Subway tunnel disease intelligent detection method based on YOLO-T

Abstract

The invention discloses an intelligent subway tunnel disease detection method based on YOLO-T, and belongs to the field of intelligent nondestructive detection. The method comprises the following steps: acquiring tunnel disease image data through an area array industrial camera, and constructing a data set through sliding window screenshot, filtering and denoising, contrast stretching and binarization preprocessing; a YOLO-T model is improved, an SCConv module and an NAM attention mechanism are introduced, the structures of a backbone network, a neck network and a head network are optimized, the model parameter quantity is reduced, and the detection precision is improved; a model is trained by adopting a data division proportion of 8: 1: 1 and hyper-parameters; an intelligent disease recognition system is developed, the functions of danger level analysis, visual positioning and result management are integrated, image / video real-time detection and parameter dynamic adjustment are supported, and accurate disease positioning is achieved in combination with ArcGIS. The problems that an existing detection method is low in efficiency and insufficient in precision are solved, and the automation level and reliability of subway tunnel disease detection are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent nondestructive testing, specifically relating to intelligent detection methods and software for subway tunnel defects. Background Technology

[0002] During subway operation, various structural defects can occur due to a combination of internal and external factors, including issues with the tunnel and station's structural materials, equipment performance deterioration, construction quality, untimely tunnel maintenance, and the surrounding environment. Research indicates that defects can occur at any stage of a tunnel's lifecycle. Tunnel construction not only alleviates urban traffic congestion but also addresses the challenges of road construction by traversing rock masses. However, subways inevitably experience various defects during operation, such as cracks, water leakage, concrete spalling, segment misalignment, joint opening, uneven longitudinal settlement, and circumferential shrinkage deformation of the pipe diameter. These defects interact with each other; the combined effect of multiple defects can lead to the emergence of new defects, and the occurrence of one defect can exacerbate the harm caused by others. For example, cracks in the lining, voids behind the lining, and segment misalignment provide pathways for water leakage. Water seeps into the lining structure along the cracks, causing water leakage. Long-term water leakage accelerates the corrosion and rust of the tunnel concrete and steel reinforcement, thus affecting the strength of the tunnel lining. Among various subway tunnel defects, cracks are a common and significant concern, often being prioritized for inspection. Therefore, the rapid and efficient detection of surface defects in tunnels is a primary concern in the field of urban rail transit safety. Tunnel defect detection and maintenance are crucial steps in ensuring tunnel safety and reliability. Failure to conduct timely and regular defect detection and maintenance of subway tunnels during operation may reduce their service life and pose a significant threat to public safety and property.

[0003] Currently, the main methods for detecting defects in subway tunnels include manual inspection, digital photogrammetry, laser scanning, and sensor-based detection. While manual inspection is intuitive, it is labor-intensive, subjective, and easily affected by environmental factors, making it difficult to guarantee the accuracy and comprehensiveness of the results. Digital photogrammetry, although capable of acquiring richer image information, requires high-quality light sources and only provides partial information about the tunnel's inner surface, necessitating professional personnel and specialized image processing software for stitching and processing. Laser scanning utilizes three-dimensional laser scanning to obtain high-resolution spatial coordinates, texture, color, and echo reflection intensity, enabling the assessment of tunnel structural deformation states, such as convergence deformation. However, it suffers from high equipment costs, data processing software unsuitable for domestic applications, and speed limitations imposed by the laser profiler's rotation speed. Sensor-based detection technologies include the use of distributed fiber optic sensors, wireless tilt sensors, and temperature sensors for tunnel defect detection; however, most of these technologies are still in the research and initial application stages and have not yet formed a mature detection system.

[0004] With the development of artificial intelligence technology, intelligent detection of tunnel defects based on deep learning and computer vision has become a research hotspot and a future development trend, but many problems still exist. For example, conventional deep learning-based tunnel defect detection methods often use rectangular boxes to describe the target positions of the detector at each stage, which leads to a decrease in the quality of extracted features and affects the final detection performance. At the same time, the visual environment inside the tunnel is extremely complex, and there are many interfering factors when using computer vision technology to identify defects. For example, cracks are easily interfered with by scratches on the inner wall of the subway tunnel, resulting in a decrease in recognition quality.

[0005] In summary, existing methods for detecting defects in subway tunnels have many shortcomings, necessitating the development of more efficient, accurate, and intelligent detection technologies to improve the efficiency and accuracy of subway tunnel defect detection. Summary of the Invention

[0006] To address the aforementioned issues, this invention provides an intelligent detection method for subway tunnel defects, aiming to overcome the limitations of existing detection technologies and offer a high-precision, efficient detection solution tailored to domestic needs. By improving the defect identification algorithm based on deep learning and computer vision, the accuracy and stability of detection are enhanced. Furthermore, the detection process is optimized through intelligent defect identification software, achieving intelligent and automated operation and providing more precise technical support for the maintenance and management of subway tunnels.

