Research on Intelligent Recognition and Characteristic Analysis of Water and Mud Bursting in Wangjiazhai Tunnel

Dang Shengshuo¹,², Ding Wenqi¹,², Zhang Qingzhao¹,², Wang Qiushi¹,², Tang Jixiang¹,²

¹ Department of Geotechnical Engineering, Tongji University, Shanghai 200092, China
² Key Laboratory of Geotechnical and Underground Engineering of Ministry of Education, Tongji University, Shanghai 200092, China

Abstract

Based on the water-mud bursting events in the Wangjiazhai semi-diagenetic tunnel in Yunnan Province, this study analyzes the video characteristics of the entire disaster process and selects a deep learning model suitable for identifying tunnel bursting features. Using the Mask R-CNN network, the classification and segmentation of water-mud bursting in the Wangjiazhai tunnel were achieved. The main research contents are as follows: (1) The causes of water-mud bursting disasters in Wangjiazhai Tunnel were analyzed, and treatment measures for bursting sections were summarized: comprehensive dewatering to reduce the confined water head at the tunnel face, in-tunnel curtain grouting, densely-packed pipe sheds, and small advance conduits to increase the bearing capacity of the arch, ensuring stability and construction safety during excavation. (2) Field video images from Wangjiazhai Tunnel were analyzed to obtain basic video characteristics of bursting: 1) suddenness of the disaster; 2) scale of the bursting; 3) high sediment content; 4) both sudden and predictive nature. A classification method for bursting severity levels was also proposed. (3) A typical image dataset of water-mud bursting in Wangjiazhai Tunnel was created, and the Mask R-CNN network was used to train and learn from the samples, achieving intelligent classification and segmentation of different types of bursting. The test set achieved a mAP of 90.0%, demonstrating that Mask R-CNN has stronger feature extraction and representation capabilities for Wangjiazhai tunnel bursting characteristics compared to Faster R-CNN, along with higher accuracy.

Keywords: mountain tunnel; deep learning; rating of water and mud bursting; video characteristics; semantic segmentation

Introduction

Since the beginning of the 21st century, with the continuous deepening of China's "Western Development" and "Belt and Road" strategies, investment in western region construction has increased significantly, leading to the construction of numerous highway and railway tunnels in western China. Mountain tunnels in western regions typically face complex geological conditions, long construction periods, high construction difficulty, and numerous uncertainties, often resulting in various geological hazards during construction. Common tunnel construction geological disasters include large deformation, ground subsidence, collapse, mud bursting, water inrush, and roof falling. Among these, water-mud bursting is a typical geological disaster that frequently occurs in numerous mountain tunnel projects [1~2].

Currently, bursting accidents are primarily discovered through manual detection. However, due to individual factors, limited space and vision in tunnels, and difficulties in sound wave transmission, disasters are often not detected in time, frequently causing casualties. When disasters occur, workers' lives are threatened, and their instinctive reaction is to flee the danger zone, making it difficult to report accidents promptly and effectively. Therefore, developing intelligent recognition and monitoring technology for bursting disasters is critically important.

In recent years, artificial intelligence technology has received widespread attention for tunnel bursting identification [3]. Zhang and Wang [4] combined neural network models to propose a coal mine water hazard prediction system, providing an effective means for identifying water inrush disasters in mines. Yang and Ma [5] selected six main factors including groundwater level and rock stratum dip angle as evaluation indicators, using the BP neural network method to assess and identify water bursting risks in karst tunnels. Bai [6] analyzed disaster-causing factors for different types of tunnel water bursting and developed an intelligent prediction system for karst tunnel water and mud bursting based on machine learning. Cao [7] proposed a low-resolution water bursting identification method based on superpixels and texture features, using texture features to identify water bursting and solving the identification problem under low-resolution conditions. However, current intelligent analysis for tunnel bursting disaster monitoring and identification primarily focuses on structured data, with limited research on unstructured data such as surveillance video images.

When bursting occurs, the time for sediment and water to flood the working face or tunnel depends on factors such as water pressure, inrush volume, and flow velocity, ranging from several seconds to several minutes or even lasting for hours. It is necessary to capture the engineering characteristics of bursting, identify it promptly, and then take corresponding treatment and drainage measures, evacuate workers to prevent accident escalation. Intelligent identification of bursting and timely information transmission are significant for managers to make correct decisions effectively, thereby minimizing personnel and property losses.

