A substation non-contact non-inductive power transformation operation risk identification method

Abstract

The application provides a substation non-contact non-inductive power transformation operation risk identification method, based on sensors such as laser radars and panoramic cameras, a high-precision three-dimensional point cloud model of a power transformation site is constructed; then a two-dimensional image semantic segmentation technology is adopted to complete the division of dangerous areas in the visual space, and a three-dimensional world space coordinate is established, and the operation personnel are non-inductively positioned through visible light imaging and three-dimensional remote sensing imaging; then a substation operation personnel safety risk identification model is constructed through a yolo v5 network structure, and real-time identification of the operation site is realized based on terminal equipment such as on-site cameras and control balls; finally, the on-site safety risk notification and early warning are realized through on-site playing.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method for identifying risks in non-contact, sensorless substation operations. Background Technology

[0002] Currently, real-time monitoring at substation work sites still faces many challenges in terms of both promotion costs and monitoring accuracy. The main problems are as follows:

[0003] (1) The cost of personnel positioning at the work site is relatively high. At present, personnel positioning is mainly achieved through technologies such as satellite differential, UWB and radar. However, due to the need to set up base stations or purchase radar equipment at the work site, and the need to carry out equipment power supply modification construction, the cost restricts the large-scale promotion and application of such solutions. Positioning equipment that relies on battery power supply is difficult to support long-term on-site operations due to its limited battery life. Furthermore, the reliability and stability of positioning systems that rely on external hardware equipment are difficult to guarantee.

[0004] (2) The acceptance of traditional positioning technology by on-site personnel is low. Currently, within the State Grid, positioning technology requires construction workers to wear devices such as positioning chips, for example, smart safety helmets, which are heavy and increase the burden on on-site workers. Issues such as radiation and battery life also affect the acceptance of front-line personnel, making it difficult to promote and use on a large scale.

[0005] (3) The accuracy of on-site operation risk identification is low. The power grid operation site relies on traditional pure video image recognition methods and generally uses two-dimensional planar monitoring probes for identification. Although it can achieve a certain degree of personnel operation safety supervision, it lacks a high-precision on-site operation environment model and three-dimensional spatial location information, and cannot identify the safe operation distance in time, resulting in a lack of hazard identification ability. Summary of the Invention

[0006] To address the shortcomings of the existing technology, the present invention aims to provide a non-contact, sensorless risk identification method for substation operations. First, a high-precision three-dimensional point cloud model of the substation operation site is constructed. Then, the dangerous areas within the visual space are divided and three-dimensional world space coordinates are established. Next, a real-time intelligent safety risk identification model for the substation operation site is constructed using a YOLO v5 network structure, thereby achieving real-time identification of the operation site and on-site safety risk notification and early warning.

[0007] This invention adopts the following technical solution: a non-contact, sensorless substation operation risk identification method based on cross-modal fusion, comprising the following steps:

[0008] Step 1: Based on LiDAR, obtain the 3D point cloud depth information of the substation; based on panoramic camera, acquire optical image data of the substation from different positions and angles to form a 3D model of the substation; using the ground as a reference point, jointly calibrate the 3D point cloud data and the 3D model of the substation, and fuse them to construct a high-precision 3D model of the substation.

[0009] Step 2: Using the coordinates of the panoramic camera worn by the operator as the operator, the spatial position is transformed based on the panoramic camera image and the high-precision 3D model of the substation to achieve seamless positioning of the operator.

[0010] Step 3: For substation operation scenarios, construct a substation operation safety risk sample library; and use a YOLO v5 network to construct a substation operator safety risk identification model.

[0011] Step 4: Compress and quantize the safety risk identification model for substation workers, and deploy edge computing devices at the work site; connect panoramic cameras and surveillance cameras to the edge computing devices, input field data into the edge computing devices, and send the data into the safety risk identification model for substation workers to identify personnel safety behaviors, achieve real-time identification at the work site, and output safety risk identification results. If there are safety risk behaviors, transmit the risk identification results to terminal devices such as surveillance cameras at the site, and prompt the risk results through voice broadcast to achieve real-time notification and early warning.