[0007] To address the problems existing in the prior art, this invention proposes an intelligent detection method for subway tunnel defects based on YOLO-T, characterized by the following steps: Step 1: Data Acquisition: Multiple area array industrial cameras move longitudinally along the subway tunnel to capture images of cracks and water leakage defects within the tunnel; Step 2: Data Preprocessing and Dataset Construction: The acquired images are processed by sliding window cropping, spatial domain filtering for noise reduction, block contrast stretching, and binarization to construct a dataset containing three categories of data: cracks, water leakage, and no defects; Step 3: Model Building: An improved YOLO-T model is constructed, comprising a backbone network, a neck network, and a head network; the backbone network includes the SCConv module and the NAM attention mechanism; the neck network includes the CBLinear module and the CBFuse module; the head network includes the RepNCSPELAN module and the upsampling module; Step 4: Model Training and Result Prediction: The dataset is divided into training, validation, and test sets in an 8:1:1 ratio for model training, and the classification accuracy is evaluated using a confusion matrix. Step 5: Development of an intelligent disease identification system: Integrating the YOLO-T model, including a disease identification module, a hazard level analysis module, a visualization positioning module, and a result management module; the system supports image / video uploading, camera access, dynamic parameter adjustment, and precise disease location in conjunction with ArcGIS.

[0008] The specific parameters for the sliding window screenshot in step 2 are: window side length d, sliding step size s, obtaining Tx images in the width direction of the original image, and Ty images in the height direction, wherein the width w and height h of the original image satisfy... , The total number of images after the sliding window screenshot is .

[0009] The backbone network includes, in sequence, the first to fifth SCConv modules, the first to fourth RepNCSPELAN modules, and the first NAM module; the neck network includes, in sequence, the first to third CBLinear modules, the sixth to tenth SCConv modules, the fifth to eighth RepNCSPELAN modules, and the first to third CBFuse modules; the head network includes, in sequence, the ninth to thirteenth RepNCSPELAN modules, the first to fourth connection modules, and the eleventh to twelfth SCConv modules.

[0010] The SCConv module consists of a spatial reconstruction unit (SRU), a SiLu activation function, and a channel reconstruction unit (CRU) to compress spatial and channel redundancy. The NAM attention mechanism suppresses non-salient features through weight sparsity penalty. The neck network introduces programmable gradient information (PGI), including auxiliary invertible branches and multi-level auxiliary information, to solve the gradient information bottleneck problem.

[0011] The YOLO-T model is trained according to the multi-class cross-entropy loss function, and the multi-class cross-entropy loss function is specifically as follows: M—the number of categories, yic—the sign function (0 or 1, taking 1 if the true category of the sample is equal, otherwise taking 0), pic—the predicted probability that the observed sample belongs to the category.

[0012] The risk level analysis module divides the risk level based on the crack length, width, and leakage area: The crack risk level is divided into five levels according to the crack length L and crack width a, including slight (a < 1mm), medium (L < 5m and a < 3mm), relatively heavy (L < 5m and 3mm ≤ a ≤ 5mm), severe (5m ≤ L ≤ 10m and a > 5mm), extremely severe (L > 10m and a > 5mm); the leakage risk level is divided into four levels according to the area A, including slight (A < 0.5m²), general (0.5m² ≤ A ≤ 2m²), severe (2m² < A ≤ 5m²), critical (A > 5m²).

[0013] In step 5, the disease location calculates the displacement through an incremental encoder. The encoder is installed on the hub of the detection platform. The pulse signal combines with the encoder parameters to generate the absolute position coordinates in the tunnel, and aligns with the BIM model or GIS map to generate a 3D maintenance path planning diagram.

[0014] The disease intelligent recognition system further includes: API interface module: providing a standardized interface and documentation for third-party system integration; Interactive user manual: including software operation video tutorials and model training guides; Model weight real-time switching function: supporting dynamic loading of different weight files without restarting the system.

[0015] In the video processing module: Mode 1: Intercept key frames of the video at a fixed frame rate and convert them into an image sequence for batch detection; Mode 2: Use a super-resolution enhancement algorithm to improve the video quality and then analyze each frame. The enhancement algorithm is implemented based on a deep learning model.

[0016] The length calculation method of the disease adopts skeleton pixel point statistics or least squares method to fit the center line, the width calculation adopts multi-point measurement in the vertical direction or distance transformation method, and the leakage area is realized by counting the number of connected domain pixel points.

[0017] The present invention has the following advantages compared with traditional detection methods: 1. This invention proposes to add a self-calibrating convolution module and a NAM attention mechanism to the YOLO-T architecture, which effectively reduces the number of parameters and computational cost, and improves model accuracy.

[0018] 2. This invention utilizes a self-designed subway tunnel defect software to achieve real-time defect identification and hazard level assessment by recognizing the length, width, and area of ​​defects such as cracks and water leakage. Attached Figure Description

[0019] Figure 1 Flowchart of this invention.

[0020] Figure 2 The SSConv architecture diagram of this invention.

[0021] Figure 3 The BottleNeck network architecture diagram of this invention.

[0022] Figure 4 The RepNCSP network architecture diagram of this invention.

[0023] Figure 5 The RepNCSPELAN network architecture diagram of this invention.

[0024] Figure 6 The YOLO-T network architecture diagram of this invention.

[0025] Figure 7 This invention provides a preprocessed image of subway defect data.

[0026] Figure 8 The confusion matrix diagram of the classification model of this invention.

[0027] Figure 9 The training loss curve of the model of this invention.

[0028] Figure 10 The result diagram of the model of this invention.