2.1 Engineering Overview

The Wangjiazhai Tunnel is located with its entrance in Linxiang District, Lincang City, and its exit in Shuangjiang Lahu, Wa, Blang, and Dai Autonomous County. The right tunnel has a total length of 8,040 meters, from chainage K21+440 to K29+480, with a maximum burial depth of approximately 1,022 meters. The left tunnel runs from chainage ZK21+460 to ZK29+470, with a total length of 8,010 meters and a maximum burial depth of approximately 1,002 meters.

The left tunnel section ZK23+025~ZK22+874 passes through a weakly weathered granite area, located in the alteration zone at the transition between Tertiary quartz sandstone and Tertiary quartz sandy conglomerate, representing a contact zone between hard rock and soft-hard rock, and a strongly to moderately water-rich alteration zone with complex and variable hydrogeological conditions. The remaining unexcavated section has experienced multiple large-scale mud and water bursting events, forming extensive "remolded soil" geological conditions with disturbed and damaged strata. Since construction began in March 2018, over twenty large-scale water-mud bursting geological disasters have occurred, with more than ten causing roof collapse. This has significantly impacted construction progress, with average daily advance less than 1 meter, posing a prominent constraint on the overall project schedule. The three adverse factors of "water-rich, high-pressure, and poor geology" have made construction difficulty and safety risks beyond imagination, with bursting prevention technology reaching unprecedented difficulty levels and attracting high attention from the engineering community.

2.2 Cause Analysis of Water-Mud Bursting in Wangjiazhai Tunnel

Based on statistical analysis of water-mud bursting cases in Wangjiazhai Tunnel, the causes are summarized as follows:

(1) Stratum lithology: The surrounding rock at the Linxiang end consists of Tertiary sandstone interbedded with claystone semi-diagenetic rock, characterized by poor cementation, extremely soft rock quality, large porosity, high saturation, cracking upon water loss, easy softening when encountering water, and prone to bursting (gushing) in a "paste" state when disturbed. Meanwhile, the tunnel face is susceptible to piping disasters, entraining sandy particles, which hollows out the stratum and subsequently develops into large-scale mud and water bursting disasters.

(2) Surface water and groundwater: Due to geological structure, the sandstone layer has high porosity, low density, and good water-bearing capacity. The stratum water is strongly water-rich with high water head, showing obvious distribution characteristics in the form of "water bags." The groundwater content in the sandstone layer is nearly saturated, and the sandstone and clay layers soften and lose strength, making geological disasters of mud and water bursting highly likely.

(3) Structural design: The Wangjiazhai Tunnel employs adverse slope construction, which largely causes water flowing back above the tunnel face and infiltrating water to continuously accumulate. As the water level gradually rises, water pressure will continuously increase, destroying the stability of the rock mass itself and continuously increasing the overlying soil pressure.

(4) Surface collapse and roof falling: The surrounding rock has extremely poor self-stability, making it difficult to form a caving arch effect. The concept of "fully utilizing surrounding rock self-stabilization capability" in the New Austrian Tunneling Method (NATM) is no longer applicable. Therefore, when water-mud bursting disasters occur, they often cause surface roof falling.

2.3 Analysis of Anti-Bursting Measures in Wangjiazhai Tunnel

Based on the analysis of bursting causes, anti-bursting measures were summarized to address each adverse factor:

(1) Comprehensive dewatering: To reduce water pressure above the tunnel face and control bursting risk, dewatering is needed to reduce the confined water head. However, single dewatering schemes are difficult to achieve expected results due to large water volume per unit area, small influence range, and long time required for dewatering before excavation. Therefore, a comprehensive dewatering scheme was adopted, focusing on surface deep wells, supplemented by in-tunnel well-point dewatering, with advance drilling drainage at the tunnel face and pipe-jacking advance pilot tunnel drainage, combining drainage and blocking to ensure construction safety and schedule requirements.

(2) In-tunnel curtain grouting: To reinforce and improve soil, block water seepage, and improve excavation conditions, while preventing hole collapse and drilling failure during pipe shed construction, advance curtain grouting was performed before pipe shed installation. Using a drilling-grouting integrated machine, two rings of advance grouting holes were drilled at 40 cm and 80 cm positions within the upper bench excavation contour line, with grouting pressure controlled at 3-4 MPa. Grouting materials were cement slurry and cement-sodium silicate double slurry. After achieving the expected reinforcement effect from in-tunnel curtain grouting, densely-packed pipe sheds could be installed, as shown in [FIGURE:2].

