Construction fence detection method and system based on graph convolution network

Abstract

The application provides a construction fence detection method and system based on a graph convolution network, which comprises the following steps: collecting high-definition images of a power construction site and preprocessing (denoising, enhancing contrast, and normalizing), constructing graph structure data (fences and environmental features as nodes, and spatial relationships as edges), inputting the graph convolution network to extract deep features, calculating standard detection loss through cross-entropy loss, optimizing and enhancing in combination with adjacent node features, designing a joint loss function containing a graph structure context loss, updating parameters through back propagation, and realizing construction fence state detection. Through graph structure modeling and a joint loss function, the application improves detection performance in a complex environment, simplifies knowledge transmission, does not increase the model size and inference complexity, and is lightweight and suitable for edge devices, thereby providing an effective solution for deploying a high-performance and high-robustness target detection model on a resource-limited device.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, specifically a construction fence detection method and system based on graph convolutional networks. Background Technology

[0002] Object detection, as a core technology of computer vision, plays a crucial role in identifying violations and abnormal equipment conditions at power construction sites. Compared to general scenarios, power construction sites require precise location of minute targets such as cross-boundary operations in energized areas, unauthorized use of safety tools, and workers not wearing protective gear. Furthermore, the complex environment of power construction sites, with factors such as strong electromagnetic fields and equipment obstruction, poses challenges to detection accuracy. While YOLO series models based on deep learning have achieved a balance between accuracy and speed in general monitoring, their direct application in power construction scenarios still struggles to meet the dual demands of high-precision feature recognition and lightweight deployment of edge devices. Therefore, developing a construction fence detection method adapted to power construction scenarios is particularly important to improve the accuracy and efficiency of fence safety management. Graph Convolutional Networks (GCNs), as an emerging deep learning technology, offer new ideas and effective methods for solving the fence detection problem at power construction sites due to their excellent feature learning capabilities and ability to process graph-structured data.

[0003] However, high-performance object detection models typically have a large number of parameters and computational complexity, which limits their deployment on resource-constrained devices such as edge computing platforms or mobile devices. Especially in power construction sites, real-time monitoring and analysis often require low-power devices; therefore, lightweight and efficient models are crucial. To address this issue, Graph Convolutional Networks (GCNs), as an emerging deep learning technique, offer an effective solution. GCNs possess excellent feature learning capabilities and can effectively handle graph-structured data, making them ideal for complex feature extraction, particularly suitable for fence detection tasks in power construction scenarios.

[0004] Despite the significant success of graph convolutional networks (CNNs) in image classification and other fields, they still face several challenges in target detection applications for power safety monitoring. The complex and variable environment of power construction projects, including electromagnetic interference and equipment obstruction, makes it difficult to accurately extract features of fences and their surroundings. For example, strong electromagnetic fields can increase image noise, and equipment obstruction can render parts of the fence invisible, both of which increase the difficulty of detection. Furthermore, traditional convolutional neural networks (CNNs) perform poorly when handling small targets and subtle features. Construction fences are often small and may blend into the background, making traditional methods prone to false positives or false negatives in such cases. Summary of the Invention

[0005] This invention aims to overcome the shortcomings of existing technologies in construction fence detection within power safety inspection scenarios. It provides a construction fence detection method and system based on graph convolutional networks, acquiring the graph structure features of the fence and its surrounding environment in a novel way to achieve efficient and accurate construction fence status detection. This invention is applicable to the complex environmental conditions of power construction sites, significantly improving detection accuracy while maintaining the model's efficiency and lightweight design.

[0006] According to one aspect of the present invention, a construction fence detection method based on graph convolutional networks is provided, comprising the following steps:

[0007] Data collection and preprocessing: Collect high-definition image data from power construction sites, perform noise reduction, contrast enhancement, and normalization preprocessing on the high-definition image data, and output preprocessed images;

[0008] Graph structure data construction: Based on the preprocessed image, features of the fence and its surrounding environment are extracted as nodes, and the nodes are connected by edges that reflect spatial relationships, interactions, and feature similarities to form graph structure data;

[0009] Forward propagation of graph convolutional networks: The constructed graph structure data is input into the graph convolutional network, and deep features of the fence and environment are extracted through multiple layers of graph convolution, and the fence feature representation is output.