[0012] In step 2 of this invention, image semantic segmentation technology is used to divide the high-precision three-dimensional model of the substation into multiple regions, including working areas, safe areas and dangerous areas, and to divide the regions from a visual spatial perspective.

[0013] In step 2 of this invention, based on the position and rotation angle of the panoramic camera, all high-precision 3D models of substations within the current viewpoint are loaded, and the high-precision 3D models of substations are rasterized. Using the camera's intrinsic parameters and its position information in the high-precision 3D models of substations, the rasterized points are projected onto the camera sensor plane to obtain the 3D coordinates of each pixel in the 2D image of the camera corresponding to the high-precision 3D model of the substation. The obtained image pixel 3D coordinate mapping is then saved.

[0014] In step 2 of this invention, the unit direction vector of the center pixel relative to the camera's optical center is obtained using the camera's intrinsic parameters and the two-dimensional coordinates of the operator; the unit direction vector of the operator's center in the global coordinate system of the high-precision three-dimensional model of the substation is obtained using the camera's extrinsic parameters; the angle between the obtained camera-operator center vector and the ground is obtained using the obtained camera-operator center vector; the height difference between the camera and the operator's center is obtained using the camera pose, the nearby ground height coordinates, and the preset average height of the human body; the absolute distance from the camera to the operator's center is obtained using the height difference and the ground angle, thus obtaining the actual vector from the camera to the operator's center; the three-dimensional coordinates of the operator's center can be calculated using the actual vector from the camera to the operator's center and the camera pose; the Z-axis value of this three-dimensional coordinate is set as the ground height coordinate value near the actual space, thus obtaining the final three-dimensional coordinate position of the operator and the spatial position relationship of the substation equipment.

[0015] In step 3 of this invention, the substation operation safety risk sample library includes images of workers in the substation, images of power equipment, images of personnel crossing the line and entering, and images of personnel climbing.

[0016] In step 4 of this invention, the compression and quantization process is as follows: For the substation operator safety risk identification model formed through training, firstly, the model is pruned using a network pruning strategy to transform the original network into a sparse network, achieving initial compression; secondly, K-Means++ clustering is used to obtain the cluster centers of the weights of each layer of the deep network, and the cluster center values ​​are used to represent the original weight values ​​of the network to achieve weight sharing and reduce the number of weights; finally, the weights of each layer of the network are quantized, and the cluster centers are updated through retraining. The quantization process reduces the number of bits used to represent the weights, thus achieving the final compression of the deep network.

[0017] Compared with the prior art, the present invention has the following technical effects:

[0018] This invention uses lidar and panoramic cameras to construct a high-precision three-dimensional point cloud model of the substation operation site. It also divides the dangerous areas within the visual space and establishes three-dimensional world space coordinates. Through the substation operator safety risk identification model, it realizes real-time identification of the operation site and provides on-site safety risk notification and early warning. Attached Figure Description

[0019] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0020] Figure 1 This is a schematic diagram of the overall process of the method in this invention;

[0021] Figure 2Flowchart for building a high-precision 3D model of a substation based on lidar and panoramic cameras;

[0022] Figure 3 For seamless location tracking of substation workers;

[0023] Figure 4 The relevant knowledge contained in the power image knowledge graph;

[0024] Figure 5 This is a diagram illustrating various image annotation methods. Specific implementation methods

[0025] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0026] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0027] like Figure 1 As shown in the figure, this embodiment discloses a method for risk identification in non-contact, sensorless substation operations, including the following steps:

[0028] Step 1: Based on LiDAR, obtain the 3D point cloud depth information of the substation; based on a panoramic camera, acquire optical image data of the substation from different positions and angles to form a 3D model of the substation; using the ground as a reference point, jointly calibrate the 3D point cloud data and the 3D model of the substation, and fuse them to construct a high-precision 3D model of the substation, such as... Figure 2 As shown.

[0029] Step 2: Using the coordinates of the panoramic camera worn by the operator as the operator, the spatial position is transformed based on the panoramic camera image and the high-precision 3D model of the substation to achieve seamless positioning of the operator.

