Keywords

1 Introduction

With the vigorous development of the global economy, countries have invested in promoting infrastructure construction, with the construction of highway bridges becoming a significant trend. This development is not only a reflection of economic strength, but also an inevitable requirement for the modernization of the transportation system. While increasing support for transportation infrastructure, the country has also accelerated the pace of bridge construction to improve the density and efficiency of the transportation network [1].

However, with the rapid increase in the number of motor vehicles, road traffic is facing a series of serious problems. Overloading, overloading, and large traffic flow have become the main bottlenecks that constrain the efficiency of highway transportation [2]. This not only puts an excessive burden on roads, but also gradually exposes the limitations of design, construction, and maintenance levels for bridges that have been used for a long time. When facing these pressures, bridges may cause various diseases, thereby threatening the smooth operation of traffic.

In some countries, the impact of natural disasters further exacerbates the vulnerability of highway bridges. Natural disasters such as earthquakes, mudslides, and strong winds may cause catastrophic damage to bridges, placing a heavy burden on transportation infrastructure. Therefore, the maintenance and management of highway bridges has become crucial, and countries are gradually shifting the focus of maintenance to highway transportation infrastructure, focusing on strengthening specialization and informatization, and improving the technical system of data collection, detection and diagnosis, and maintenance and treatment.

Bridge damage not only brings direct social impacts, but also comes with huge economic cost losses [3]. The collapse events of the I-10 highway bridge in the United States in July 2015 and the Morandi bridge in Italy in August 2018 were typical examples, leading to a sharp decline in traffic efficiency, irreversible loss of life, and significant economic losses. Poor bridge maintenance may also lead to corrosion and other issues, and the corrosion problem of the Morandi Bridge ultimately led to the collapse of the bridge.

Bridge collapse not only directly limits the availability of infrastructure, but may also cause strong traffic interference and negative impacts on the surrounding road network, including indirect costs such as loss of life, user delays, and alternative route planning. Therefore, the detection and maintenance of highway bridges are particularly important.

Although natural disasters are difficult to avoid, by improving inspection standards, developing non-destructive testing technology, and applying intelligent testing equipment, bridges can be evaluated, tested, and maintained in a timely manner, effectively reducing the direct and potential indirect costs of repairing bridges [4]. In this context, this study focuses on the current status and problems of intelligent detection technology and equipment for highway bridges in the form of literature review, aiming to propose a series of ideas for optimizing the process of highway bridge detection.

Overall, the construction and management of highway bridges have become a global focus of attention. In the era of high connectivity, ensuring the safety and stable operation of highway bridges is not only related to national economic development, but also to the convenience of people's lives. Therefore, we need to delve deeper into and understand the construction and maintenance management of highway bridges in order to cope with more complex and ever-changing challenges in the future.

2 Research Status

In order to promote the intelligent development of bridge detection technology and combine the spatial layout characteristics of bridge structural components, the research and development of intelligent detection equipment for bridges has become a hot topic in this research field. Among them, in order to achieve the detection of bridge surface diseases, mechanical equipment equipped with image acquisition equipment has made significant progress in image acquisition and analysis in recent years.

Liu et al. conducted crack and surface damage detection on bridge piers based on traditional commercial drone equipped with image acquisition equipment (Fig. 1), and reconstructed three-dimensional images based on the detection results. The results showed that drone image acquisition can achieve free perspective switching, adapt to a wide range of bridge types, and have good detection flexibility [5,6,7,8]. However, this method usually requires manual control of drones, and the detection results are greatly affected by the drone's flight attitude and path.

Fig. 1.
figure 1

Application of drone technology in bridge surface disease detection.

Full size image

To overcome the problems of blurred image acquisition and difficult positioning caused by unstable drone platforms in drone image acquisition, Cuevas et al. developed an attachable drone monitoring device, which can achieve deformation monitoring and detection of high-altitude bridge beams on the basis of bridge surface disease inspection, as shown in Fig. 2. The detection device consists of four parts: ① unmanned aerial vehicle carrying platform; ② Attached device; ③ Deformation testing device; ④ Computer control system.

Fig. 2.
figure 2

Attachable drone monitoring device.

Full size image

Phung et al. proposed using point cloud scanning technology to plan the detection path of unmanned aerial vehicles (UAVs) in response to the difficulty of locating under the bridge and the possibility of collision between the UAVs and the bridge when using UAV detection, and conducted practical engineering verification.

