Deep foundation pit slope deformation monitoring method based on YOLOv5 and convolutional neural network

Abstract

The invention provides a deep foundation pit side slope deformation monitoring method based on YOLOv5 and a convolutional neural network, and belongs to the technical field of side slope deformation monitoring, and the method comprises the following steps: arranging a plurality of targets on a side slope supported by a truss, continuously collecting real-time images of the side slope through a high-definition camera, designing a GANs model to carry out data enhancement, and carrying out real-time monitoring on the real-time images of the side slope; an image data set is prepared, a YOLO v5 model is used for carrying out preliminary identification on an image, a CNN model is used for carrying out accurate center positioning on a target image, and real-time displacement calculation and monitoring are carried out. High precision and automation are realized, automatic identification and accurate positioning of the target are realized by combining the YOLOv5 model and the CNN model, and the limitation that a traditional method depends on manual calibration is overcome. In a complex environment, the system can quickly and accurately complete target positioning and displacement monitoring, and the monitoring precision and efficiency are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of slope deformation monitoring, and particularly to a deep foundation pit slope deformation monitoring method based on YOLOv5 and convolutional neural network. Background Art

[0002] Slope stability monitoring is a key link in geological disaster prevention and engineering safety assurance. With the continuous development of engineering construction, problems such as slope landslides and displacements bring great risks to the safe operation of projects. Although traditional slope monitoring methods can monitor slope displacements to a certain extent, they have problems such as long data collection cycles, low accuracy, and susceptibility to external interference, and cannot meet the requirements of modern projects for the efficiency and accuracy of slope stability monitoring. Therefore, it is particularly important to develop more accurate and highly automated slope monitoring technologies.

[0003] Traditional foundation pit deformation monitoring technologies mainly include methods such as total station, inclinometer, and anchor monitoring. Although these technologies are widely used in engineering, they generally have problems such as low monitoring frequency, poor real-time performance, dependence on manual operation, and susceptibility to environmental interference, and are difficult to meet the requirements of modern construction for efficient and accurate monitoring. For example, total station and inclinometer are weak in dynamic deformation capture ability and are relatively slow to respond to small-scale deformations and short-term changes. Anchor monitoring can analyze the slip surface, but its adaptability is poor under complex working conditions. Digital image correlation (DIC) technology has been taken seriously in displacement and strain monitoring in recent years, has the advantages of non-contact and full-field measurement, and can provide high-resolution deformation data, but its computational complexity is high and the processing efficiency is low. Especially in complex construction sites or environments with large changes in lighting conditions, the measurement accuracy and real-time performance are insufficient.

[0004] Existing slope displacement monitoring technologies mainly rely on digital image correlation method (DIC) or traditional measurement methods such as total station and GPS. Although these technologies can provide basic displacement monitoring functions, they have significant disadvantages. First, these technologies rely on manual selection and calibration of targets, are susceptible to environmental and human factors, and lead to a decrease in the accuracy of monitoring results. Second, these technologies are less sensitive in detecting small-scale displacement changes and are difficult to achieve early warning. In addition, the monitoring efficiency of traditional methods is low, and it is difficult to provide real-time and accurate monitoring results under complex terrains and harsh conditions. Summary of the Invention

[0005] The purpose of the present invention is to provide a deep foundation pit slope deformation monitoring method based on YOLOv5 and convolutional neural network to solve the technical problems of low existing monitoring efficiency and providing real-time and accurate monitoring results under complex terrains and harsh conditions.

[0006] YOLO is a real-time object detection system that can quickly identify and locate target objects in images, while CNN is good at image classification and feature extraction. Combining YOLO and CNN for slope monitoring can automatically identify the targets on the slope and accurately locate the center position of the targets, and then calculate the displacement of the slope. It can automatically identify and locate the targets in the image without manual intervention. This function greatly simplifies the process of slope displacement measurement, reduces the errors and workload caused by manual operations, and improves the automation degree and efficiency of measurement.

