12] Mauludin L M, Zhuang X, Rabczuk T. Computational modeling of fracture in encapsulation-based self-healing concrete using cohesive elements. Composite Structures, 2018, 196: 63–75

[13] Quayum M S, Zhuang X, Rabczuk T. Computational model generation and RVE design of self-healing concrete. Frontiers of Structural and Civil Engineering, 2015, 9(4): 383–396

[14] Zhu H, Wu X, Luo Y, Jia Y, Wang C, Fang Z, Zhuang X, Zhou S. Prediction of early compressive strength of ultrahigh-performance concrete using machine learning methods. International Journal of Computational Methods, 2023, 20(8): 2141023

[15] Chen J S, Lin K Y, Young S Y. Effects of crack width and permeability on moisture-induced damage of pavements. Journal of Materials in Civil Engineering, 2004, 16(3): 276–282

[16] Khan S U R, Zhao M, Asif S. Hybrid-NET: A fusion of DenseNet169 and advanced machine learning classifiers for enhanced brain tumor diagnosis. International Journal of Imaging Systems and Technology, 2024, 34(1): e22975

[17] Al-Masni M A, Kim D H, Kim T S. Multiple skin lesions diagnostics via integrated deep convolutional networks for segmentation and classification. Computer Methods and Programs in Biomedicine, 2020, 190: 105351

[18] HuangGLiuZvan der MaatenL. Densely connected convolutional networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Honolulu: IEEE Computer Society, 2017, 4700–4708

[19] HowardA GZhuMChenB. Mobilenets: Efficient convolutional neural networks for mobile vision applications. 2017, arXiv: 1704.04861

[20] HeKZhangXRenS. Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas, NV: IEEE Computer Society, 2016, 770–778

[21] CholletF. Deep learning with separable convolutions. 2016, arXiv: 1610.2357

[22] QuangN H. Automatic skin lesion analysis towards melanoma detection. In: Proceedings of 2017 21st Asia Pacific Symposium on Intelligent and Evolutionary Systems (IES). Hanoi: Springer, 2017, 106–111

[23] ShahinA HKamalAElattarM A. Deep ensemble learning for skin lesion classification from dermoscopic images. In: Prodeeings of 2018 9th Cairo International Biomedical Engineering Conference (CIBEC). Cairo: IEEE, 2018, 150–153

[24] SzegedyCIoffeSVanhouckeV. Inception-V4, inception-ResNet and the impact of residual connections on learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. San Francisco, CA: AAAI Press, 2017

[25] TanMLeQ. Efficientnet: Rethinking model scaling for convolutional neural networks. In: Proceedings of International Conference on Machine Learning. Beach, CA: PMLR, 2019, 6105–6114

[26] ChhabraMKumarR. A smart healthcare system based on classifier DenseNet 121 model to detect multiple diseases. In: Proceedings of Second MRCN 2021. Singapore: Springer Nature Singapore, 2022, 297–312

[27] Ashqar B A, Abu-Naser S S. Identifying images of invasive hydrangea using pre-trained deep convolutional neural networks. International Journal of Academic Engineering Research, 2019, 3(3): 28–36

[28] Wang M, Wang E, Liu X, Wang Z, Wang C. Influence of neural network structure on rock intelligent classification based on structural and tectonic features of rocks. Rock Mechanics and Rock Engineering, 2022, 55(9): 5415–5432

[29] Li Y, Wang P, Feng Q, Ji X, Jin D, Gong J. Landslide detection based on shipborne images and deep learning models: A case study in the three gorges reservoir area in China. Landslides, 2023, 20(3): 547–558

[30] TolstayaEEgorovA. Segmentation of seismic images. In: Processing of 13th International Conference in Computer and Graphic Visualization, Computer Vision and Image Processing. Hangzhou: IEEE, 2021, 564–570

[31] Jeon J, Lee J, Shin D, Park H. Development of dam safety management system. Advances in Engineering Software, 2009, 40(8): 554–563

[32] Zhang H, Yang G, Li H, Du W, Wang J. Pixel-wise detection algorithm for crack structural reconstruction based on rock CT images. Automation in Construction, 2023, 152: 104895

[33] Li Y, Bao T, Huang X, Chen H, Xu B, Shu X, Zhou Y, Cao Q, Tu J, Wang R, Zhang K. Underwater crack pixel-wise identification and quantification for dams via lightweight semantic segmentation and transfer learning. Automation in Construction, 2022, 144: 104600

[34] StrickerREisenbachMSesselmannM. Improving visual road condition assessment by extensive experiments on the extended gaps dataset. In: Proceedings of 2019 International Joint Conference on Neural Networks (IJCNN). Budapest: IEEE, 2019, 1–8

