Substation field safety dynamic perception and safety management method and system

Abstract

The application discloses a kind of substation field safety dynamic perception and safety management and control method and system, belong to power plant field safety monitoring and early warning technical field.The method includes collecting the image data of typical maintenance field operation vehicle, personnel, fence, establishes image database;Based on image database, utilize the lightweight deep learning network of CenterNet to the safety violation behavior in real-time collected video image is dynamically identified, realizes the dynamic perception of power station field safety;Based on the dynamic perception information of power station field safety, the safety violation behavior in video image is early warned, and the accuracy of fence safety measure layout map of checking safety measure management and control platform is tracked, the state of field safety measure robot is realized, and the safety control of power station field is realized.The application realizes the dynamic perception and control of power station field safety, effectively improves the safe operation level of power station.

Description

Technical Field

[0001] This invention relates to the field of power plant on-site safety monitoring and early warning technology, specifically to a method and system for dynamic perception and safety control of substation on-site safety. Background Technology

[0002] Real-time dynamic monitoring and early warning of safe distances for vehicles, personnel, and fencing at substation sites is one of the core technologies for safety early warning at maintenance sites. Power production sites are dynamic, complex, and changeable, easily leading to unlicensed operations, operations beyond permitted scope, and unplanned operations, resulting in frequent violations. Although power grids and power companies have established comprehensive power safety production regulations, violations still occur frequently due to lax power safety management and weak safety awareness among workers. For example, in substations, when partial power outages occur, temporary fencing should be installed if the safe distance between workers and non-powered equipment does not meet the safety distance regulations for equipment without power outages. Modern power safety production demands that safety monitoring and evaluation systems provide for process monitoring, result traceability, and intelligent alarms.

[0003] In existing technologies, safety measure layout diagrams can be generated from work orders, which can then be used to create corresponding fence models in 3D twin models and plan their positions and sizes. However, due to the complex dynamic environment on-site, existing safety measure management platforms, which automatically generate safety measure layout diagrams based on 3D digital models and knowledge graphs, struggle to determine whether the placement of fences and other equipment meets requirements based solely on the model. Furthermore, the operational footage of safety measure robots at maintenance sites is difficult to view comprehensively and in real-time on the safety measure management platform's backend, making it difficult to promptly identify potential safety hazards. All these safety hazards indicate that, due to the complex environment of power plants, existing safety monitoring technologies are still unable to effectively identify safety violations. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for dynamic perception and safety management of substations, to conduct safety supervision of power plants and reduce the probability of danger to a certain extent.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a method for dynamic perception and control of safety at substation sites, comprising: Image data of vehicles, personnel, and fences at typical maintenance sites are collected, and preprocessing and data augmentation techniques are used to establish an image database; the preprocessing includes spatial domain processing, frequency domain processing, etc. Based on an image database, the lightweight deep learning network CenterNet is used to dynamically identify safety violations in real-time video images, thereby achieving dynamic perception of on-site safety at the power plant. The violations include vehicles exceeding clearance limits, on-site safety guides not being in place, incorrect fence placement, vehicles or construction workers crossing fence boundaries within the work area, failure to stand in the prescribed positions, failure to wear safety helmets and work clothes, etc. Based on dynamic perception information of power plant site safety, early warnings are issued for safety violations in video images, and the accuracy of the fence safety measure layout diagram of the safety measure management platform is verified, and the status of the on-site safety measure robot is tracked to achieve on-site safety management of the power plant.

[0006] According to the above technical solution, the lightweight deep learning network method of CenterNet includes the following steps: The collected data is augmented using methods including translation, rotation, mirroring transformation, brightness and contrast transformation, noise addition, and random cropping. The enhanced data is used to extract features through the backbone network of CenterNet, which includes ResNet and DLA networks. The features extracted by the backbone network are input into network 1, network 2 and network 3; network 1 is responsible for predicting the horizontal and vertical coordinates of the center point of the target, network 2 is responsible for predicting the offset of the horizontal and vertical coordinates of the center point, and network 3 is responsible for the length and width of the bounding box centered on the center point. The target at the current center point is calculated using regression analysis of the output data from Network 1, Network 2, and Network 3, thus completing target identification. The method of calculating the target at the current center point using regression analysis is existing technology and will not be described in detail here.