[0007] First, by combining the YOLO and CNN models, automated target detection and positioning are achieved, eliminating the dependence on manual operations, thereby improving the monitoring accuracy and reducing human errors. Using a generative adversarial network (GAN) to expand the dataset enhances the detection sensitivity to small displacement changes, improves the monitoring robustness in complex environments, and meets the requirements of high precision and rapid response in practical engineering applications.

[0008] To achieve the above objectives, the technical solutions adopted in the present invention are as follows:

[0009] A deep foundation pit slope deformation monitoring method based on YOLOv5 and convolutional neural network, the method comprising the following steps:

[0010] Step 1: Arrange a number of targets on the slope supported by trusses, and continuously collect real-time images of the slope through a high-definition camera;

[0011] Step 2: Design a GANs model for data augmentation;

[0012] Step 3: Prepare the image dataset;

[0013] Step 4: Use the YOLO v5 model to perform a preliminary identification of the images;

[0014] Step 5: Use the CNN model to accurately locate the center of the target image;

[0015] Step 6: Perform real-time displacement calculation and monitoring.

[0016] Furthermore, in Step 1: The images are large-scale engineering images in the actual construction site environment, the targets are patterns with obvious geometric shapes and a contrast higher than the set value, and the collected pictures are labeled with the target using the annotation software labelme, and the dataset pictures are divided into a training set and a test set according to a ratio of 7:3.

[0017] Furthermore, in step 2, since the number of target images actually collected on-site is limited, directly training the model will lead to overfitting and insufficient robustness. The original target images are input into the generative adversarial network (GANs) to generate synthetic target images similar to the actual construction site scenarios. The images generated by GANs enrich the diversity of the dataset and enhance the generalization ability of the model in complex environments. The extended training dataset is used to train the CNN model, thereby improving the accuracy of target center localization and the detection ability for small displacement changes.

[0018] Furthermore, in step 3, to measure the displacement of the foundation pit support slope, it is first necessary to establish an image dataset containing targets. The targets are arranged at several positions on the slope for accurately monitoring displacement changes. The image data includes two types: one is the on-site collected construction site images containing targets, and the other is the local target images intercepted by the YOLOv5 model. The central position of the target is marked as the main measurement index for monitoring the displacement changes of the target at different times, providing data support for model training and subsequent monitoring.

[0019] Furthermore, in step 4, the YOLO v5 model directly processes the actual engineering images and can detect the target positions in the images in real time. The YOLO v5 model can accurately locate the bounding boxes of all targets in the entire image, extract all the targets in the predicted image, and generate a preliminary detection result containing the target positions.

[0020] Furthermore, in step 5, the CNN model is used to accurately locate the center of the target images extracted by the YOLO v5 model. By further processing the image details of the target, the CNN model can determine the coordinates of the center point of the target. By analyzing the target area extracted by YOLO v5, the range is further narrowed to ensure the accuracy of the target center.

[0021] Furthermore, in step 6, after the CNN model completes the accurate localization of the target center, the displacement is calculated according to the formulas Δx = x2 - x1 and Δy = y2 - y1. By comparing the changes in the target center positions in two consecutive frames of images, the displacement of the slope is calculated in real time. By comparing the position differences of the target center points in the images taken at different times, the displacement of the target is calculated. The displacement reflects the actual displacement changes of the truss-supported slope. The calculation of the displacement is achieved by the coordinate differences of the target center in consecutive images, thereby obtaining the displacement data of the slope at different time points. The calculation results are stored in real time and compared with the previous data for analysis to evaluate the displacement trend of the slope.

[0022] Furthermore, in step 4, the YOLOv5 model framework includes an input end, a backbone network, a neck network, and a prediction layer.

[0023] Furthermore, in step 1, the adversarial network includes a generator and a discriminator. The generator starts from a random noise vector and gradually generates a grayscale image of 80×80 pixels through deconvolution layers. Each layer uses batch normalization and the LeakyReLU activation function to ensure the stability of the network and accelerate convergence. The discriminator consists of several convolutional network layers, using the LeakyReLU activation function and Dropout regularization layers to prevent overfitting. The discriminator finally outputs a scalar to judge the authenticity of the image. To prevent gradient explosion or disappearance, a gradient penalty term is added during the training process.