[0010] Standard detection loss calculation: Based on the fence feature representation and the ground truth label, the standard detection loss is calculated using the cross-entropy loss function;

[0011] Feature optimization and enhancement: By combining the feature information of neighboring nodes represented by fence features, a feature fusion strategy is implemented to improve the robustness of graph convolutional networks;

[0012] No new loss function design: A joint loss function is designed based on graph structure data, which combines the standard detection loss with the graph structure context loss in a weighted manner;

[0013] Model parameter update: Based on the joint loss function, the weight parameters of the graph convolutional network are updated through the backpropagation mechanism to obtain the trained graph convolutional network;

[0014] Construction fence detection: Input new power construction site image data into a trained graph convolutional network, and output the state detection results of the construction fence.

[0015] Furthermore, the nodes of the graph structure data include fence components, construction tools, equipment, and background objects, and the weights of the edges are determined by calculating the spatial distance and feature similarity between nodes.

[0016] Furthermore, the graph convolutional network adopts a multi-layer structure, with each layer dynamically adjusting node weights through a self-attention mechanism to enhance the expression of key features.

[0017] Furthermore, the formula for calculating the joint loss function is as follows:

[0018] Total loss = α × Standard detection loss + (1-α) × Graph structure context loss

[0019] Where α is the weight coefficient, and 0 < α < 1; the graph structure context loss is calculated by the difference in node spatial coordinates and the deviation of edge feature similarity.

[0020] Furthermore, the high-definition image data includes images of construction fences under complex environmental conditions, such as high light, low light, and rainy weather.

[0021] A construction fence detection system based on graph convolutional networks includes:

[0022] The data collection and preprocessing module is used to collect high-definition image data from power construction sites, perform noise reduction, contrast enhancement and normalization preprocessing on the high-definition image data, and output preprocessed images.

[0023] The graph structure data construction module is used to extract features of the fence and its surrounding environment as nodes based on preprocessed images, and connect the nodes through edges that reflect spatial relationships, interactions, and feature similarities to form graph structure data.

[0024] The graph convolutional network forward propagation module is used to input graph structure data into the graph convolutional network, extract deep features of the fence and environment through multi-layer graph convolution, and output fence feature representation;

[0025] The standard detection loss calculation module is used to calculate the standard detection loss based on the fence feature representation and the ground label using the cross-entropy loss function.

[0026] The feature optimization and enhancement module is used to combine the feature information of neighboring nodes represented by the fence features to implement a feature fusion strategy and improve the robustness of fence detection.

[0027] A new loss function design module is used to design joint loss functions based on graph structure data, which weights and combines the standard detection loss with the graph structure context loss.

[0028] The model parameter update module is used to update the weight parameters of the graph convolutional network based on the joint loss function through the backpropagation mechanism, so as to obtain the trained graph convolutional detection network.

[0029] The construction fence detection module is used to input new power construction site image data into a trained graph convolutional detection network and output the status detection results of the construction fence.

[0030] Furthermore, the nodes of the graph structure data include fence components, construction tools, equipment, and background objects, and the weights of the edges are determined by calculating the spatial distance and feature similarity between nodes.

[0031] Furthermore, the graph convolutional network adopts a multi-layer structure, with each layer dynamically adjusting node weights through a self-attention mechanism to enhance the expression of key features.

[0032] Furthermore, the formula for calculating the joint loss function is as follows:

[0033] Total loss = α × Standard detection loss + (1-α) × Graph structure context loss

[0034] Where α is the weight coefficient, and 0 < α < 1; the graph structure context loss is calculated by the difference in node spatial coordinates and the deviation of edge feature similarity.

[0035] Furthermore, the high-definition image data includes images of construction fences under complex environmental conditions, such as high light, low light, and rainy weather.

[0036] The beneficial effects of this invention are mainly reflected in the following aspects:

[0037] I. Improve detection precision and accuracy

[0038] By constructing graph structure data of the fence and its surrounding environment (nodes represent fence parts, tools, equipment, etc., and edges reflect spatial relationships, feature similarity, etc.), the graph convolutional network can capture complex relationships. Combined with feature fusion strategies, it enhances the ability to extract deep features and significantly improves the detection accuracy of the construction fence status (normal, missing, damaged, etc.).

[0039] The design incorporates a joint loss function that includes standard detection loss (cross-entropy loss) and graph structure context loss, enhancing the model's ability to learn target features, especially improving sensitivity to fence damage and missing states, thereby further ensuring detection accuracy.