(3) Densely-packed pipe sheds: To improve arch soil and effectively form a reinforcement ring, after advance grouting, Φ89 pipe sheds were installed on the upper bench with a circumferential spacing of 25 cm and length of 9-15 meters. During pipe shed installation, supplementary grouting was performed based on drilling conditions to target weak grouting zones ahead of the tunnel face. The densely-packed pipe sheds formed a row of pre-control shed protection, increasing arch bearing capacity, ensuring stability during excavation and controlling initial support deformation, reducing the risk of sliding and bursting, while avoiding low efficiency, high cost, and uncontrollable risks from large-area curtain grouting reinforcement.

(4) Advance small conduits: To control over-excavation and ensure construction safety, by encrypting advance small conduits, adjusting their length, and excavating and supporting each frame, advance small conduits were installed for each steel frame. This effectively reduced disturbance and block falling from milling excavation, ensured excavation shaping, guaranteed construction safety, and reduced bursting risk.

(5) Rapid emergency measures: When small-scale local bursting occurs, to prevent deterioration, timely closure is required. Steel-reinforced stone cages are used to block water inrush promptly, while local precise grouting around the area compacts and densifies the surrounding rock in the bursting range.

3.1 Video Characteristic Analysis of Water-Mud Bursting in Wangjiazhai Tunnel

Unlike typical mountain karst tunnel water bursting, bursting in Wangjiazhai Tunnel often appears as dripping, showering, or flowing states during excavation, which may cause large-scale sudden bursting hazards when passing through local sections. Based on field investigation and monitoring of surveillance videos, the following characteristics of typical bursting in Wangjiazhai Tunnel were summarized:

(1) Suddenness of bursting: The sudden collapse of the tunnel face is one of the significant characteristics of Wangjiazhai Tunnel bursting. The suddenness of face collapse differs significantly from typical surrounding rock failure. The direct cause is that after excavation, when settlement control fails, the high water pressure and soil pressure accumulated in the mountain mass cause water-blocking measures such as horizontal jet grouting piles to instantly fail. Mud and water gush out, generating strong water wave vibrations that lead to frequent changes in surface ripple characteristics.

(2) Scale of disaster: Wangjiazhai Tunnel bursting exhibits certain scale characteristics. When disasters occur, large water flows rush into the tunnel interior, forming drastic changes in flow velocity and volume, with complex and violent changes in macroscopic morphological features such as covered area and action characteristics, which is one of the main video features of bursting events.

(3) High sediment content: When large-scale bursting occurs in Wangjiazhai Tunnel, it often entrains large amounts of sediment. Water flow with high sediment content typically appears more turbid, with color characteristics different from clear water, turning gray-brown or light brown.

(4) Both sudden and predictive nature: Once bursting deteriorates in Wangjiazhai Tunnel, its scale is often large and develops rapidly. However, it also shows certain predictability. Before deterioration, bursting often exhibits sharp increase in initial support settlement, cracking and deformation, sudden increase in water inflow, and increased sediment content at water outlets. By detecting color changes in videos, sediment content changes can be reflected, thereby judging the severity of bursting events.

(5) Often causes roof falling and surface collapse: Since construction began, among the 16 large-scale bursting geological disasters that occurred over 5 years, 10 caused roof falling and surface collapse. For example, on June 2, 2018, the Linxiang end left tunnel face ZK21+554 collapsed with sediment outflow, causing roof collapse. On June 15, two collapses occurred, forming a surface pit 21 meters long, 22.5 meters wide, and 11 meters deep, with approximately 2,887 m³ of mud burst.

In summary, compared with typical water bursting, Wangjiazhai Tunnel bursting has both similarities and obvious differences: (1) Both have characteristics of suddenness, scale, and mud-water mixing, showing video features such as frequent surface ripple changes and irregular complex macroscopic morphology, but Wangjiazhai Tunnel bursting emphasizes short-term sudden mud-water gushing. (2) Both have certain predictability, but before Wangjiazhai Tunnel bursting deteriorates, it often shows increased sediment content leading to different color video features, while typical water bursting focuses more on changes in water inflow volume and flow velocity.