[0024] Furthermore, in step 5, the convolutional neural network includes three convolutional layers, a fully connected layer, and an output layer. The three convolutional layers use 32, 64, and 128 3×3 convolutional kernels. After each layer of convolution, a ReLU activation function and a 2×2 max-pooling layer are connected to gradually extract the spatial features of the image, reduce the dimension of the feature map, and retain key information at the same time. The feature maps extracted by the convolutional layer are flattened and then input into two fully connected layers, each containing 256 and 128 neurons, and 50% Dropout regularization is used to prevent overfitting. The output layer is used for the final output layer, which contains 2 neurons for predicting the two-dimensional coordinates of the target.

[0025] Due to the adoption of the above technical solutions, the present invention has the following beneficial effects:

[0026] (1) High precision and automation. By combining the YOLOv5 and CNN models, automatic identification and precise positioning of the target are achieved, overcoming the limitation of traditional methods relying on manual calibration. In a complex environment, the system can quickly and accurately complete target positioning and displacement monitoring, significantly improving the monitoring accuracy and efficiency.

[0027] (2) Real-time monitoring and early warning. Real-time calculation and monitoring of slope displacement are realized, and combined with an automatic early warning system, an alarm can be issued in a timely manner when abnormal displacement is detected. Compared with the prior art, the present invention not only improves the real-time performance but also provides more efficient safety protection, meeting the monitoring requirements under complex working conditions and changing environments.

[0028] (3) Automatic identification and precise positioning of the target on the foundation pit support slope are realized in a complex environment, reducing manual intervention, improving the identification efficiency and accuracy. The generative adversarial network (GANs) is used to expand the target dataset, enhancing the sensitivity of the model in detecting small displacement changes and the robustness in complex environments, improving the adaptability of the monitoring system. By comparing the changes in the target center position in consecutive images, real-time calculation of slope displacement is realized, and combined with an automatic early warning system, an efficient safety early warning function is provided, supporting remote control and real-time data update. Through efficient data transmission and processing, remote viewing and real-time adjustment of monitoring results are realized, improving the operation convenience of the system. Description of the Drawings

[0029] Figure 1 is the flowchart of the method of the present invention;

[0030] Figure 2 is the flowchart of image recognition of the present invention;

[0031] Figure 3 is the framework diagram of the YOLOv5 model of the present invention;

[0032] Figure 4 is the structure diagram of GANS of the present invention;

[0033] Figure 5 is the structure diagram of the CNN model of the present invention;

[0034] Figure 6 is the on-site test diagram of the construction site of the present invention;

[0035] Figure 7 is the prediction result diagram of yolo v5 of the present invention;

[0036] Figure 8 is the predicted target center diagram of CNN of the present invention;

[0037] Figure 9 is the diagram for calculating the displacement of each target of the present invention. Detailed Embodiments

[0038] To make the objectives, technical solutions and advantages of the present invention more clear and understandable, the following preferred embodiments are given with reference to the accompanying drawings, and the present invention is further described in detail. However, it should be noted that many details listed in the specification are only for the reader to have a thorough understanding of one or more aspects of the present invention, and these aspects of the present invention can be implemented even without these specific details.

[0039] As Figure 1-2 shown, a deep foundation pit slope deformation monitoring method based on YOLOv5 and convolutional neural network, the method includes the following steps:

[0040] Step 1: Image acquisition:

[0041] Arrange multiple targets on the slope supported by the truss, and continuously collect real-time images of the slope through a high-definition camera. These images are large-scale engineering images in the actual construction site environment, and the targets are patterns with obvious geometric shapes and high contrast, such as circles or crosses, to ensure clear recognition under various lighting conditions. Use the annotation software labelme to annotate the targets in the collected pictures, and divide the dataset pictures into a training set and a test set according to a ratio of 7:3.

[0042] Step 2: GANs model data augmentation:

[0043] Since the number of target images actually collected on site is limited, directly training the model may lead to overfitting and insufficient robustness. Therefore, the original target images are input into a generative adversarial network (GANs) to generate synthetic target images similar to the actual construction site scenarios. The images generated by GANs enrich the diversity of the dataset and significantly improve the generalization ability of the model in complex environments. The extended training dataset is used to train the CNN model, thereby improving the accuracy of target center localization and the detection ability for small displacement changes.