[0029] Figure 11 The software flowchart of this invention. Detailed Implementation

[0030] A YOLO-T-based intelligent detection method for subway tunnel defects includes: Step 1: Data Collection Four 12-megapixel area array industrial cameras were moved along the longitudinal direction of the tunnel to collect images of the defects. The data collection was completed during the "maintenance window" when the subway tunnel was not in operation. The defect image data included information on cracks and water leakage. Step 2: Data preprocessing and dataset construction A high-quality dataset was obtained by performing a series of preprocessing steps on the collected disease image data, including sliding window screenshotting, spatial domain filtering for noise reduction, block contrast stretching, and binarization. The classification data mainly includes three categories: cracks, water seepage, and no disease. The width and height of the original image are represented by w and h, respectively. When performing sliding window screenshotting, a square window with a side length of d is used, and the step size of each slide is s. Through this sliding method, Tx images are obtained in the width direction of the original image, and Ty images are obtained in the height direction. Thus, each original tunnel image can yield a total of S images after sliding window screenshotting. The formula is as follows: Step 3: Model Building The structure of Yolo-T consists of three parts: the backbone network, the head network, and the neck network; The backbone network is responsible for extracting key features from the input image. These features provide the necessary information for subsequent network layers to perform object detection. The Silence module, located in the first layer of the YOLO-T network before the first SCConv module, facilitates auxiliary branches to access the original image input to the network. Its main function is as a pass-through layer, simply passing the input to the next network layer without performing any operations. The module itself does not perform any computations, and the output remains unchanged from the input. The backbone network sequentially includes the first SCConv module, the second SCConv module, the first RepNCSPELAN module, the third SCConv module, the second RepNCSPELAN module, the fourth SCConv module, the third RepNCSPELAN module, the fifth SCConv module, the fourth RepNCSPELAN module, and the first NAM module, wherein the first SCConv module serves as the input to the improved YOLO-T model. The neck network, located between the backbone and head networks, is used for feature fusion and enhancement to ensure that the head network receives richer and more effective feature information. It consists of auxiliary reversible branches and multi-level auxiliary information. The neck section sequentially includes a first CBLinear module, a second CBLinear module, a third CBLinear module, a sixth SCConv module, a seventh SCConv module, a fifth RepNCSPELAN module, an eighth SCConv module, a first CBFuse module, a sixth RepNCSPELAN module, a ninth SCConv module, a second CBFuse module, a seventh RepNCSPELAN module, a tenth SCConv module, a third CBFuse module, and an eighth RepNCSPELAN module; the output of the first CBLinear module serves as the input of the first CBFuse module; the output of the second CBLinear module serves as the input of the second CBFuse module; the output of the third CBLinear module serves as the input of the third CBFuse module; the output of the sixth RepNCSPELAN module generates a first detection layer (31) after a convolution operation; the output of the seventh RepNCSPELAN module generates a second detection layer (34) after a convolution operation; and the output of the eighth RepNCSPELAN module generates a third detection layer (37) after a convolution operation. Head network: Generates the final detection results based on the extracted features, including the target's location, category, and confidence level. The header section sequentially includes a ninth RepNCSPELAN module, a first upsampling module, a first connection module, a tenth RepNCSPELAN module, a second upsampling module, a second connection module, an eleventh RepNCSPELAN module, an eleventh SCConv module, a third connection module, a twelfth RepNCSPELAN module, a twelfth SCConv module, a fourth connection module, and a thirteenth RepNCSPELAN module. The output of the ninth RepNCSPELAN module serves as the input of the thirteenth RepNCSPELAN module. The output of the tenth RepNCSPELAN module serves as the input of the third connection module. The output of the second connection module serves as the input of the eleventh RepNCSPELAN module. The output of the eleventh RepNCSPELAN module is convolved to generate a fourth detection layer (16). The output of the twelfth RepNCSPELAN module is convolved to generate a fifth detection layer (19). The output of the thirteenth RepNCSPELAN module is convolved to generate a sixth detection layer (22). Auxiliary reversible branches and multi-level auxiliary information exist only in training mode to help the backbone network achieve better performance. The YOLO-T model is trained using a multivariate cross-entropy loss function. During training, six output feature maps are generated from the first, second, third, fourth, fifth, and sixth detection layers, serving as the outputs of the improved YOLO-T model. These six output feature maps are fed into the detection head to obtain the predicted labels. Furthermore, to improve model accuracy and reduce parameter computation, the YOLO-T model is improved by adding a self-calibrating convolutional (SC) module and a NAM attention mechanism, which improves model accuracy through the contribution of weights. The specific multivariate cross-entropy loss function is as follows: M—Number of categories, y ic —Sign function (0 or 1), 1 if the true class of the sample is equal, 0 otherwise, p ic —Predicted probability that the observed sample belongs to a category Step 4: Result Prediction The preprocessed data was divided into three parts: training set, validation set, and test set, with the collected image data divided into three parts in an 8:1:1 ratio. The improved YOLO-T model was used to train the classification model on the dataset, and the confusion matrix was used as a method to evaluate the prediction results. Step 5: Intelligent Disease Identification System The vision-based acquisition system can collect and store high-quality tunnel surface images. The processing and analysis of these images requires interactive software. It includes tunnel image display, visualization of the defect identification process, and result storage. The identification process combines image processing algorithms and deep learning. This hardware-software integration simplifies operations for subway inspection personnel and significantly reduces the burden of daily maintenance. The intelligent identification system can complete the entire process of tunnel defect identification. It displays the original tunnel images, ensuring detailed observation of key tunnel surface images. It can identify defect areas in tunnel images and select optimal parameters for different tunnel environments. It can calculate defect parameters, including length, width, and area, from the identified defect areas. It not only provides a simple and easy-to-use interface but also supports batch uploading and dataset management. The software allows users to upload images and videos of subway tunnel defects. It can also use industrial cameras for real-time tunnel monitoring. By integrating ArcGIS, it can pinpoint the location of identified defects, achieving precise location of surface defects in subway tunnels and facilitating maintenance by staff. Example