[0040] II. Enhancing robustness and environmental adaptability

[0041] For complex environments at power construction sites (such as electromagnetic interference, equipment obstruction, high / low light, and different weather conditions), the model has strong robustness through graph structure feature learning and training with multi-environment data, and can still maintain high detection performance under challenging conditions such as environmental changes and feature interference.

[0042] The input data covers images under various environmental conditions, which improves the model's generalization ability and can adapt to the actual needs of different construction sites.

[0043] III. Maintaining Model Efficiency and Lightweight

[0044] The feature extraction method based on graph convolutional networks simplifies the original image feature processing process. At the same time, by optimizing the model structure (such as multi-layer GCN and self-attention mechanism to dynamically adjust node weights), the model is lightweight while ensuring accuracy, making it suitable for deployment on edge computing devices and meeting real-time monitoring needs.

[0045] IV. Supporting Intelligent and Automated Safety Monitoring Applications

[0046] Real-time monitoring capabilities enable unattended automated safety management, significantly reducing labor costs at power construction sites and improving construction safety.

[0047] By combining graph convolutional network technology with the actual needs of power safety monitoring, this study lays the foundation for building an intelligent and automated power safety monitoring solution, and promotes the industry towards a safer and more efficient direction. Attached Figure Description

[0048] Figure 1 This is a diagram of the convolutional network detection framework according to an embodiment of the present invention;

[0049] Figure 2 This is a diagram of the convolutional network training framework according to an embodiment of the present invention;

[0050] Figure 3 This is a flowchart of the convolutional network testing process according to an embodiment of the present invention;

[0051] Figure 4 This is a flowchart of a construction fence detection method based on graph convolutional networks according to an embodiment of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Please see Figure 4 This invention provides a construction fence detection method based on graph convolutional networks, comprising the following steps:

[0054] a) Data collection and preprocessing

[0055] Collect and prepare image data of the power construction site. These images should include examples of various fence conditions, including normal, missing, and damaged fences. The collected data undergoes preprocessing, including noise reduction, brightness and contrast enhancement, and normalization, to output preprocessed images to ensure image quality.

[0056] b) Graph Structure Data Construction: The preprocessed image is transformed into graph structure data. In this process, features of the fence and its surrounding environment are extracted as nodes, and the relationships between these nodes are represented by edges. For example, each part of the fence can be considered an independent node, while other elements in the environment (such as tools, equipment, obstacles, etc.) can also be considered nodes. This graph structure construction provides a rich information foundation for subsequent graph convolutional network models.

[0057] At the end of this stage, a dataset containing rich graph structure features will be obtained, laying the foundation for subsequent model training.

[0058] c) Forward propagation of graph convolutional networks:

[0059] The model is trained using Graph Convolutional Networks (GCNs). First, the structure of the GCN needs to be defined, and appropriate layers and node features need to be selected to effectively extract fence features. During this process, algorithms such as Stochastic Gradient Descent (SGD) or the Adam optimizer are used to optimize the model parameters. This framework provides a complete set of tools for training, validating, and deploying GCN models, ensuring effective handling of fence detection tasks in complex construction environments.

[0060] During training, the graph structure data constructed in step b) is input into the graph convolutional network. The model's forward propagation process generates feature representations of the fenced-off states.

[0061] d) Standard Detection Loss Calculation: Calculate the standard detection loss, including classification loss and regression loss, to ensure the model can accurately identify the fence state. By retaining the base loss, ensure the model has the ability to fit real samples. At the end of this stage, a set of standard detection loss data is obtained, including the total loss value and details of each sub-item loss, providing a basis for subsequent loss optimization.

[0062] e) Feature optimization and enhancement: Combining the feature information of neighboring nodes represented by the fence features in step c), a feature fusion strategy is implemented to improve the robustness of the graph convolutional network;

[0063] f) Design of New Loss Functions: In this stage, the performance of graph convolutional networks is analyzed and optimized. Based on the standard detection loss, specific loss functions are designed to enhance the model's learning ability. For example, a joint loss combining graph structure features can be used to combine the standard detection loss with the contextual information loss to improve the model's detection accuracy in complex environments. The joint loss can be a weighted combination of the standard detection loss and the graph structure context loss.

[0064] g) Model parameter update: The model parameters are updated based on the optimized loss function to continuously improve the detection performance of construction fences. The weights of the graph convolutional network are gradually adjusted through backpropagation to better fit the training data and improve detection accuracy.