[0007] This invention employs Mosaic data augmentation and the Mixup method to enhance data. The Mosaic method is used to augment scarce background data, accelerating training and convergence. The Mixup method, which models different categories, effectively increases the model's generalization ability. Furthermore, CenterNet's construction adopts an anchor-free approach, eliminating the need for candidate boxes as precursors for detection and post-processing.

[0008] According to the above technical solution, the convolutional layer of the DLA network includes a ResNet network. The DLA network has several advantages, which are reflected in: (1) the addition of a feature fusion mechanism in the network structure to internally fuse low-order, mid-order, and high-order features, resulting in strong robustness. (2) the realization of multi-level feature fusion through a tree structure to achieve the fusion of multi-level features of different depths by using different residual units. (3) the implementation of an IDA structure for iterative feature fusion.

[0009] According to the above technical solution, the ResNet network uses the GSC attention mechanism to focus on the foreground part of the digital image; The GSC attention mechanism uses the NL (Non-Local) module to extract global channel features of the image. The generated feature map is then input into the SE (Squeeze and Excitation) module. The output feature map is then superimposed point by point with the original input to obtain the final output feature map. Channel attention can learn the correlation between channels in the feature map, achieving better channel feature focusing. The spatial attention module is connected to the channel attention mechanism, improving the model's accuracy by focusing on effective information in the feature map in a multi-dimensional and multi-level manner.

[0010] The SE (Squeeze and Excitation) module includes a GMP (Global Mean Pooling) module and a GAP (Global Feature Pooling) module. The feature map output from the NL (Non-Local) module is simultaneously input into both GMP and GAP for dimensionality compression, resulting in a feature map combining global average pooling and global feature pooling. Then, an FC (Full Functional) layer maps the feature map to spatial dimensions, forming a spatial attention mechanism that interacts with the original image to obtain the final features. The GSC (Global Squeeze and Excitation) attention mechanism can perform global focusing in the channel direction and pixel-level focusing in the spatial direction, deeply mining multi-dimensional and multi-level information to identify and focus on the features that contribute most to the model.

[0011] Since a network model based on a multi-dimensional and multi-layered attention mechanism is generated by the PyTorch framework and cannot be directly converted into a TensorRT model, the PyTorch model is first converted into the Microsoft ONNX (Open Neural Network Exchange) open neural network exchange, and then the ONNX model is converted into a TensorRT model, which can accelerate the inference of the model on the embedded device.

[0012] According to the above technical solution, CenterNet's lightweight deep learning network is used to collect the three-dimensional position information of on-site work vehicles, personnel and fences in real time to obtain real-time three-dimensional point cloud data of the on-site physical space; the real-time three-dimensional point cloud data is integrated into the three-dimensional model of the substation, and based on the on-site safety ticket information, a three-dimensional spatial safety distance collision algorithm is adopted. When the distance between the two is less than the specified safety distance, an audible and visual alarm is sent to the work site and a warning is issued to the background safety platform. CenterNet's lightweight deep learning network is used to collect real-time data on the behavior of vehicles and personnel at the work site and match it with violations in the image database. If a violation occurs, an audible and visual alarm is triggered at the work site and a warning is issued to the back-end safety platform.