[0031] Step 1: Collect data The system integrates a data acquisition vehicle, an area array industrial camera, an incremental encoder, LED lighting, temperature and humidity sensors, and an expansion dock, with all components being detachable. The internal layout of the acquisition system is standardized, with neatly routed data and power cables. Different modules within the inspection vehicle and image acquisition system were designed and selected separately to meet the specific needs of subway tunnel inspection. The machine vision system first needs to calculate the camera's field of view (FOV), i.e., the range of the tunnel lining cross-section to be acquired. For a shield tunnel with a diameter D = 5400 mm, the acquisition range L of the tunnel cross-section is: The acquisition system needs to capture a 260° area of ​​the entire tunnel, which is calculated to be 12252.2 mm. To avoid omissions and ensure subsequent stitching of full-section images, a 20% overlap area is set between adjacent camera acquisition areas. It is initially considered to use four area-array cameras to complete the acquisition task; whether this will be sufficient will be determined based on further calculations. Therefore, the total area to be acquired is x: Calculations show that the field of view (FOV) of each area array industrial camera is x = 3603.6 mm, and the required acquisition accuracy is R = 1 mm / pixel. Therefore, the resolution of each camera is: A 4000-pixel area scan camera is sufficient to meet the above data acquisition requirements. An industrial area scan camera developed by MedVision Technologies was selected. Figure 2 The resolution is 4112×3088, which meets the circumferential acquisition accuracy requirements for tunnel lining. To ensure no frame drops in the lining images during acquisition, the selected sensor has a maximum frame rate of 30fps. The maximum speed of vscan for rapid inspection is: Calculations show that the maximum speed for rapid inspection is 92 m / s. This plan proposes to use four industrial cameras, which can simultaneously capture video and automatically take photos at set intervals. The four area-array industrial cameras will photograph defects in the tunnel's arch, arch waist, and sidewalls. The four cameras will also record video for subsequent defect location and identification. To balance inspection efficiency and defect detection accuracy, the inspection platform will use the MindVision-MV-SUC1206GC area-array camera model.

[0032] The lens focal length is calculated by determining the object distance for lens selection. Considering the camera module is fixed on a rectangular disk with a radius r = 500mm at the center of the tunnel, the camera object distance WD = D / 2 - r = 2200mm. Given that the pixel size of the IMX226 is m = 1.85μm, the camera focal length can be calculated as follows: Therefore, a 12mm lens was chosen. At a distance of 2.8m from the tunnel, the horizontal distance of the 12mm lens is 3.36m. The industrial camera can be set to take pictures at intervals, and based on a calculation of 3 photos per second, the maximum speed of the inspection vehicle capable of stitching together a complete tunnel image is 36.29km / h. In shield tunnel image acquisition, certain blind spots exist due to the illumination angle of the tunnel lighting and obstruction from some auxiliary facilities. In these blind spots, the CMOS sensor cannot obtain sufficient light, resulting in the loss of image details and insufficient contrast between defects and the lining, which is detrimental to subsequent image preprocessing and defect identification. Therefore, supplementary lighting equipment is needed to ensure the stability of image acquisition and the efficiency of defect identification.

[0033] During actual data acquisition, since there are no GNSS signals inside the tunnel, an encoder is required to position the detection platform. Encoders are divided into absolute encoders and incremental encoders. Incremental encoders have a simpler structure, stronger anti-interference capability, and a service life of over 50,000 hours. Considering that absolute position coordinates are not required during actual data acquisition, an incremental encoder was chosen and installed on the detection platform's hub. The encoder generates pulses as the hub rotates. These pulse signals are input to a pulse counter, and the displacement can be calculated by combining the number of signals counted by the counter with the encoder parameters.

[0034] Step 2: Data Processing The actual environment of subway tunnels is dimly lit, and the lining texture is complex, containing a large amount of water seepage and noise from unidentified obstacles, making direct classification and identification difficult. Therefore, the overall theoretical model includes image preprocessing algorithms such as gray-level erosion, block contrast stretching, and binarization to improve image contrast and image quality; and a hierarchical filtering preprocessing algorithm based on the zero-order moment of connected regions, rectangularity, and shape extrema to further filter out scattered noise, speckle noise, and extended large-area block noise, while preserving and enhancing crack features.

[0035] Two typical linear and nonlinear filtering methods in spatial domain filtering—mean filtering and median filtering—are used for noise reduction of tunnel crack images. Mean filtering is simple and effective, replacing the center pixel value with the average value of its neighborhood. It has a certain denoising effect on Gaussian noise and salt-and-pepper noise. Median filtering denoises by replacing the pixel value with the median value of its neighborhood. It is particularly effective against salt-and-pepper noise because it does not change the relative order of the pixels.