Due to the difficulty in image positioning caused by the installation of image acquisition devices on drones, Diaz et al. developed a mobile device for synchronous image acquisition and ranging, and conducted experimental tests.

Xie et al. installed a multi-point 3D camera and laser irradiation system on the robotic arm of the inspection vehicle (Fig. 3) and designed an intelligent detection device based on 2D point images and line positioning. The device has good applicability in rapid detection of beam bridge diseases.

Fig. 3.
figure 3

Multi point 3D camera and laser irradiation intelligent detection robot

Full size image

In addition, Wang et al. developed an adsorption type steel bridge detection device based on the magnetic adsorption characteristics of steel in steel bridge structures [9]. The device consists of a magnetic shaft and sensors. In response to the electromagnetic characteristics changes caused by steel bridge defects, a strong magnetoresistance array is used to detect steel structure bridge damage based on the magnetic leakage characteristics of the steel bridge.

To achieve performance detection of bridge bearings, Peel et al. developed a detection tool that combines mobile robots and image acquisition technology. This device can achieve aging and cracking detection of bridge bearings.

3 Highway Bridge Disease Detection Technology Under the Background of Artificial Intelligence

3.1 Deep Learning

Deep learning is an important branch of machine learning and artificial intelligence. Its core idea is to train a large number of data samples layer by layer and construct multi-layer neural networks for sample detection. This method has shown enormous advantages and potential in fields such as computer vision, speech recognition, and natural language processing [10]. Especially in the application of semantic segmentation models, a deep learning branch, it can be successfully applied to the detection of highway bridge diseases. The publicly available dataset provides good data support for this detection task, providing a solid foundation for model training and performance improvement. Against the backdrop of the rapid development of artificial intelligence, deep learning technology has played an important role in constructing efficient semantic segmentation models through meticulous preprocessing and effective feature extraction of disease images [11, 12]. This model has significant application significance in highway bridge disease detection, and provides a powerful tool for improving the accuracy and efficiency of detection. This highlights the widespread applicability of deep learning technology in solving practical problems, providing people with more reliable and intelligent solutions.

3.2 PSPNet Semantic Segmentation Network

The semantic segmentation of images plays a crucial role in deep learning research, especially in scene understanding. In the field of driverless driving, achieving high-quality semantic segmentation of road scenes is crucial to ensure the safe driving of autonomous vehicle. This technology is not only limited to unmanned driving, but also plays an important role in areas such as highway bridge disease detection. By quickly and accurately segmenting bridges, the overall safety of the road can be effectively improved. In deep learning, especially in the improvement of semantic segmentation networks, PSPNet (Pyramid Scene Segmentation Network) is a noteworthy model. PSPNet is an improvement on the basis of fully convolutional network (FCN), introducing PSP module and pyramid pooling module to improve the ability to obtain global information. The key to the PSP structure is to divide the feature layer into grids of different sizes, and average pooling is performed within each grid to effectively aggregate contextual information from different regions. Specifically, in the PSPNet model, the input feature layers are cleverly divided into grids of different sizes (6 × 6. 3 × 3. 2 × 2. 1 × 1) Subsequently, perform an average pooling operation within each grid to obtain the final output result. The design of this structure enables the network to better capture semantic information at different scales in images, thereby improving the accuracy and effectiveness of semantic segmentation. Overall, the introduction of PSPNet has injected new vitality into the field of image semantic segmentation, providing powerful tools and methods for solving complex problems in practical scenarios. The steps are shown in Fig. 4.

Fig. 4.
figure 4

Basic structure of PSPNet

Full size image

4 Highway Bridge Disease Detection Based on PSPNet Model

4.1 Data Production

Personnel using artificial intelligence technology to replace highway bridge disease detection need sufficient image data support. In this process, the size of the semantic segmentation label should be consistent with the image size to ensure that each pixel value corresponds to a different target area. However, there are relatively few label sets for highway bridge disease images in the current public dataset, which requires us to actively collect existing public data and use Labelme tools for point labeling. The scope of point marking covers multiple aspects of highway bridge diseases, including exposed reinforcement, cracks, honeycomb and pitted surfaces, etc. In order to meet the requirements of semantic segmentation, this process is not only a simple labeling of the image, but also precise target classification for each pixel. In order to improve efficiency, we can also adopt methods such as cutting to reduce resource occupancy. At the same time, data augmentation through geometric transformations or color adjustments can help improve the robustness and generalization ability of the model. The organic integration of these steps will provide a more reliable data foundation for the application of artificial intelligence in the field of highway bridge disease detection.