[0013] Based on the above technical solution, verifying the accuracy of the fence safety measure layout diagram of the safety measure control platform includes: Based on the work order, the safety measures work tasks are defined by the work scope. The substation on-site camera equipment is automatically linked to collect images. The image stitching and fusion technology is used to generate a panoramic image of the fence. Based on the photos collected on-site, the CenterNet lightweight deep learning network is used to obtain basic data of the substation on-site for physical modeling. A 3D engine is then used for secondary rendering and optimization. The fence layout model generated by video image rendering is substituted into the layout model automatically generated by the safety management platform. The data matching algorithm is used to calculate and determine whether the fence meets the design requirements after being introduced into the real scene. Based on the actual site conditions, the generated safety measure layout diagram was checked, adjusted, and optimized, and then written into the safety measure layout diagram database; The optimized safety measure layout results are fed back to the safety measure management platform to form the basic data and training samples for secondary fine-tuning, thereby realizing intelligent iterative optimization of the safety measure layout and guiding the generation of the next safety measure layout.

[0014] According to the above technical solution, on-site safety monitoring robot status tracking includes: Real-time image acquisition of the Ancuo robot is performed. The lightweight deep learning network of CenterNet is used to identify the status of the Ancuo robot and match it with the Ancuo robot's violations in the image database. If the Ancuo robot violates the rules, an audio-visual safety regulation warning is issued. The violations of the Ancuo robot are stored in sequence according to time nodes to achieve full traceability and automatic archiving of safety behaviors. Real-time image acquisition is performed on the Ancuo robot. The lightweight deep learning network of CenterNet is used to identify the position of the Ancuo robot. The three-dimensional spatial sensing data overlay technology is used to display the safe collision avoidance status between the Ancuo robot and the substation equipment, the robot's own fault / limit state, and the real-time spatiotemporal trajectory dynamics online, so as to realize the real-time safety behavior analysis and early warning of the robot.

[0015] The substation on-site safety dynamic perception and safety control system includes: The database creation module collects image data of vehicles, personnel, and fences at typical maintenance sites, and uses preprocessing and data augmentation techniques to establish an image database. The dynamic perception module, based on an image database, uses CenterNet's lightweight deep learning network to dynamically identify safety violations in video images in real time, thereby achieving dynamic perception of on-site safety at the power plant. The safety management module, based on dynamic perception information of power plant site safety, provides early warnings for safety violations in video images, verifies the accuracy of the fence safety measure layout diagram of the safety measure management platform, and tracks the status of the on-site safety measure robot to achieve on-site safety management of the power plant.

[0016] Compared with the prior art, the beneficial effects achieved by the present invention are: 1. This invention utilizes CenterNet's lightweight deep learning network to monitor the behavior of on-site vehicles and personnel in real time, and can dynamically identify safety violations, thereby improving the efficiency of on-site safety management and providing a rich data foundation for subsequent safety control.

[0017] 2. This invention employs the GSC attention mechanism, which can focus attention on both the spatial and channel network layers, as well as the compatibility between different channels, effectively capturing target information and better distinguishing between foreground and background. Furthermore, it introduces efficient data augmentation methods, augmenting data from multiple angles, including the data source itself, background enhancement, and data mixing, effectively increasing data diversity and making model learning more robust. This achieves focus on the foreground portion of digital images, improving the accuracy and efficiency of image processing.

[0018] 3. This invention utilizes CenterNet's lightweight deep learning network to perform real-time identification and analysis of image data of on-site operating vehicles, personnel, and fences, providing early warnings for safety violations, verifying the accuracy of the fence safety measure layout diagram on the safety measure management platform, and tracking the status of on-site safety measure robots. This enables dynamic perception and control of on-site safety at power plants, effectively improving the safety operation level of power plants. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the substation on-site safety dynamic perception and safety control method of the present invention; Figure 2 This is a schematic diagram of the CenterNet network structure of the present invention; Figure 3 This is a schematic diagram of the DLA backbone network structure of the present invention; Figure 4 This is a schematic diagram of the GSC attention mechanism structure of the present invention; Figure 5 This is a schematic diagram illustrating the verification and optimization of the safety measures layout in this embodiment; Figure 6 This is a safety status monitoring diagram of the Ancuo robot in this embodiment. Detailed Implementation

[0020] 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, and 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.