[0036] Block contrast stretching first divides the image into multiple equal-sized blocks (or blocks of varying sizes as needed). For each image block, the minimum and maximum brightness values ​​are calculated. For all pixels within each block, a linear transformation is used to map the minimum brightness value to a predetermined minimum output value, and the maximum brightness value to a predetermined maximum output value. The formula is as follows: Where s' is the stretched pixel value, s is the original pixel value, lmin is the minimum brightness value within the block, a is the slope, and b is the intercept.

[0037] Binarization is a technique in image processing that converts an image into one containing only two pixel values, typically black and white. This conversion process divides all pixels into two categories: one category is assigned a value (usually white, represented as 255), and the other category is assigned a different value (usually black, represented as 0). The binarization process requires determining a threshold value; all pixels above this threshold are set to white, and those below are set to black. Binarization is a fast and efficient image preprocessing method that can significantly reduce the amount of image data and facilitate further image analysis and processing. The binarization operation is illustrated by the following formula: Where I(x,y) is the gray value at point (x,y) in the binary image, and T is the threshold for the binarization operation.

[0038] In creating the dataset, tunnel lining images with cracks, water leakage, and spalling were specifically selected. Professionals manually screened the collected tunnel lining images to identify those with water leakage, cracks, spalling, and those without defects. The selected images were cropped to highlight the target cracks and water leakage. Large image sizes would significantly increase the computational load on the network, hindering training. The image resolution was set to 640×640 pixels. The final dataset consisted of 4115 images for tunnel lining defect classification, including 1375 images of water leakage, 1360 images of cracks, and 1380 images of no defects. 80% of the total image data for each class was randomly selected as the training dataset, 10% as the validation dataset, and the remaining 10% as the test set. Therefore, 3292 images were used as the training set, and the remaining 823 images were used as the validation test set. It can be seen that the proportions of images without defects, images with water leakage, and images with cracks are similar, the sample distribution is relatively uniform, and the test set and training set meet the basic requirements for training deep learning networks.

[0039] Step 3: Overall Network Structure and Improvements The Yolo-T architecture consists of a backbone network, a head network, and a neck network. The backbone network is the foundation of the model, responsible for extracting features from the input image. These features form the basis for subsequent network layers to perform object detection. The head network is the decision-making part of the object detection model, responsible for generating the final detection results. The neck network, located between the backbone and head networks, performs feature fusion and enhancement. Compared to other YOLO models, Yolo-T adjusts the anchor box size and aspect ratio, the number of network layers and convolutional kernels, the learning rate and optimizer parameters, and data augmentation strategies to reduce computation, improve detection accuracy, and enhance generalization ability. Yolo-T introduces the concept of Programmable Gradient Information (PGI) to address the various variations required for deep networks to achieve multiple objectives. PGI can provide complete input information for the target task to compute the objective function, thereby obtaining reliable gradient information to update network weights. PGI consists of auxiliary invertible branches and multi-level auxiliary information. Auxiliary invertible branches address the problems arising from deepening neural networks. Deeper networks can create information bottlenecks, preventing the loss function from generating reliable gradients. Multi-level auxiliary information aims to address the error accumulation problem caused by deep supervision, particularly in architectures with multiple prediction branches and lightweight models. Furthermore, a novel lightweight network architecture based on gradient path planning—Generalized Efficient Layer Aggregation Network (GELAN)—was designed. GELAN combines two neural network architectures designed using gradient path planning: CSPNet and ELAN. Specifically, it uses the RepNCSPELAN module within the structure. Compared to state-of-the-art methods based on deep convolution, GELAN achieves better parameter utilization using only traditional convolution operators. PGI can be used for a wide range of models, from lightweight to large. It can capture complete information, allowing models trained from scratch to achieve better results than existing models pre-trained on large datasets. The SSCconv module replaces the traditional convolution module. SSCconv is an efficient convolution module used to compress redundant features in convolutional neural networks, reducing computational load and improving model performance. It consists of three units: a Spatial Reconstruction Unit (SRU), a SiLu activation function, and a Channel Reconstruction Unit (CRU). The SRU is used to suppress spatial redundancy, and the CRU is used to reduce channel redundancy. Experimental results show that embedding the SCConv model can achieve better performance by reducing redundant features, and significantly reduces complexity and computational cost. A normalized attention module (NAM) is added to the backbone network, which suppresses less significant weights. It imposes a weight sparsity penalty on the attention module, thus making it more computationally efficient while preserving similar performance.