3.2 Classification of Water-Mud Bursting Severity in Wangjiazhai Tunnel

GB12329-90 Karst Geology Terminology defines "karst water inrush" as: groundwater flow stored and moving in karst aquifers that suddenly produces large water inflow when artificially exposed or affected by natural factors, often accompanied by sand and mud gushing. By water volume change, it is divided into concentrated inrush and constant inrush; by inflow rate, it is divided into extra-large, large, and small inrush, as shown in [TABLE:1]; by gushing size, karst water inrush is divided into small gushing (<100 m³/d), medium gushing (100-1000 m³/d), large gushing (1000-10000 m³/d), and extra-large gushing (>10000 m³/d).

[TABLE:1]

Many problems remain in field response to bursting in Wangjiazhai Tunnel. When there is no obvious impact on construction, monitoring and treatment are usually not performed. Only when significant impacts occur on the project are mud and water inrush quantities monitored. The degree requiring high vigilance and remedial measures from field personnel remains unclear.

Based on practical engineering needs, we propose a bursting severity classification method that better reflects disaster characteristics of typical bursting such as in Yunnan's Wangjiazhai Tunnel. Based on field-collected video images, Wangjiazhai bursting is divided into three levels, as shown in [TABLE:2].

[TABLE:2]

Regarding sediment content, based on comparative analysis of different types of bursting from field monitoring videos, the characteristic features of each bursting level in Wangjiazhai Tunnel were summarized, as shown in [TABLE:3].

[TABLE:3]

How to identify and classify bursting disasters with low cost, high efficiency, and intelligence is an important current challenge. With deep learning development, convolutional neural network-based image recognition algorithms provide possibilities for solving this problem. Target detection algorithms can locate bursting occurrence but cannot calculate parameter information such as coverage area; instance segmentation algorithms can segment all bursting locations but cannot distinguish bursting categories. Therefore, this paper proposes a bursting classification and segmentation method based on the Mask R-CNN [8] model, which can classify different severity levels of bursting and complete segmentation of different levels.

4.1 Data Preparation

(1) Bursting image collection: The dataset in this study was obtained from videos captured by field cameras. To prevent high similarity between adjacent frames, images were extracted every 4 frames to form the dataset. High-quality images were manually selected, and typical bursting samples were classified by severity level.

(2) Data processing: Due to limited bursting videos, 150 samples were generated after selecting high-quality images. To enrich the dataset and further improve recognition efficiency, this study used the OpenCV open-source library to perform data augmentation for image samples, thereby improving model generalization and preventing overfitting. After processing, the bursting sample dataset used for training and testing contained 960 images, with 720 for training and 240 for testing. Each image was within 2 MB.

(3) Bursting dataset annotation: After processing bursting image data, LabelMe software was used to annotate bursting in dataset images. The labels in this study were mud1, mud2, and mud3, representing high-risk, medium-risk, and low-risk bursting, respectively, for network training. Partial annotated images are shown in [FIGURE:4].

4.2 Bursting Severity Recognition

[FIGURE:5] shows the model's recognition results for these three types of bursting. The figure demonstrates that Mask R-CNN performs well in bursting classification, successfully identifying the defined levels. However, the model's segmentation of bursting contours is not entirely accurate in some areas, possibly because when bursting contains high sediment content, target features are similar to tunnel environment features.

4.3 Recognition Accuracy Analysis

The Precision-Recall curve (PR curve) effectively addresses the limitation of single-threshold performance evaluation. By setting a series of different thresholds to provide different classification results, it demonstrates the overall performance of the classifier and allows selection of appropriate classification thresholds based on specific needs, making it more practical in real scenarios.

Average Precision (AP) represents the area under the Precision-Recall curve, calculated through integration. Generally, larger AP values indicate better detection performance for that category. Mean Average Precision (mAP) represents global model performance, being the mean of AP values across all categories.

This study used ResNet101 as the feature extraction network for model training and testing. [FIGURE:6] shows PR curves for Mask R-CNN and Faster R-CNN [9] recognition of different bursting types on the test set. The area under the PR curve represents recognition effectiveness for each category.

[TABLE:4] compares the two models under several evaluation metrics. The results show that Mask R-CNN achieved a test set mAP of 90.0%, demonstrating higher overall accuracy than Faster R-CNN and reflecting stronger feature extraction and representation capabilities for bursting characteristics. In terms of recognition effectiveness, it performs better on medium-risk and high-risk bursting, while low-risk bursting recognition is relatively weaker. Faster R-CNN also shows unsatisfactory performance on low-risk bursting.