[0021] like Figure 1-6 As shown, the substation on-site safety dynamic perception and safety control method of the present invention includes the following steps: S1. Collect image data of vehicles, personnel, and fences at typical maintenance sites, and establish an image database using preprocessing and data augmentation techniques; S2. Based on an image database, utilize a lightweight deep learning network, CenterNet. Figure 2 Dynamically identify safety violations in real-time video images to achieve dynamic perception of on-site safety at the power plant; The lightweight deep learning network method of CenterNet includes the following steps: The collected data is augmented using methods including translation, rotation, mirroring transformation, brightness and contrast transformation, noise addition, and random cropping. The enhanced data is processed through the CenterNet backbone network to extract features. The CenterNet backbone network includes a ResNet network and a DLA network. The convolutional layers of the DLA network include ResNet networks. The DLA backbone network structure is as follows: Figure 3 As shown; The features extracted by the backbone network are input into network 1, network 2 and network 3; network 1 is responsible for predicting the horizontal and vertical coordinates of the center point of the target, network 2 is responsible for predicting the offset of the horizontal and vertical coordinates of the center point, and network 3 is responsible for the length and width of the bounding box centered on the center point. The target at the current center point is calculated by using regression analysis of the output data of network 1, network 2 and network 3, thus completing the target identification.

[0022] The ResNet network employs the GSC attention mechanism to focus on the foreground portion of digital images. The GSC attention mechanism uses the NL (Non-Local) module to extract global channel features of the image. The generated feature map is then input into the SE (Squeeze and Excitation) module, and the output feature map is superimposed point by point with the original input to obtain the final output feature map. The SE (Squeeze and Excitation) module includes a GMP module and a GAP module. The feature map output by the NL (Non-Local) module is simultaneously input into GMP and GAP for dimensionality compression, resulting in a feature map combining global average pooling and global feature pooling. Then, the FC layer is used to map the feature map to the spatial dimension, forming a spatial attention mechanism to perform point-to-point conversion with the original image, thus obtaining the final feature.

[0023] The network structure of the GSC attention mechanism is as follows: Figure 4 As shown, in the network structure of the GSC attention mechanism This represents matrix multiplication, and also represents the broadcast multiplication of a matrix and a vector (the vector is expanded to the same dimension as the matrix and then multiplied point by point). This indicates matrix broadcast addition (vectors are expanded to the same dimension as the matrix and then multiplied point by point), GMP indicates global max pooling, GAP indicates global average pooling, and FC indicates fully connected layer.

[0024] S3. Based on dynamic perception information of power plant site safety, it provides early warning of safety violations in video images, verifies the accuracy of the fence safety measure layout diagram of the safety measure management platform, tracks the status of the on-site safety measure robot, and realizes on-site safety management of the power plant.

[0025] The safety violation warning mechanism includes: using CenterNet's lightweight deep learning network to collect real-time 3D position information of on-site work vehicles, personnel, and fences to obtain real-time 3D point cloud data of the on-site physical space; integrating the real-time 3D point cloud data into the substation's 3D model; and using a 3D spatial safety distance collision algorithm based on on-site safety ticket information. When the distance between the two is less than the specified safety distance, an audible and visual alarm is sent to the work site, and a warning is issued to the background safety platform. CenterNet's lightweight deep learning network is used to collect real-time data on the behavior of vehicles and personnel at the work site and match it with violations in the image database. If a violation occurs, an audible and visual alarm is triggered at the work site and a warning is issued to the back-end safety platform.