[35] Li Z. Global sensitivity analysis of the static performance of concrete gravity dam from the viewpoint of structural health monitoring. Archives of Computational Methods in Engineering, 2021, 28(3): 1611–1646

[36] Dong J, Wang N, Fang H, Hu Q, Zhang C, Ma B, Ma D, Hu H. Innovative method for pavement multiple damages segmentation and measurement by the Road-Seg-CapsNet of feature fusion. Construction and Building Materials, 2022, 324: 126719

[37] Kang D, Benipal S S, Gopal D L, Cha Y J. Hybrid pixel-level concrete crack segmentation and quantification across complex backgrounds using deep learning. Automation in Construction, 2020, 118: 103291

[38] Xiang Y, Sheng J B, Wang L, Cai Y B, Meng Y, Cai W. Research progresses on equipment technologies used in safety inspection, repair, and reinforcement for deepwater dams. Science China. Technological Sciences, 2022, 65(5): 1059–1071

[39] Kov’ari K, Peter G. Continuous strain monitoring in the rock foundation of a large gravity dam. Rock Mechanics and Rock Engineering, 1983, 16(3): 157–171

[40] Li Y, Bao T, Gong J, Shu X, Zhang K. The prediction of dam displacement time series using STL, extra-trees, and stacked LSTM neural network. IEEE Access: Practical Innovations, Open Solutions, 2020, 8: 94440–94452

[41] Wang W, Wang M, Li H, Zhao H, Wang K, He C, Wang J, Zheng S, Chen J. Pavement crack image acquisition methods and crack extraction algorithms: A review. Journal of Traffic and Transportation Engineering, 2019, 6(6): 535–556

[42] Li Y, Bao T, Gao Z, Shu X, Zhang K, Xie L, Zhang Z. A new dam structural response estimation paradigm powered by deep learning and transfer learning techniques. Structural Health Monitoring, 2022, 21(3): 770–787

[43] Li Y, Bao T, Shu X, Gao Z, Gong J, Zhang K. Data-driven crack behavior anomaly identification method for concrete dams in long-term service using offline and online change point detection. Journal of Civil Structural Health Monitoring, 2021, 11(5): 1449–1460

[44] Ishfaque M, Salman S, Jadoon K Z, Danish A A K, Bangash K U, Dai Q W. Understanding the effect of hydro-climatological parameters on dam seepage using shapley additive explanation (SHAP): A case study of earth-fill Tarbela dam, Pakistan. Water, 2022, 14(17): 2598

[45] Yamaguchi T, Hashimoto S. Fast crack detection method for large-size concrete surface images using percolation-based image processing. Machine Vision and Applications, 2010, 21(5): 797–809

[46] Chun P, Izumi S, Yamane T. Automatic detection method of cracks from concrete surface imagery using two-step light gradient boosting machine. Computer-Aided Civil and Infrastructure Engineering, 2021, 36(1): 61–72

[47] Hsieh Y A, Tsai Y J. Machine learning for crack detection: Review and model performance comparison. Journal of Computing in Civil Engineering, 2020, 34(5): 04020038

[48] Shirhatti V, Borthakur A, Ray S. Effect of reference scheme on power and phase of the local field potential. Neural Computation, 2016, 28(5): 882–913

[49] Pathirage C S N, Li J, Li L, Hao H, Liu W, Ni P. Structural damage identification based on autoencoder neural networks and deep learning. Engineering Structures, 2018, 172: 13–28

[50] Ni F, Zhang J, Chen Z. Pixel-level crack delineation in images with convolutional feature fusion. Structural Control and Health Monitoring, 2019, 26(1): e2286

[51] Kim B, Cho S. Image-based concrete crack assessment using mask and region-based convolutional neural network. Structural Control and Health Monitoring, 2019, 26(8): e2381

[52] Huyan J, Li W, Tighe S, Xu Z, Zhai J. CrackU-net: A novel deep convolutional neural network for pixelwise pavement crack detection. Structural Control and Health Monitoring, 2020, 27(8): e2551

[53] Zhao C, Ding D, Du Z, Shi Y, Su G, Yu S. Analysis of perception accuracy of roadside millimeter-wave radar for traffic risk assessment and early warning systems. International Journal of Environmental Research and Public Health, 2023, 20(1): 879

[54] Zhao C, Song A, Zhu Y, Jiang S, Liao F, Du Y. Data-driven indoor positioning correction for infrastructure-enabled autonomous driving systems: A lifelong framework. IEEE Transactions on Intelligent Transportation Systems, 2023, 24(4): 3908–3921

[55] Zhao C, Song A, Du Y, Yang B. TrajGAT: A map-embedded graph attention network for real-time vehicle trajectory imputation of roadside perception. Transportation Research Part C, Emerging Technologies, 2022, 142: 103787