[0040] The backbone component sequentially includes a first SCConv module, a second SCConv module, a first RepNCSPELAN module, a third SCConv module, a second RepNCSPELAN module, a fourth SCConv module, a third RepNCSPELAN module, a fifth SCConv module, a fourth RepNCSPELAN module, and a first NAM module, wherein the SCConv module serves as the input terminal of the improved YOLO-T model; The neck section sequentially includes a first CBLinear module, a second CBLinear module, a third CBLinear module, a sixth SCConv module, a seventh SCConv module, a fifth RepNCSPELAN module, an eighth SCConv module, a first CBFuse module, a sixth RepNCSPELAN module, a ninth SCConv module, a second CBFuse module, a seventh RepNCSPELAN module, a tenth SCConv module, a third CBFuse module, and an eighth RepNCSPELAN module. The output of the first CBLinear module serves as the input of the first CBFuse module; the output of the second CBLinear module serves as the input of the second CBFuse module; the output of the third CBLinear module serves as the input of the third CBFuse module; the output of the sixth RepNCSPELAN module generates a first detection layer after a convolution operation; the output of the seventh RepNCSPELAN module generates a first detection layer after a convolution operation; the output of the eighth RepNCSPELAN module generates a first detection layer after a convolution operation. The header section sequentially includes a ninth RepNCSPELAN module, a first upsampling module, a first connection module, a tenth RepNCSPELAN module, a second upsampling module, a second connection module, an eleventh RepNCSPELAN module, an eleventh SCConv module, a third connection module, a twelfth RepNCSPELAN module, a twelfth SCConv module, a fourth connection module, and a thirteenth RepNCSPELAN module. The output of the ninth RepNCSPELAN module serves as the input of the thirteenth RepNCSPELAN module; the output of the tenth RepNCSPELAN module serves as the input of the third connection module; the output of the second connection module serves as the input of the eleventh RepNCSPELAN module; the output of the eleventh RepNCSPELAN module is convolved to generate a fourth detection layer; the output of the twelfth RepNCSPELAN module is convolved to generate a fifth detection layer; and the output of the thirteenth RepNCSPELAN module is convolved to generate a sixth detection layer. Auxiliary reversible branches and multi-level auxiliary information exist only in training mode and are used to help the backbone network achieve better performance. Multi-level auxiliary information is... Figure 6 In the network architecture diagram, branches #23-25 ​​and auxiliary invertible branches #26-37 together constitute the neck network. During the training phase, there are a total of 6 output feature maps, namely the first detection layer, the second detection layer, the third detection layer, the fourth detection layer, the fifth detection layer, and the sixth detection layer, which serve as the outputs of the improved YOLO-T model. These 6 output feature maps are fed into the detection head to obtain the predicted label.

[0041] Step 4: Train the model and predict the results. The results of the preprocessed crack, water leakage, and defect-free classification database were imported into the improved YOLo-T image classification model for training.

[0042] The selection of hyperparameters plays a crucial role in determining the training efficiency of deep learning models. Hyperparameters are not settings learned from the data, but rather set before the training process begins. This study primarily investigates the impact of three critical hyperparameters: learning rate, batch size, and epoch (one complete traversal of the training dataset), and selects their optimal values. In deep learning models, the learning rate determines the step size taken during gradient descent, which affects the speed at which the model updates weights. The batch size determines the number of samples used to update parameters in each iteration and significantly affects the model's generalization performance. As the number of epochs increases, the number of weight update iterations also increases, and the curve progresses from an unfit state to an optimal fit state, eventually leading to overfitting. After repeated tuning, the model's hyperparameters were determined to be learning rate = 0.0001, batch size = 32, and epoch = 100.

[0043] The loss curves from the training and validation sets show that the model's loss plateaus after 150 epochs and reaches its minimum at 200 epochs, indicating the highest accuracy. The improved model performs well in classifying cracks, leaks, and safety, achieving accuracies of 0.99, 0.94, and 0.99, respectively. The confusion matrix is ​​a tool used in machine learning to evaluate the performance of classification models. In the graph, each row corresponds to the true class, and each column corresponds to the predicted class. Each cell in the table contains two values: the horizontal axis represents the predicted class, and the vertical axis represents the true class. The value in each cell indicates the number of samples correctly or incorrectly classified within the predicted class. The confusion matrix of the model shows that out of 898 crack images, 10 were incorrectly predicted. Specifically, 7 cracks were incorrectly predicted as leaks, 3 cracks were incorrectly predicted as safe, 32 leaks were incorrectly classified as cracks, 3 leaks were incorrectly predicted as safe, and 2 safe samples were incorrectly classified as leaks. The confusion matrix can help us understand the performance of the classification model and fully demonstrate the accuracy of the model's predictions.

[0044] Step 5: Development of an intelligent disease identification system This software integrates a YOLO-T model for intelligent hazard classification of tunnel cracks and water seepage. Model training and optimization ensure accuracy and generalization. A user interface displays the model's hazard classification results for uploaded images. Visual analysis of the model's predictions is presented using charts and graphs to help users understand and interpret the model's working principles more intuitively. A simple and intuitive user interface allows both professionals and non-professionals to easily operate the software. Detailed user manuals and training materials are provided to help users better understand the software's functions and operating procedures. Software performance is optimized to ensure efficient operation when processing large-scale datasets.

[0045] The aforementioned intelligent detection and classification system for tunnel defects based on the YOLO-T model includes the following modules: Model training and optimization module: By integrating the YOLO-T model, target detection of tunnel cracks and water leakage is performed, and the model parameters are optimized through training to improve detection accuracy and generalization.

[0046] The user interaction module provides a visual interface that supports image / video uploads, camera access, model parameter settings, and preprocessing operations, and displays a real-time comparison between the original image and the model prediction results. This module uses a three-column layout: the left hidden column contains file loading buttons (image / video / camera), preprocessing function options, and dataset upload path configuration; the middle display is divided into upper and lower sections, with the upper section showing detection result statistics (category, quantity, FPS, model version), and the lower section displaying the original image and the predicted result image side-by-side; the right hidden column provides the model parameter setting interface, including model weight selection, IOU threshold, confidence threshold, latency adjustment, and result saving options.