[0044] Since the original dataset is small in scale and single in sample, this paper enhances the dataset through a generative adversarial network (GAN) to improve the generalization ability and prediction accuracy of the model. The generative adversarial network consists of two parts: a generator and a discriminator, and Wasserstein GAN with Gradient Penalty (WGAN-GP) is used as the training framework. The GAN structure is as Figure 4 shown.

[0045] Generator: The generator starts from a random noise vector (with a length of 100) and gradually generates a grayscale image of 80×80 pixels through deconvolution layers. Batch normalization and the LeakyReLU activation function are used in each layer to ensure the stability of the network and accelerate convergence.

[0046] Discriminator: The discriminator consists of a multi-layer convolutional network, using the LeakyReLU activation function and a Dropout regularization layer to prevent overfitting. The discriminator finally outputs a scalar to judge the authenticity of the image. To prevent gradient explosion or disappearance, a gradient penalty term is added during the training process.

[0047] Training and optimization: The Adam optimizer is adopted, and the learning rates of both the generator and the discriminator are 0.001. At the end of each epoch, the loss values of the generator and the discriminator are recorded to monitor the stability of the training process. Table 1 shows the structural parameters of the GANs model.

[0048] Table 1 Structural parameters of the GANs model

[0049]

[0050] Step 3: Dataset preparation:

[0051] To achieve displacement measurement of the foundation pit support slope, it is first necessary to establish an image dataset containing targets. These targets are arranged at multiple positions on the slope for accurately monitoring displacement changes. The image data includes two categories: one is the on-site collected construction site images containing targets, and the other is the local images of the targets intercepted by the YOLOv5 model. In these images, the central positions of the targets are marked as the main measurement indicators to monitor the displacement changes of the targets at different times, providing data support for model training and subsequent monitoring.

[0052] Step 4: Initial identification by the YOLO v5 model:

[0053] The YOLO v5 model directly processes actual engineering images and can detect the positions of the targets in the images in real time. With its fast and efficient characteristics, the YOLOv5 model can accurately locate the bounding boxes of all targets in the entire image. Extract all the targets in the predicted image to generate a preliminary detection result containing the target positions.

[0054] As Figure 3 shown, the YOLOv5 model framework consists of four main parts: the input end, the backbone network, the neck network, and the prediction layer.

[0055] Input end: YOLOv5 introduces an adaptive image scaling technique to dynamically adjust the size of the input image according to the target size. The default image size is 640×640 pixels. When the input image is smaller than the standard size, the network enlarges the image; otherwise, it shrinks, thus enhancing the model's robustness to targets of different scales.

[0056] Backbone network: YOLOv5 uses CSPDarknet53 as the backbone network and combines the Focus structure and the CSP structure. The CSP structure improves the combination effect of low-level detailed features and high-level abstract features by dividing the feature map into two parts, processing them through sub-networks and directly, and then fusing them. The Focus structure effectively reduces the computational amount and retains important feature information while improving the feature extraction efficiency by dividing the input feature map into four sub-maps and using channel splicing for downsampling and feature compression.

[0057] Neck network: The neck network is used to enhance the features extracted by the backbone network and improve the detection performance. In YOLOv5, the neck network adopts the SPP (Spatial PyramidPooling) structure and the PAN (Path Aggregation Network) structure. The SPP structure enhances the perception ability of the feature map, and the PAN structure improves the detection accuracy of the model through multi-scale feature fusion.

[0058] Prediction layer: The prediction layer uses multi-scale output to generate prediction results at different scales (80×80, 40×40, 20×20). The output includes bounding box coordinates, classes, and confidence levels, enabling accurate detection of objects of different sizes.

[0059] Compared with YOLOv4, the main improvement of YOLOv5 lies in the adaptive image scaling technology, which enables the network to adapt to different input sizes and avoids size limitations. In addition, the backbone network introduces the Focus structure, which not only reduces the computational complexity but also effectively retains important feature information. Table 2 shows the training parameters of the YOLOv5 model.