[0065] Next, the model is evaluated using a validation set during training to ensure its generalization ability. Cross-validation is used to monitor the model's performance on the validation set to avoid overfitting. Based on the validation results, hyperparameters such as the learning rate and batch size are adjusted as needed to further optimize model performance. Finally, the model's final output is generated, enabling real-time monitoring and early warning of construction fence status in practical applications.

[0066] During model training (e.g.) Figure 2 As shown, the method of this invention extracts and learns fence features through a graph convolutional network (GCN). During training, algorithms such as stochastic gradient descent (SGD) or the Adam optimizer are used to optimize model parameters. Simultaneously, appropriate learning rates and batch sizes are set to achieve optimal training results. For the loss function, the embodiment defines a composite loss that combines graph structure features, integrating the standard detection loss with the contextual information loss to enhance the model's ability to detect fence states.

[0067] f) Construction Fence Detection: Once the model training is complete, the inference phase begins. In this phase, the trained graph convolutional network is used to perform real-time detection on new construction site images (e.g., ...). Figure 1 (As shown). The input image is converted into graph-structured data, which is then fed into a trained model for forward inference to generate detection results for the fence state.

[0068] During the validation phase, cross-validation is used to ensure the model's generalization ability. By dividing the dataset into training and validation sets, the model's performance on the validation set is continuously monitored to prevent overfitting.

[0069] During inference, the model can identify the status of construction fences in real time and issue timely alarms when anomalies are detected. This process achieves efficient and accurate monitoring of construction fence status, providing strong support for safety management in the power industry. The method of this invention can run in real time on edge computing devices, and combined with a lightweight graph convolutional network structure, ensures the efficiency and accuracy of construction fence status detection. By monitoring the status of construction site fences in real time, the model can issue timely warnings, enhancing the effectiveness of construction safety management.

[0070] Embodiments of this invention can be implemented in common deep learning frameworks, such as PyTorch. The implementation steps of this invention, from data preparation to model inference, form a complete graph convolutional neural network method for detecting construction fences in power safety inspection scenarios. By utilizing graph-structured data, calculating and optimizing standard detection loss, and performing real-time model inference, this invention provides an efficient and accurate solution, promoting intelligent and automated management in the power industry.

[0071] When training using the method of this invention, firstly, a Graph Convolutional Network Trainer (GCNTrainer) class is instantiated, and necessary configuration parameters are passed in, such as the teacher model path, temperature T, and alpha value. Next, the framework's training function (such as PyTorch's training loop) is called, and the GCNTrainer instance is used as the trainer. The training process is managed by GCNTrainer, covering data loading, forward propagation of the student model, loss calculation (through a rewritten loss function, internally including forward calls to the teacher model and knowledge distillation loss calculation), backpropagation (only updating the weights of the student model), and the optimizer step.

[0072] In this way, the model can fully utilize the knowledge of the teacher model to improve the performance of the student model in construction fence detection. To verify the effectiveness of the method, a series of experiments were conducted, particularly testing on challenging power safety inspection scenario datasets (such as...). Figure 3 (As shown). This dataset contains images of construction fences taken under different lighting and weather conditions, and aims to evaluate the robustness of object detection models in complex environments.

[0073] In the experimental environment, suitable computing hardware for graph convolutional networks must first be configured. High-performance GPUs (such as NVIDIA RTX series or Tesla V100) are recommended to accelerate model training and inference. Optimized computing resource configuration can significantly improve training efficiency and support the processing of large-scale datasets. The experimental environment of this invention, through high-performance computing hardware, precise data processing, and optimized model training strategies, provides comprehensive technical support for construction fence detection in power safety inspection scenarios, promoting the intelligent and automated process of safety management in the power industry.

[0074] Baseline Model: A lightweight student model architecture (e.g., yolov10n.pt) is trained on a power safety inspection scenario dataset using standard training procedures, without any knowledge distillation techniques. Distillation Model (Invention): The same student model architecture (yolov10n) is used, but trained via the GCNTrainer of this invention. A larger, more powerful model (e.g., yolov10x.pt) is selected as the teacher model, and distillation parameters are set, for example, T=4.0 and α=0.5.

[0075] Performance metrics: Evaluation is conducted using standard target detection and assessment metrics, including:

[0076] Precision

[0077] Recall

[0078] mAP@50 (mean precision when IoU threshold is 0.5)

[0079] mAP@50-95 (IoU threshold from 0.5 to 0.95, average mAP in steps of 0.05)

[0080] Regarding the dataset, the implementation example utilizes image data captured at actual power construction sites, including instances of normal, missing, and damaged fences. These images were pre-processed and labeled as graph-structured data, including nodes and edges of the fences and their surrounding environment. The labeling was performed using specialized labeling tools to ensure the accuracy and validity of the data.