[0047] Hazard Level Analysis Module: Based on the disease segmentation results, extract skeleton features, calculate crack length / width and leakage area, and output hazard level according to preset classification standards. Specific steps are as follows: Perform semantic segmentation on the classified disease image to extract the crack skeleton and leakage area contour; calculate the maximum crack width, extension length, and leakage area; classify into three hazard levels according to preset thresholds: Level I (Emergency), Level II (High Risk), and Level III (Warning).

[0048] Visual Positioning Module: This module integrates with the ArcGIS platform to pinpoint the location of defects on the tunnel's operational route map, achieving precise positioning. Specifically, it aligns the coordinate information of the inspection results with the coordinate system of the tunnel's BIM model or GIS map; marks the location and severity of defects on the tunnel route map; and generates an interactive 3D maintenance path planning map.

[0049] Video processing module: Provides two modes: Mode 1: Extracts keyframes from the video at a fixed frame rate, converts them into image sequences for batch detection; Mode 2: Improves video quality using a deep learning-based super-resolution enhancement algorithm, then analyzes the video frame by frame.

[0050] Parameter setting interface: Supports dynamic slider adjustment of IOU threshold (range 0.3-0.9) and confidence threshold (range 0.5-0.99); real-time switching and loading of model weight files without system restart; delay time setting (0-500ms) to adapt to different hardware performance.

[0051] Results Management Module: Automatically saves test result images, label files, and hazard level reports to a specified database; supports historical data comparison and analysis, and filters for differences in results between different test batches via a timeline.

[0052] Data processing flow: Supports fixed frame capture or super-resolution enhancement algorithm preprocessing when video input is selected; automatic contrast stretching, binarization and filtering denoising operations are performed during the image preprocessing stage; label files are generated synchronously when the detection results are output, and the threshold can be dynamically adjusted through the progress bar and the prediction effect can be refreshed in real time.

[0053] Supporting materials: An interactive user manual, including video tutorials on software operation and a model training guide; API interface documentation, supporting third-party system integration of disease detection functions.

[0054] The entire identification process involves opening the tunnel defect intelligent identification software, first selecting the model to use, and then choosing the previously trained YOLO-T defect classification and identification model. The trained YOLO-T model weights are then invoked, and the corresponding "Upload Dataset" button is clicked. After clicking "Load File," the main interface will display the input image, video path, and content. The appropriate tunnel defect dataset is selected based on the requirements. Once the selection is complete, the dataset upload step is finished. There are two options for uploading videos: one sets a fixed frame rate to extract images from the video for detection, and the other uses a super-resolution enhancement algorithm to improve the video quality and obtain video data. For images, the "Preprocess" button is clicked, and the software performs a series of image preprocessing operations on the uploaded dataset, including contrast stretching, binarization, and filtering / denoising. Alternatively, the "Select Camera" button can be clicked to activate an industrial camera and begin identification. The original image is displayed on the left side of the window, while the predicted classification image is displayed on the right side, allowing for easy comparison of the identification results and viewing of the defect type and quantity. The progress bar allows you to select the IOU and confidence thresholds. Setting different thresholds allows the model to produce different detection results. You can click the save button in the lower right corner to save the identified images and label files for later comparison and analysis. Select the images with completed defect classification for further hazard level analysis. After clicking hazard level analysis, the defects will be automatically segmented. Based on the segmented images, the skeleton of the defect will be extracted, and the length, width, and leakage area of ​​the cracks will be calculated. Hazard level analysis will be performed on the defects based on the hazard level classification criteria set for length, width, and area. Finally, using ArcGIS, the locations of the detected defects will be marked on the tunnel's operating line, achieving precise defect location and facilitating maintenance by staff.

[0055] The risk level of cracks is divided into five grades: Grade 1 (Minor): a < 1 mm (general cracking or no development); Grade 2 (Moderate): Length L < 5 m and width a < 3 mm; Grade 3 (Slightly Severe): L < 5 m and 5 mm ≥ a ≥ 3 mm; Grade 4 (Severe): L is 5~10 m, but a > 5 mm; Grade 5 (Extremely Severe): Length L > 10 m and width a > 5 mm. The standard for classifying the risk level of water leakage is as follows: Grade 1 (Minor): Area A < 0.5 m² 2 The leakage is small, the location is not critical, and the impact on structural safety is minimal. Level II, 0.5m. 2 ≤A≤2m 2(General): Moderate leakage, relatively important location, some impact on structural safety. Level 3 (Severe): 2m 2 <A≤5m 2 The leakage is large, the location is critical, and it has a significant impact on structural safety, potentially requiring emergency measures. Level IV (Critical) A > 5m 2 The leakage is extremely large and the location is critical, posing a serious threat to structural safety. Immediate measures are needed to prevent an accident.

[0056] The calculation process for crack length, width, and leakage area is as follows: First, the input image is preprocessed, including grayscale conversion, filtering, and edge detection. These steps enhance the features of crack leakage, reduce noise interference, and provide better input data for subsequent deep learning models. Crack leakage contour skeleton extraction: Extract the crack contour from the crack image. This can be achieved through morphological operations (dilation, erosion) and edge detection algorithms (Canny edge detection). Length and width calculation: Calculate the crack length and width based on the extracted crack contour. Length calculation: The length can be calculated by counting the number of pixels in the contour or by fitting the crack centerline using a fitting algorithm (least squares method) and then calculating the length of the centerline.