[0060] Table 2 Training parameters of the YOLOv5 model

[0061]

[0062] Step 5: Precise localization of the CNN model:

[0063] Next, use CNN to precisely locate the center of the target image extracted by the YOLO v5 model. By further processing the image details of the target, CNN can determine the center point coordinates of the target. This step further narrows down the range by analyzing the target area extracted by YOLO v5 to ensure the accuracy of the target center.

[0064] To further improve the accuracy of object detection and localization, this paper designs a multi-layer model architecture based on convolutional neural network (CNN), as Figure 5 shown. This model combines the functions of feature extraction and coordinate prediction.

[0065] Convolutional layer: The CNN model consists of three convolutional layers, using 32, 64, and 128 3×3 convolutional kernels. After each layer of convolution, a ReLU activation function and a 2×2 max pooling layer are connected to gradually extract the spatial features of the image, reduce the dimension of the feature map, and retain key information at the same time.

[0066] Fully connected layer: The feature maps extracted by the convolutional layer are flattened and then input into two fully connected layers, each containing 256 and 128 neurons, and 50% Dropout regularization is used to prevent overfitting.

[0067] Output layer: The final output layer contains 2 neurons for predicting the two-dimensional coordinates (x, y) of the target.

[0068] Training and optimization: The model uses the mean squared error (MSE) as the loss function, the Adam optimizer, and the learning rate is 0.001. Through the StepLR learning rate scheduler, the learning rate decays by 5% every 100 epochs of training to accelerate the convergence of the model. Table 3 shows the training parameters of the CNN model.

[0069] Table 3 Hyperparameters for CNN Model Training

[0070]

[0071] The implementation steps for measuring the displacement of the truss-supported slope based on the combination of YOLO and CNN generative adversarial networks are as follows Figure 2 shown

[0072] This paper aims to solve the problems of poor real-time performance, low efficiency, and dependence on manual operation in traditional foundation pit support structure monitoring technologies, and proposes an automated monitoring method based on deep learning to improve monitoring accuracy and efficiency

[0073] Step 6: Real-time displacement calculation and monitoring

[0074] After the CNN model completes the precise positioning of the target center, the displacement is calculated according to the formulas Δx = x2 - x1 and Δy = y2 - y1. By comparing the changes in the positions of the target centers in two consecutive frames of images, the displacement of the slope is calculated in real time. By comparing the position differences of the target center points in the images taken at different times, the displacement of the target is calculated. These displacements reflect the actual displacement changes of the truss-supported slope. The calculation of the displacement is achieved by the coordinate differences of the target center in consecutive images, thereby obtaining the displacement data of the slope at different time points. The calculation results are stored in real time and compared with the previous data for analysis to evaluate the displacement trend of the slope

[0075] The specific experimental setup and experimental verification are as follows

[0076] To verify the effectiveness of the proposed method, this study conducted experiments at the site of a foundation pit project in Nanning, Guangxi (as Figure 6 shown). The foundation pit has dimensions of 12.6 meters in width, 24.5 meters in length, and 7 meters in depth, and adopts an assembled cable-stayed space steel truss multi-layer internal support system. The earth excavation is carried out in a "layered and symmetric excavation" manner, with slope release layer by layer and simultaneous installation of the support structure. When excavated to different depths, the construction of the capping beam, installation of the steel truss and waist beam are completed in sequence, and vertical and horizontal prestresses are applied to balance the earth pressure and stabilize the structure. Targets are arranged in different areas on the left slope (as Figure 6 (c)), covering the slope top (L11 - L16), the middle part of the slope surface (L21 - L23), and the capping beam position (L31 - L34) to monitor the slope top stability, the deformation characteristics of the middle section, and the stress state of the capping beam respectively. The targets are designed as black and white circular patterns (as Figure 6 (a)), fixed by metal brackets, and have high-contrast patterns to enhance the recognition rate and detection robustness under complex lighting conditions. Two high-resolution industrial cameras are arranged on both sides of the foundation pit and connected to the construction site computer through Gigabit Ethernet to collect the target displacement data in real time (as Figure 6(b)). The camera supports external trigger mode synchronous acquisition, providing high-quality input for subsequent YOLOv5 target detection and CNN precise positioning, and enabling remote monitoring and parameter adjustment.