4.2 Network Training

In the process of deep learning network training for highway bridge disease detection, we adopted the Keras framework, which allows us to flexibly adjust network parameters to optimize performance. In this process, key adjustments mainly include network classification settings and learning rate settings. Firstly, we divide the network into two categories to address the dual nature of highway bridge image data: one is diseased image data, and the other is disease-free image data. This grouping strategy helps the network better learn and distinguish between two situations, improving the accuracy of disease detection.

During the parameter adjustment process, we set the initial learning rate to 0.0001. This setting has been carefully considered and aims to make the model converge more stably in the early stages of training. By selecting an appropriate learning rate, we can accelerate the training process and improve the model's adaptability to data.

In order to complete the entire network training process, we conducted 45 iterations. The selection of this iteration number is based on balancing the training effect and computational cost, ensuring that the model can achieve sufficient convergence in a limited time and achieve satisfactory detection performance. Overall, by adjusting the parameters of the Keras deep learning framework reasonably, we have successfully established a deep learning model that can effectively detect highway bridge diseases.

4.3 Concrete Step

We first successfully created an effective dataset and conducted comprehensive training. Next, we used drones to collect video images of highway bridges in the field, and these images were transmitted to a computer for processing. On the computer side, we adopted the PSPNet semantic segmentation model for further training and calculation, in order to accurately identify and classify various diseases that may exist on highway bridges. The goal of the entire process is to quickly and accurately detect potential problems. Once our system detects any diseases on the bridge, we will immediately record these issues and take necessary measures to eliminate or repair them. This comprehensive method not only improves the efficiency of highway bridge inspection, but also ensures timely measures to maintain and strengthen the structural integrity of the bridge. By integrating advanced technology and data processing methods, we provide an efficient and reliable solution for bridge management and maintenance. The specific steps are shown in Fig. 5.

Fig. 5.
figure 5

Steps for Highway Bridge Disease Detection Based on PSPNet Model

Full size image

5 Experiment and Analysis

5.1 Dataset Collection

The core objective of this paper is to achieve intelligent detection methods by studying the use of PSPNet for highway bridge disease detection. To achieve this goal, a dataset containing highway bridge diseases was first constructed. These diseases mainly include various types such as cracks, exposed reinforcement, honeycomb and pitted surfaces. In order to obtain sufficient and diverse data, this article collected a total of 1000 images of highway bridge diseases through the internet and on-site photography. This dataset is divided into two parts, with 900 sheets used as the training set and the remaining 100 sheets forming the test set.

In order to generate meaningful results for subsequent model training and testing, this article used the labelme tool to annotate the images in detail. The process of annotation is complex and detailed, and the specific steps and annotation methods can be referred to Fig. 6 in the paper.

Fig. 6.
figure 6

Dataset production

Full size image

In summary, through this study, we not only established a large-scale dataset of highway bridge diseases, but also laid a solid foundation for the training and testing of subsequent PSPNet models. The collection and annotation of this dataset provides rich experimental materials for research, and provides strong support for the development of intelligent highway bridge disease detection methods.

5.2 Experimental Process and Analysis

This study aims to achieve effective crack detection by training on the highway bridge crack dataset using the PSPNet semantic segmentation platform. During the training process, we adopted a learning rate of 0.0001, batch_ Set the size to 4 and epoch to 50, and successfully generate the corresponding weight file.

Through the analysis of the results, we conclude that this method performs well in crack detection and successfully identifies cracks in highway bridges. This achievement not only proves the effectiveness of the PSPNet semantic segmentation platform in this task, but also emphasizes the rationality of the training settings used. Overall, this method not only accurately detects cracks, but also significantly improves the convenience and efficiency of detection, providing strong technical support for personnel in related fields.

6 Conclusion

The detection of highway bridge diseases plays a crucial role in management, but traditional manual detection has problems such as high cost and danger. In the current context of artificial intelligence, it is particularly important to study the detection of highway bridge diseases. This paper takes the PSPNet semantic segmentation platform as an example to verify the accuracy of the model in crack detection of highway bridges through experiments. With the increasing usage time of highway bridges, adopting deep learning semantic segmentation models for disease detection not only helps to reduce maintenance costs, improve detection efficiency, but also ensures the safety of operators. Therefore, this study provides an advanced, effective, and safe detection method for highway bridge management, and a feasible solution for long-term bridge maintenance.