[0081] Key Results: The experimental results are compared in Table 1 below:

[0082] Table 1 Experimental Results

[0083]

[0084] Results analysis:

[0085] The baseline model yolov10n.pt performed poorly on the power security inspection scenario dataset, with all indicators significantly lower than expected, demonstrating the limitations of traditional target detection models in complex environments.

[0086] In comparison, the model trained using the GCNTrainer proposed in this invention achieved significant performance improvements across all metrics. For example, mAP@50 improved from 0.0082 to 0.675, and mAP@50-95 improved from 0.0045 to 0.450, representing an improvement of over 80 times.

[0087] This result demonstrates the effectiveness of the method of the present invention, especially in complex power security inspection environments, where it can successfully transfer knowledge from a powerful teacher model to a lightweight student model. By directly utilizing the output of the teacher model for knowledge distillation, the model's performance in construction fence detection tasks is significantly enhanced, overcoming the detection difficulties of traditional methods under suboptimal imaging conditions.

[0088] The core of this invention lies in utilizing the characteristics of graph convolutional networks, combined with the specific features of construction sites, to achieve efficient and accurate detection of construction fences. Specifically, the method constructs a graph structure of the fence and its surrounding environment to capture complex spatial relationships and feature associations, thereby improving the robustness and accuracy of the detection.

[0089] Furthermore, considering the specific needs of power construction sites, this invention also takes into account detection performance under electromagnetic interference and low-light conditions. By introducing data augmentation and model optimization techniques, the model's adaptability to various environments is improved, ensuring real-time monitoring and safety management of construction fences. This method not only meets the high-precision feature recognition requirements of safety inspections but also adapts to the lightweight deployment of edge devices, promoting the advancement of safety management technology at power construction sites.

[0090] In summary, as the power industry's requirements for safety management technology continue to increase, the construction fence detection method based on graph convolutional networks provides a new solution for power safety supervision. This not only helps improve the safety of construction sites but also lays the foundation for future intelligent inspection and automated management, promoting the power industry towards a safer and more efficient direction.

[0091] This invention also provides a construction fence detection system based on graph convolutional networks, comprising:

[0092] The data collection and preprocessing module is used to collect high-definition image data from power construction sites, perform noise reduction, contrast enhancement and normalization preprocessing on the high-definition image data, and output preprocessed images.

[0093] The graph structure data construction module is used to extract features of the fence and its surrounding environment as nodes based on preprocessed images, and connect the nodes through edges that reflect spatial relationships, interactions, and feature similarities to form graph structure data.

[0094] The graph convolutional network forward propagation module is used to input graph structure data into the graph convolutional network, extract deep features of the fence and environment through multi-layer graph convolution, and output fence feature representation;

[0095] The standard detection loss calculation module is used to calculate the standard detection loss based on the fence feature representation and the ground label using the cross-entropy loss function.

[0096] The feature optimization and enhancement module is used to combine the feature information of neighboring nodes represented by the fence features to implement a feature fusion strategy and improve the robustness of fence detection.

[0097] A new loss function design module is used to design joint loss functions based on graph structure data, which weights and combines the standard detection loss with the graph structure context loss.

[0098] The model parameter update module is used to update the weight parameters of the graph convolutional network based on the joint loss function through the backpropagation mechanism, so as to obtain the trained graph convolutional detection network.

[0099] The construction fence detection module is used to input new power construction site image data into a trained graph convolutional detection network and output the status detection results of the construction fence.

[0100] This invention proposes an improved graph convolutional network detection method for construction fence detection in power security inspection scenarios. By directly calling the forward method of the teacher model in the trainer to obtain the original discriminants and using them to calculate the loss, this invention simplifies the knowledge transfer process and effectively guides the learning of inference. Experimental results show that the proposed method achieves a significant performance improvement compared to the baseline model on a power security inspection scenario dataset, demonstrating its potential to improve detection performance in complex environments. Furthermore, this performance improvement is achieved without increasing the size of the student model or the complexity of inference, providing an effective solution for deploying high-performance, robust object detection models on resource-constrained devices, especially suitable for handling scenarios with non-ideal conditions such as power construction.