[0026] Verifying the accuracy of the safety measure control platform's fence layout diagram includes: Based on the work order, the safety measures work tasks are defined within the work area. The system automatically links the substation's on-site camera equipment to acquire images and uses image stitching and fusion technology to generate panoramic images around the fence. Based on the photos acquired on-site, the system uses CenterNet's lightweight deep learning network to obtain basic data of the substation site for physical modeling and then uses a 3D engine for secondary rendering and optimization. The image stitching and fusion technology mentioned is existing technology and will not be described in detail here. The fence layout model generated by video image rendering is substituted into the layout model automatically generated by the safety management platform. The data matching algorithm is used to calculate and determine whether the fence meets the design requirements after being introduced into the real scene. Based on the actual site conditions, the generated safety measure layout diagram was checked, adjusted, and optimized, and then written into the safety measure layout diagram database; The optimized safety measure layout results are fed back to the safety measure management platform to form the basic data and training samples for secondary fine-tuning, thereby realizing intelligent iterative optimization of the safety measure layout and guiding the generation of the next safety measure layout.

[0027] On-site safety measures robot status tracking includes: Real-time image acquisition of the Ancuo robot is performed. The lightweight deep learning network of CenterNet is used to identify the status of the Ancuo robot and match it with the Ancuo robot's violations in the image database. If the Ancuo robot violates the rules, an audio-visual safety regulation warning is issued. The violations of the Ancuo robot are stored in sequence according to time nodes to achieve full traceability and automatic archiving of safety behaviors. Real-time image acquisition is performed on the Ancuo robot. The robot's position is identified using a lightweight deep learning network based on CenterNet. Three-dimensional spatial sensing data overlay technology is employed to display online the safe collision avoidance status between the Ancuo robot and substation equipment, as well as the robot's own fault / limit states (e.g., ...). Figure 6 (and real-time spatiotemporal trajectory dynamics) enable real-time safety behavior analysis and early warning for robots.

[0028] Another embodiment includes a substation on-site safety dynamic perception and safety control system, comprising: The database creation module collects image data of vehicles, personnel, and fences at typical maintenance sites, and uses preprocessing and data augmentation techniques to establish an image database. The dynamic perception module, based on an image database, uses CenterNet's lightweight deep learning network to dynamically identify safety violations in video images in real time, thereby achieving dynamic perception of on-site safety at the power plant. The safety management module, based on dynamic perception information of power plant site safety, provides early warnings for safety violations in video images, verifies the accuracy of the fence safety measure layout diagram of the safety measure management platform, and tracks the status of the on-site safety measure robot to achieve on-site safety management of the power plant.

[0029] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0030] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for dynamic safety perception and safety control at substation sites, characterized in that, include: Collect video image data of vehicles, personnel, and fences at typical maintenance sites, and establish an image database; Based on an image database, the lightweight deep learning network CenterNet is used to dynamically identify safety violations in real-time acquired video images, thereby achieving dynamic perception of on-site safety at the power plant. Based on dynamic perception information of power plant site safety, early warnings are issued for safety violations in video images, and the accuracy of the fence safety measure layout diagram of the safety measure management platform is verified, and the status of the on-site safety measure robot is tracked to achieve on-site safety management of the power plant. Warnings for safety violations include: CenterNet's lightweight deep learning network is used to collect the three-dimensional position information of on-site work vehicles, personnel and fences in real time, and obtain real-time three-dimensional point cloud data of the on-site physical space. The real-time three-dimensional point cloud data is integrated into the three-dimensional model of the substation, and based on the on-site safety ticket information, a three-dimensional spatial safety distance collision algorithm is adopted. When the distance between the two is less than the specified safety distance, an audible and visual alarm is sent to the work site and a warning is issued to the background safety platform. CenterNet's lightweight deep learning network is used to collect real-time data on the behavior of vehicles and personnel at the work site and match it with violations in the image database. If a violation occurs, an audible and visual alarm is triggered at the work site and a warning is issued to the back-end safety platform. Verifying the accuracy of the safety measure control platform's fence layout diagram includes: Based on the work order, the safety measures work tasks are defined by the work scope. The substation on-site camera equipment is automatically linked to collect images. The image stitching and fusion technology is used to generate a panoramic image of the fence. Based on the photos collected on-site, the CenterNet lightweight deep learning network is used to obtain basic data of the substation on-site for physical modeling. A 3D engine is then used for secondary rendering and optimization. The fence layout model generated by video image rendering is substituted into the layout model automatically generated by the safety management platform. The data matching algorithm is used to calculate and determine whether the fence meets the design requirements after being introduced into the real scene. Based on the actual site conditions, the generated safety measure layout diagram was checked, adjusted, and optimized, and then written into the safety measure layout diagram database; The optimized safety measure layout results are fed back to the safety measure management platform to form the basic data and training samples for secondary fine-tuning, thereby realizing intelligent iterative optimization of the safety measure layout and guiding the generation of the next safety measure layout. On-site safety measures robot status tracking includes: Real-time image acquisition of the Ancuo robot is performed, and the lightweight deep learning network of CenterNet is used to identify the status of the Ancuo robot. The status is matched with the Ancuo robot's violations in the image database. If the Ancuo robot violates the rules, an audio-visual safety regulation warning is issued. The violations of the Ancuo robot are stored in chronological order according to time nodes, so as to realize the full-process traceability and automatic archiving of safety behaviors.