[56] Cha Y J, Choi W, Suh G, Mahmoudkhani S, Büyüköztürk O. Autonomous structural visual inspection using region-based deep learning for detecting multiple damage types. Computer-Aided Civil and Infrastructure Engineering, 2018, 33(9): 731–747

[57] Huyan J, Li W, Tighe S, Zhai J, Xu Z, Chen Y. Detection of sealed and unsealed cracks with complex backgrounds using deep convolutional neural network. Automation in Construction, 2019, 107: 102946

[58] Du Y, Pan N, Xu Z, Deng F, Shen Y, Kang H. Pavement distress detection and classification based on YOLO network. International Journal of Pavement Engineering, 2021, 22(13): 1659–1672

[59] Liang J, Chen B, Shao C, Li J, Wu B. Time reverse modeling of damage detection in underwater concrete beams using piezoelectric intelligent modules. Sensors, 2020, 20(24): 7318

[60] Chen J, Xiong F, Zhu Y, Yan H. A crack detection method for underwater concrete structures using sensing-heating system with porous casing. Measurement, 2021, 168: 108332

[61] Si J, Xiong W, Zhong D, Yan A, Wang P, Liu Z. Piezoelectric-based damage-depth monitoring method for underwater energy-relief blasting technique. Journal of Civil Structural Health Monitoring, 2021, 11(2): 251–264

[62] KönigJJenkinsM DBarrieP. A convolutional neural network for pavement surface crack segmentation using residual connections and attention gating. In: Proceedings of 2019 IEEE International Conference on Image Processing (ICIP). Taipei, China: IEEE, 2019, 1460–1464

[63] Zhang J, Zhang J. An improved nondestructive semantic segmentation method for concrete dam surface crack images with high resolution. Mathematical Problems in Engineering, 2020, 2020: 1–14

[64] Dung C V, Sekiya H, Hirano S, Okatani T, Miki C. A vision-based method for crack detection in gusset plate welded joints of steel bridges using deep convolutional neural networks. Automation in Construction, 2019, 102: 217–229

[65] Sward D, Monk J, Barrett N. A systematic review of remotely operated vehicle surveys for visually assessing fish assemblages. Frontiers in Marine Science, 2019, 6: 134

[66] Capocci R, Dooly G, Omerdić E, Coleman J, Newe T, Toal D. Inspection-class remotely operated vehicles—A review. Journal of Marine Science and Engineering, 2017, 5(1): 13

[67] Lund-Hansen L C, Juul T, Eskildsen T D, Hawes I, Sorrell B, Melvad C, Hancke K. A low-cost remotely operated vehicle (ROV) with an optical positioning system for under-ice measurements and sampling. Cold Regions Science and Technology, 2018, 151: 148–155

[68] GuoHZhuangXRabczukT. A deep collocation method for the bending analysis of Kirchhoff plate. 2021, arXiv: 2102.02617

[69] Samaniego E, Anitescu C, Goswami S, Nguyen-Thanh V M, Guo H, Hamdia K, Zhuang X, Rabczuk T. An energy approach to the solution of partial differential equations in computational mechanics via machine learning: Concepts, implementation and applications. Computer Methods in Applied Mechanics and Engineering, 2020, 362: 112790

[70] Zhuang X, Guo H, Alajlan N, Zhu H, Rabczuk T. Deep autoencoder based energy method for the bending, vibration, and buckling analysis of Kirchhoff plates with transfer learning. European Journal of Mechanics-A/Solids, 2021, 87: 104225

[71] Guo H, Zhuang X, Chen P, Alajlan N, Rabczuk T. Stochastic deep collocation method based on neural architecture search and transfer learning for heterogeneous porous media. Engineering with Computers, 2022, 38(6): 5173–5198

[72] Guo H, Zhuang X, Fu X. Physics-informed deep learning for three-dimensional transient heat transfer analysis of functionally graded materials. Computational Mechanics, 2023, 72(3): 1–12

[73] Guo H, Zhuang X, Alajlan N, Rabczuk T. Physics-informed deep learning for melting heat transfer analysis with model-based transfer learning. Computers & Mathematics with Applications, 2023, 143: 303–317

[74] KwasigrochAMikołajczykAGrochowskiM. Deep neural networks approach to skin lesions classification—A comparative analysis. In: Proceedings of 2017 22nd International Conference on Methods and Models in Automation and Robotics (MMAR). Miedzyzdroje: IEEE, 2017, 1069–1074

[75] BonaccorsoG. Machine Learning Algorithms. London: Packt Publishing Ltd., 2017

[76] SaitoTRehmsmeierM. Basic evaluation measures from the confusion matrix. 2017. Available at the website of Classeval

RIGHTS & PERMISSIONS

Higher Education Press