[0057] Where n and m represent the distance between non-zero pixels on the skeleton as 1 or a pixel interval, respectively. Width calculation: The width can be calculated by measuring the width at multiple locations along the vertical direction of the crack and then taking the average. Alternatively, morphological operations (distance transformation) can be used to calculate the crack width.

[0058] The area of ​​water leakage is determined by counting the number of pixels, that is, by counting the total number of pixels within the connected domain.

[0059] For a binary image that has undergone morphological processing, the background pixel value is 0 and the seepage water pixel value is 1. Count the number of gray values ​​of 1 in each connected component, i.e., f(x,y)=1, which is the area of ​​the seepage water.

[0060] This invention effectively overcomes the shortcomings of traditional manual inspection (low efficiency), laser scanning (high cost), and conventional deep learning models (high redundancy), providing a high-precision, automated, and low-cost solution for the health monitoring of subway tunnel structures. It significantly improves the reliability of defect identification and the speed of operation and maintenance response, and has important practical value and broad application prospects.

Claims

1. A method for intelligent detection of defects in subway tunnels based on YOLO-T, characterized in that, It includes the following steps: Step 1, data collection: Collect image data of crack and leakage diseases in the tunnel; Step 2, data preprocessing and dataset construction: Perform sliding window screenshot, spatial domain filtering denoising, block contrast stretching and binarization on the collected images, and construct a dataset containing three types of data: cracks, leakage and no diseases; Step 3, model construction: Construct an improved YOLO-T model, which includes a backbone network, a neck network and a head network; Step 4, model training and result prediction: Divide the dataset into a training set, a validation set and a test set, perform model training, and evaluate the classification accuracy through a confusion matrix; Step 5, development of a disease intelligent recognition system: Integrate the YOLO-T model, including a disease recognition module, a danger level analysis module, a visualization positioning module and a result management module.

2. The method according to claim 1, characterized in that, The specific parameters for the sliding window screenshot in step 2 are: window side length d, sliding step size s, obtaining Tx images in the width direction of the original image, and Ty images in the height direction, wherein the width w and height h of the original image satisfy... , The total number of images after the sliding window screenshot is .

3. The method according to claim 1, characterized in that, The backbone network contains an SCConv module and a NAM attention mechanism, and successively includes the first to fifth SCConv modules, the first to fourth RepNCSPELAN modules and the first NAM module; the neck network contains a CBLinear module and a CBFuse module, and successively includes the first to third CBLinear modules, the sixth to tenth SCConv modules, the fifth to eighth RepNCSPELAN modules and the first to third CBFuse modules; the head network contains a RepNCSPELAN module and an upsampling module, and successively includes the ninth to thirteenth RepNCSPELAN modules, the first to fourth connection modules and the eleventh to twelfth SCConv modules.

4. The method according to claim 1, characterized in that, The SCConv module consists of a spatial reconstruction unit SRU, a SiLu activation function and a channel reconstruction unit CRU, and is used to compress spatial and channel redundancy; the NAM attention mechanism suppresses non-significant features through weight sparse penalty; The neck network introduces programmable gradient information PGI, including an auxiliary reversible branch and multi-level auxiliary information, and is used to solve the gradient information bottleneck problem.

5. The method according to claim 1, characterized in that, Train the YOLO-T model according to the multi-class cross-entropy loss function, and divide the dataset into a training set, a validation set and a test set according to 8:1:

1. The multi-class cross-entropy loss function is specifically as follows: M - the number of categories; yic - a sign function, taking 0 or 1. If the true category of the sample is equal to, take 1, otherwise take 0; pic - the predicted probability that the observed sample belongs to the category.

6. The method according to claim 1, characterized in that, The danger level analysis module divides the danger level based on the crack length, width and leakage area: The crack danger level is divided into five levels according to the crack length L and the crack width a, including a < 1mm being minor, L < 5m and a < 3mm being medium, L < 5m and 3mm ≤ a ≤ 5mm being relatively serious, 5m ≤ L ≤ 10m and a > 5mm being serious, L > 10m and a > 5mm being extremely serious; the leakage danger level is divided into four levels according to the area A, including A < 0.5m² being minor, 0.5m² ≤ A ≤ 2m² being general, 2m² < A ≤ 5m² being serious, A > 5m² being critical.

7. The method according to claim 1, characterized in that, In step 5, the location of the defect is calculated by an incremental encoder. The encoder is installed on the hub of the detection platform. The pulse signal is combined with the encoder parameters to generate the absolute position coordinates inside the tunnel, and then aligned with the BIM model or GIS map to generate a 3D maintenance path planning map.

8. The method according to claim 1, characterized in that, The intelligent disease identification system also includes: API Interface Module: Provides standardized interfaces and documentation for third-party system integration; Interactive user manual: Includes video tutorials on software operation and a model training guide; Real-time model weight switching function: Supports dynamically loading different weight files without restarting the system.

9. The method according to claim 1, characterized in that, In the video processing module: Mode 1: Extract keyframes from the video at a fixed frame rate, convert them into image sequences, and perform batch detection; Mode 2: After improving video quality using a super-resolution enhancement algorithm, the video is analyzed frame by frame. The enhancement algorithm is based on a deep learning model.

10. The method according to claim 1, characterized in that, The length of the lesion is calculated by using skeleton pixel point statistics or least squares fitting of the center line, the width is calculated by using vertical multi-point measurement or distance transformation, and the leakage area is realized by counting the number of connected domain pixels.

Images