4.4 Analysis of Bursting Development and Change

After model segmentation, visualization results are satisfactory. For bursting video images with good recognition results, this study used the OpenCV open-source library to extract specified colors from segmented bursting images, obtaining the bursting range in images. By finding contours and calculating their area, the coverage area of bursting portions was obtained. Typical bursting development video images were selected from samples. Since bursting occurs on the tunnel floor and assuming constant bursting channel depth, the coverage area can reflect bursting flow changes to some extent. Representative moments were selected: A, B, C, and D represent initial, developing, maximum scale, and receding moments, respectively, as shown in [FIGURE:7]. With time as the horizontal axis and the rate of area change at each moment relative to the initial moment as the vertical axis, the curve of area change rate over time was obtained, as shown in [FIGURE:8].

The curve shows that before bursting deterioration, area growth rate is relatively slow over time. At approximately 15 seconds in the video, bursting begins rapid development, reaching 60% of maximum scale within 5 seconds. Around 30 seconds, bursting scale reaches maximum, then gradually decreases, but coverage area remains larger than the initial moment. This curve demonstrates the disaster characteristics of Wangjiazhai Tunnel bursting, with less than 5 seconds from initial deterioration to relatively complete development, fully reflecting the suddenness of bursting and providing reference for subsequent flow measurement.

Conclusions

Based on the Wangjiazhai Tunnel project in Yunnan and combined with bursting cases occurring during construction, this study aimed to use artificial intelligence methods for bursting identification, yielding the following conclusions:

(1) By analyzing the causes of bursting disasters, treatment schemes for tunnel bursting sections were summarized: comprehensive dewatering to reduce confined water head at the tunnel face; in-tunnel curtain grouting, densely-packed pipe sheds, and small conduits to increase arch bearing capacity, ensuring stability and construction safety during excavation.

(2) Field video images were used to analyze bursting in Wangjiazhai Tunnel, summarizing video characteristics: 1) Disaster suddenness: frequent changes in surface ripple features; 2) Bursting scale: complex changes in macroscopic morphological features; 3) High sediment content: color features appearing gray-brown or light brown; 4) Both sudden and predictive nature. Based on practical engineering needs and combined with disaster characteristics, a severity classification method was proposed.

(3) A typical image dataset of water-mud bursting in Yunnan's Wangjiazhai Tunnel was created, and the Mask R-CNN network was used to train and learn from samples, achieving classification and segmentation of different bursting types. The test set achieved a mAP of 90.0%. Finally, based on segmented images, the development curve of bursting coverage area change rate over time was plotted, clearly reflecting the bursting characteristics of Wangjiazhai Tunnel.

References

[1] Fu Kailong, Zhou Yu, Wei Zhengxiong. Analysis of the 6.10 water and mud inrush event at the outlet adit of Chaoyang Tunnel on the Guiyang-Nanning High-speed Railway [J]. Carsologica Sinica, 2022, 41(06): 895-904.

[2] Wu Li, Wan Junwei, Chen Gang, et al. Analysis of the "8.5" water inrush accident in Yesanguan Tunnel of the Yichang-Wanzhou Railway [J]. Carsologica Sinica, 2009, 28(02): 212-218.

[3] Ou Guangzhao. Research on risk assessment of engineering disasters during tunnel construction based on machine learning [D]. Wuhan: China University of Geosciences, 2022.

[4] Zhang Wangjing, Li Renhou, Yan Ruiyi. An integrated expert system and neural network system for coal mine water hazard prediction [J]. Journal of Xi'an Jiaotong University, 1996, (08): 29-33+40.

[5] Yang Zhuo, Ma Chao. Risk prediction of water bursting in karst tunnels based on BP neural network method [J]. Tunnel Construction, 2016, 36(11): 1337-1342.

[6] Bai Chenghao. Research on intelligent prediction method for water and mud bursting disaster risk in karst tunnels based on machine learning [D]. Jinan: Shandong University, 2021.

[7] Cao Yuchao. Research on mine water disaster perception method based on visual computing [D]. Xuzhou: China University of Mining and Technology (Beijing), 2020.

[8] He K., Gkioxari G., Dollar P., et al. Mask R-CNN [C]// International Conference on Computer Vision. Computer Society, 2017.

[9] Ren S., He K., Girshick R., et al. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks [J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2017, 39(6): 1137-1149.