[0030] Image semantic segmentation technology is used to divide the high-precision 3D model of the substation into multiple regions, including working areas, safe areas and dangerous areas, thus dividing the regions from a visual spatial perspective.

[0031] Based on the panoramic camera's position and rotation angle, load all high-precision 3D models of substations within the current viewpoint, and rasterize the high-precision 3D models of substations. Using the camera's intrinsic parameters and its position information in the high-precision 3D models of substations, project the rasterized points onto the camera sensor plane to obtain the 3D coordinates of each pixel in the 2D image of the camera corresponding to the high-precision 3D model of the substation. The resulting image pixel 3D coordinate mapping is then saved.

[0032] The unit direction vector of the center pixel relative to the camera's optical center is obtained using the camera's intrinsic parameters and the two-dimensional coordinates of the operator; the unit direction vector of the operator's center in the global coordinate system of the high-precision three-dimensional model of the substation is obtained using the camera's extrinsic parameters.

[0033] The angle between the obtained camera-person center vector and the ground is obtained. The height difference between the camera and the person center is obtained using the camera pose, the height coordinates of the nearby ground and the preset average height of the human body.

[0034] The absolute distance from the camera to the center of the people is obtained by using the height difference and the ground angle, and then the actual vector from the camera to the center of the people is obtained.

[0035] Using the actual vector from the camera to the center of the personnel and the camera's pose, the three-dimensional coordinates of the personnel center can be calculated. Setting the Z-axis value of these three-dimensional coordinates as the ground elevation coordinates near the actual location gives the final three-dimensional coordinate position of the personnel and the spatial relationship between the personnel and the power equipment. Figure 3 As shown.

[0036] Step 3: For substation operation scenarios, construct a substation operation safety risk sample library; and use a YOLO v5 network to construct a substation operator safety risk identification model.

[0037] For substation operation scenarios, a substation operation safety risk sample library is constructed. This library includes images of personnel working inside the substation, images of power equipment, images of personnel crossing lines and intruding, and images of personnel climbing. Simultaneously, a YOLO v5 network architecture is constructed. Based on the substation operation safety risk sample library, the YOLO v5 network is trained to obtain a substation operation personnel safety risk identification model. This model can identify the behavior of personnel, equipment, and other elements in the images.

[0038] Step 4: Compress and quantize the substation worker safety risk identification model and deploy edge computing devices at the work site; connect panoramic cameras, surveillance cameras, and other terminal devices to the edge computing devices, input field data into the edge computing devices, and send the data to the substation worker safety risk identification model for personnel safety behavior identification. This enables real-time identification at the work site and outputs safety risk identification results. If a safety risk behavior is found, the risk identification result is transmitted to the on-site surveillance cameras and other terminal devices, and the risk result is announced via voice broadcast, achieving real-time notification and early warning, and improving the ability to manage safe operations. Figure 4 As shown.

[0039] Secondly, the substation worker safety risk identification model is compressed and quantized to facilitate deployment in edge devices at the work site. For the trained substation worker safety risk identification model, firstly, a network pruning strategy is used to prune the model after normal training, transforming the original network into a sparse network, achieving initial compression. Then, K-Means++ clustering is used to obtain the cluster centers of each layer's weights in the deep network. These cluster center values ​​represent the original weight values ​​of the network, achieving weight sharing and reducing the number of weights. Finally, the weights of each layer's clusters are quantized, and retraining is performed to update the cluster centers. The quantization process reduces the number of bits used to represent the weights, thus achieving the final compression of the deep network.

[0040] After the substation operator safety risk identification model is converted into OM file format, it is uploaded to the edge computing device. Taking ATLAS as an example, but not limited to it, the substation operator safety risk identification model is transmitted to the system HOST end, and then the substation operator safety risk identification model file is transmitted to the DEVICE end through the Matrix framework.

[0041] The above description is merely the preferred embodiment of the present invention and is not intended to limit other embodiments of the invention. Those skilled in the art can modify the above disclosure to create equivalent embodiments. However, any simple modifications, substitutions, and simplifications of the above embodiments without departing from the scope of the present invention should be within the protection scope of the present invention.