2. The substation on-site safety dynamic perception and safety control method according to claim 1, characterized in that, The method steps for object recognition using the lightweight deep learning network CenterNet include: Enhance the collected data; The enhanced data is used to extract features through the backbone network of CenterNet; The features extracted by the backbone network are input into network 1, network 2 and network 3. Network 1 is responsible for predicting the horizontal and vertical coordinates of the center point of the target. Network 2 is responsible for predicting the offset of the horizontal and vertical coordinates of the center point. Network 3 is responsible for the length and width of the bounding box centered on the center point. The target at the current center point is calculated by using regression analysis of the output data of network 1, network 2 and network 3, thus completing the target identification.

3. The substation on-site safety dynamic perception and safety control method according to claim 2, characterized in that, Data augmentation methods include: translation, rotation, mirroring transformation, brightness and contrast transformation, adding noise, and random cropping.

4. The substation on-site safety dynamic perception and safety control method according to claim 2, characterized in that, The backbone network of CenterNet includes the ResNet network and the DLA network.

5. The substation on-site safety dynamic perception and safety control method according to claim 4, characterized in that, The convolutional layers of the DLA network include a ResNet network.

6. The substation on-site safety dynamic perception and safety control method according to claim 5, characterized in that, The ResNet network employs the GSC attention mechanism to focus on the foreground portion of the digital image; The GSC attention mechanism uses the NL module to extract global channel features of the image. The generated feature map is then input into the SE module. The output feature map is then superimposed point by point with the original input to obtain the final output feature map. The SE module includes a GMP module and a GAP module. The feature map output by the NL module is simultaneously input into GMP and GAP for dimensionality compression to obtain a feature map combining global average pooling and global feature pooling. Then, the FC layer is used to map the feature map to the spatial dimension, forming a spatial attention mechanism to perform point-to-point conversion with the original image to obtain the final feature.

7. The substation on-site safety dynamic perception and safety control method according to claim 1, characterized in that, Also includes: CenterNet's lightweight deep learning network is used to identify the position of the Ancuo robot. Three-dimensional spatial sensing data overlay technology is used to display the safe collision avoidance status between the Ancuo robot and substation equipment, the robot's own fault / limit state, and the real-time spatiotemporal trajectory dynamics online, so as to realize the real-time safety behavior analysis and early warning of the robot.

8. A substation on-site safety dynamic perception and safety control system, characterized in that, include: The database creation module collects image data of vehicles, personnel, and fences at typical maintenance sites and establishes an image database. The dynamic perception module, based on an image database, uses CenterNet's lightweight deep learning network to dynamically identify safety violations in video images in real time, thereby achieving dynamic perception of on-site safety at the power plant. The safety management module, based on dynamic perception information of power plant site safety, provides early warnings for safety violations in video images, verifies the accuracy of the fence safety measure layout diagram of the safety measure management platform, and tracks the status of the on-site safety measure robot to achieve on-site safety management of the power plant.

Images