[0077] The YOLO v5 target prediction results are as Figure 7 shown. The positions of the targets are marked in the image in the form of bounding boxes. The part marked with a purple border is the target position on the left slope, and the green border is the target position on the right slope. The predicted confidence score shown above each target (such as "target 0.81") represents the model's confidence in the detection result, and the complex terrain and surface features around the targets do not cause obvious interference to the detection.

[0078] After the YOLOv5 detects the target positions, the detected target regions are input into the CNN model to accurately predict the central coordinates of the targets. YOLOv5 reduces the image region to an 80×80 pixel region by predicting the bounding boxes and coordinate values of the targets. The target center results predicted by the CNN are as Figure 8 shown. The red stars in the figure are the target centers predicted by the CNN model, which highly coincide with the actual target center positions. By combining the preliminary detection of YOLOv5 and the precise coordinate prediction of the CNN, this study constructs a complete target recognition and center positioning framework.

[0079] Figure 9 For the working condition of foundation pit excavation and prestress application, the displacement changes in the vertical direction of different monitoring groups (marked as L11 to L33) are recorded. The line chart in the figure represents the displacement changing with time, and the downward displacement is positive. It can be seen from the figure that the displacement values of all groups reach a peak on the fifth day and then gradually rebound to a level close to zero. This peak reflects the maximum displacement caused by the reduction of soil support and the increase of lateral pressure when the excavation reaches a certain depth. Subsequently, with the effectiveness of the support structure and the application of prestress, the displacement gradually rebounds to a level close to the original level. This rebound indicates that the prestress and the support structure effectively stabilize the soil around the foundation pit, prevent further displacement, and ensure the overall safety and stability of the foundation pit.

[0080] Automated target detection and positioning are achieved:

[0081] The present invention combines the YOLOv5 and CNN models to achieve automated recognition and precise positioning of the targets on the foundation pit support slopes in complex environments, reducing manual intervention and improving the recognition efficiency and accuracy.

[0082] Data augmentation technology:

[0083] Expanding the target dataset using Generative Adversarial Networks (GANs) enhances the sensitivity of the model in detecting small displacement changes and its robustness in complex environments, improving the adaptability of the monitoring system.

[0084] Real-time displacement calculation and monitoring:

[0085] By comparing the changes in the target center positions in consecutive images, the real-time calculation of slope displacement is achieved, and combined with an automated warning system, an efficient safety warning function is provided.

[0086] Remote monitoring and real-time feedback:

[0087] The system supports remote control and real-time data update. Through efficient data transmission and processing, remote viewing and real-time adjustment of the monitoring results are achieved, improving the operational convenience of the system.

[0088] Matters not covered by this invention are well-known techniques.

[0089] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A deep foundation pit slope deformation monitoring method based on YOLOv5 and convolutional neural network, characterized in that: The method comprises the following steps: Step 1: Arrange several targets on the slope supported by the truss, and continuously collect real-time images of the slope through a high-definition camera; Step 2: Design a GANs model for data enhancement; Step 3: Prepare image dataset; Step 4: Use the YOLO v5 model to perform preliminary recognition of the image; Step 5: Use the CNN model to accurately locate the center of the target image; Step 6: Perform real-time displacement calculation and monitoring.

2. A deep foundation pit slope deformation monitoring method based on YOLOv5 and convolutional neural network according to claim 1, characterized in that: In step 1: the image is a large-scale engineering image in the actual construction site environment. The target is a pattern with obvious geometric shape and contrast higher than the set value. The labeling software labelme is used to annotate the target on the collected image, and the data set images are divided into training set and test set in a ratio of 7:

3.