Claims

1. A method for risk identification in non-contact, sensorless substation operations, characterized in that, Includes the following steps: Step 1: Based on LiDAR, obtain the 3D point cloud depth information of the substation; based on panoramic camera, acquire optical image data of the substation from different positions and angles to form a 3D model of the substation; using the ground as a reference point, jointly calibrate the 3D point cloud data and the 3D model of the substation, and fuse them to construct a high-precision 3D model of the substation. Step 2: Using the coordinates of the panoramic camera worn by the operator as the operator, the spatial position is transformed based on the panoramic camera image and the high-precision 3D model of the substation to achieve seamless positioning of the operator. Step 3: For substation operation scenarios, construct a substation operation safety risk sample library; and use a YOLO v5 network to construct a substation operator safety risk identification model. Step 4: Compress and quantize the safety risk identification model for substation workers, and deploy edge computing devices at the work site; connect panoramic cameras and surveillance cameras to the edge computing devices, input field data into the edge computing devices, and send the data into the safety risk identification model for substation workers to identify personnel safety behaviors, achieve real-time identification at the work site, output safety risk identification results, and if there are safety risk behaviors, transmit the risk identification results to the terminal device, and prompt the risk results through voice broadcast to achieve real-time notification and early warning.

2. The method for risk identification in non-contact, sensorless substation operations according to claim 1, characterized in that, In step 2, image semantic segmentation technology is used to divide the high-precision 3D model of the substation into multiple regions, including the working area, the safe area and the dangerous area, and to divide the regions from a visual spatial perspective.

3. The method for risk identification in non-contact, sensorless substation operations according to claim 2, characterized in that, In step 2, based on the panoramic camera's position and rotation angle, all high-precision 3D models of substations within the current viewpoint are loaded, and the high-precision 3D models of substations are rasterized. Using the camera's intrinsic parameters and its position information in the high-precision 3D models of substations, the rasterized points are projected onto the camera sensor plane to obtain the 3D coordinates of each pixel in the 2D image of the camera corresponding to the high-precision 3D model of the substation. The obtained image pixel 3D coordinate mapping is then saved.

4. The risk identification method for non-contact, sensorless substation operations according to claim 3, characterized in that, In step 2, the unit direction vector of the center pixel relative to the camera's optical center is obtained using the camera's intrinsic parameters and the operator's two-dimensional coordinates. The unit direction vector of the operator's center in the global coordinate system of the high-precision three-dimensional model of the substation is obtained using the camera's extrinsic parameters. The angle between the obtained camera-operator center vector and the ground is obtained using the obtained camera pose, nearby ground height coordinates, and the preset average height of the human body. The height difference between the camera and the operator's center is obtained using the height difference and the ground angle. Thus, the actual vector from the camera to the operator's center is obtained. The three-dimensional coordinates of the operator's center are calculated using the actual vector from the camera to the operator's center and the camera pose. The Z-axis value of this three-dimensional coordinate is set as the ground height coordinate value near the actual space. Thus, the final three-dimensional coordinate position of the operator and the spatial position relationship of the substation equipment are obtained.

5. The method for risk identification in non-contact, sensorless substation operations according to claim 4, characterized in that, In step 3, the substation operation safety risk sample library includes images of workers inside the substation, images of power equipment, images of people crossing the line and entering, and images of people climbing.

6. The risk identification method for non-contact, sensorless substation operations according to claim 5, characterized in that, In step 4, the compression and quantization process is as follows: For the substation operator safety risk identification model formed by training, firstly, the substation operator safety risk identification model after normal training is pruned through a network pruning strategy to transform the original network into a sparse network, thus achieving initial compression; then, the cluster centers of the weights of each layer of the deep network are obtained through K-Means++ clustering, and the cluster center values ​​are used to represent the original weight values ​​of the network to achieve weight sharing and reduce the number of weights; finally, the weights of each layer of the network are quantized, and the cluster centers are updated through retraining.

Images