[0101] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A construction fence detection method based on graph convolutional networks, characterized in that, Includes the following steps: Data collection and preprocessing: Collect high-definition image data from power construction sites, perform noise reduction, contrast enhancement, and normalization preprocessing on the high-definition image data, and output preprocessed images; Graph structure data construction: Based on the preprocessed image, features of the fence and its surrounding environment are extracted as nodes, and the nodes are connected by edges that reflect spatial relationships, interactions, and feature similarities to form graph structure data; Forward propagation of graph convolutional networks: The constructed graph structure data is input into the graph convolutional network, and deep features of the fence and environment are extracted through multiple layers of graph convolution, and the fence feature representation is output. Standard detection loss calculation: Based on the fence feature representation and the ground truth label, the standard detection loss is calculated using the cross-entropy loss function; Feature optimization and enhancement: By combining the feature information of neighboring nodes represented by fence features, a feature fusion strategy is implemented to improve the robustness of graph convolutional networks; Novel Loss Function Design: A joint loss function is designed based on graph structure data, which combines the standard detection loss with the graph structure context loss in a weighted manner. The graph structure context loss is calculated by the difference in node spatial coordinates and the deviation of edge feature similarity. Model parameter update: Based on the joint loss function, the weight parameters of the graph convolutional network are updated through the backpropagation mechanism to obtain the trained graph convolutional network; Construction fence detection: Input new power construction site image data into a trained graph convolutional network, and output the state detection results of the construction fence.

2. The method according to claim 1, characterized in that, The nodes of the graph structure data include fence components, construction tools, equipment, and background objects. The weights of the edges are determined by calculating the spatial distance and feature similarity between nodes.

3. The method according to claim 1, characterized in that, The graph convolutional network adopts a multi-layer structure, and each layer dynamically adjusts the node weights through a self-attention mechanism to enhance the expression of key features.

4. The method according to claim 1, characterized in that, The formula for calculating the joint loss function is as follows: Total loss = α × Standard detection loss + (1-α) × Graph structure context loss Where α is the weighting coefficient, and 0 < α < 1.

5. The method according to claim 1, characterized in that, The high-definition image data includes images of construction fences under complex environmental conditions, such as high light, low light, and rainy weather.

6. A construction fence detection system based on graph convolutional networks, characterized in that, include: The data collection and preprocessing module is used to collect high-definition image data from power construction sites, perform noise reduction, contrast enhancement and normalization preprocessing on the high-definition image data, and output preprocessed images. The graph structure data construction module is used to extract features of the fence and its surrounding environment as nodes based on preprocessed images, and connect the nodes through edges that reflect spatial relationships, interactions, and feature similarities to form graph structure data. The graph convolutional network forward propagation module is used to input graph structure data into the graph convolutional network, extract deep features of the fence and environment through multi-layer graph convolution, and output fence feature representation; The standard detection loss calculation module is used to calculate the standard detection loss based on the fence feature representation and the ground label using the cross-entropy loss function. The feature optimization and enhancement module is used to combine the feature information of neighboring nodes represented by the fence features to implement a feature fusion strategy and improve the robustness of fence detection. The new loss function design module is used to design joint loss functions based on graph structure data. It combines the standard detection loss with the graph structure context loss in a weighted manner. The graph structure context loss is calculated by the difference in node spatial coordinates and the deviation of edge feature similarity. The model parameter update module is used to update the weight parameters of the graph convolutional network based on the joint loss function through the backpropagation mechanism, so as to obtain the trained graph convolutional detection network. The construction fence detection module is used to input new power construction site image data into a trained graph convolutional detection network and output the status detection results of the construction fence.

7. The system according to claim 6, characterized in that, The nodes of the graph structure data include fence components, construction tools, equipment, and background objects. The weights of the edges are determined by calculating the spatial distance and feature similarity between nodes.

8. The system according to claim 6, characterized in that, The graph convolutional network adopts a multi-layer structure, and each layer dynamically adjusts the node weights through a self-attention mechanism to enhance the expression of key features.

9. The system according to claim 6, characterized in that, The formula for calculating the joint loss function is as follows: Total loss = α × Standard detection loss + (1-α) × Graph structure context loss Where α is the weighting coefficient, and 0 < α < 1.

10. The system according to claim 6, characterized in that, The high-definition image data includes images of construction fences under complex environmental conditions, such as high light, low light, and rainy weather.

Images