3. The method for monitoring deformation of deep foundation pit slope based on YOLOv5 and convolutional neural network according to claim 1, characterized in that: In step 2, due to the limited number of target images actually collected on site, directly training the model will lead to overfitting and insufficient robustness. The original target images are input into the generative adversarial network (GANs) to generate synthetic target images similar to the actual construction site scenes. The images generated by GANs enrich the diversity of the data set and improve the generalization ability of the model in complex environments. The expanded training data set is used to train the CNN model to improve the accuracy of target center positioning and the ability to detect small displacement changes.

4. The method for monitoring deformation of deep foundation pit slope based on YOLOv5 and convolutional neural network according to claim 1, characterized in that: In step 3, in order to realize the displacement measurement of the foundation pit support slope, it is first necessary to establish an image dataset containing targets. The targets are arranged at several positions on the slope for accurate monitoring of displacement changes. The image data includes two categories: one is the construction site image containing targets collected on-site, and the other is the local image of the target captured by the YOLOv5 model. The center position of the target is marked as the main measurement indicator, which is used to monitor the displacement changes of the target at different times and provide data support for model training and subsequent monitoring.

5. The method for monitoring deformation of deep foundation pit slope based on YOLOv5 and convolutional neural network according to claim 1, characterized in that: In step 4, the YOLO v5 model directly processes the actual engineering image and can detect the target position in the image in real time. The YOLO v5 model can accurately locate the bounding boxes of all targets in the entire image, extract all targets in the predicted image, and generate a preliminary detection result including the target position.

6. The method for monitoring deformation of deep foundation pit slope based on YOLOv5 and convolutional neural network according to claim 1, characterized in that: In step 5, the CNN model is used to accurately locate the center of the target image extracted by the YOLO v5 model. The CNN model can determine the coordinates of the target center point by further processing the image details of the target. By analyzing the target area extracted by YOLO v5, the range is further narrowed to ensure the accuracy of the target center.

7. The method for monitoring deformation of deep foundation pit slope based on YOLOv5 and convolutional neural network according to claim 1, characterized in that: In step 6, after the CNN model completes the precise positioning of the target center, the displacement is calculated according to the formula Δx=x2-x1, Δy=y2-y1. The displacement of the slope is calculated in real time by comparing the position change of the target center in two consecutive frames of images. The displacement of the target is calculated by comparing the position difference of the target center point in the images taken at different times. The displacement reflects the actual displacement change of the truss-supported slope. The displacement is calculated by the coordinate difference of the target center in the consecutive images, thereby obtaining the displacement data of the slope at different time points. The calculation results are stored in real time and compared with the previous data for analysis to evaluate the displacement trend of the slope.

8. The method for monitoring deformation of deep foundation pit slope based on YOLOv5 and convolutional neural network according to claim 1, characterized in that: In step 4, the YOLOv5 model framework includes the input end, backbone network, neck network and prediction layer.

9. The method for monitoring deformation of deep foundation pit slope based on YOLOv5 and convolutional neural network according to claim 1, characterized in that: In step 1, the adversarial network includes a generator and a discriminator. The generator starts with a random noise vector and gradually generates a grayscale image of 80×80 pixels through a deconvolution layer. Each layer uses batch normalization and LeakyReLU activation function to ensure the stability of the network and accelerate convergence. The discriminator consists of several layers of convolutional networks, using LeakyReLU activation function and Dropout regularization layer to prevent overfitting. The discriminator finally outputs a scalar to judge the authenticity of the image. In order to prevent the gradient from exploding or disappearing, a gradient penalty term is added during the training process.

10. The method for monitoring deformation of deep foundation pit slope based on YOLOv5 and convolutional neural network according to claim 1, characterized in that: In step 5, the convolutional neural network includes three convolutional layers, a fully connected layer and an output layer. The three convolutional layers use 32, 64 and 128 3×3 convolution kernels. Each convolution layer is followed by a ReLU activation function and a 2×2 maximum pooling layer to gradually extract the spatial features of the image and reduce the dimension of the feature map while retaining key information. The feature map extracted by the convolutional layer is flattened and input into two fully connected layers, each containing 256 and 128 neurons, and 50% Dropout regularization is used to prevent overfitting. The output layer is used for the final output layer, which contains 2 neurons for predicting the two-dimensional coordinates of the target.

Images