Power system substation open space floating object identification method, device and equipment
Patent Information
- Application Number
- CN202610658218.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-13
- Publication Date
- 2026-10-16
AI Technical Summary
[0005]本发明实施例提供了一种电力系统变电站开阔空间的漂浮物识别方法、装置及设备，以解决现有技术存在的漂浮物识别滞后、识别精度、环境适应性差以及预警不精准的问题
[0017] This invention provides a method, apparatus, and equipment for identifying floating objects in open spaces of power system substations. By acquiring images of the target area in the open space of the substation around the clock, it overcomes the coverage blind spots and time limitations of manual inspections and traditional monitoring. Preprocessing the acquired multimodal visual images yields target images, significantly improving image quality in complex environments such as nighttime, fog, rain, snow, and strong sunlight, providing a stable and reliable data foundation for subsequent high-precision identification and ensuring monitoring effectiveness in harsh scenarios. The target multimodal image is input into a pre-trained floating object detection model to obtain initial floating object detection results containing multiple initial detection boxes. The AI visual model automatically identifies the objects, replacing manual inspection and analysis, greatly improving detection response speed and processing efficiency, and enabling real-time capture of small-sized, long-distance floating objects, fundamentally solving the problems of lag and high false negative rates in manual inspections. During the iterative training of the floating object detection model, minimizing joint... The parameters of the floating object detection model are optimized and trained using a loss function as the objective. The joint loss function includes at least bounding box localization loss, class confidence loss, and target existence loss to adapt to the detection of flexible and deformable floating objects. Through joint constraints of multiple loss terms, the model accurately adapts to the characteristics of targets such as farmland mulch, plastic bags, and kites, which are flexible, shapeless, have blurred edges, and are easily stacked and occluded, significantly improving the localization accuracy and classification accuracy of irregular floating objects. It effectively distinguishes harmful floating objects from harmless targets such as birds, fallen leaves, and ground debris, greatly reducing the false alarm rate and false negative rate. Non-maximum suppression deduplication is performed on the initial floating object detection results to obtain the final floating object recognition results. This solves the problems of multiple bounding boxes for one object, duplicate detection, and redundant labeling caused by the flexible deformation and varying scale of floating objects. It retains the effective detection boxes with the highest confidence and the most accurate location, outputting pure, unique, and reliable recognition results, avoiding duplicate warnings and interference, and improving the accuracy and efficiency of subsequent warnings and handling.
Smart Images
Figure CN122929248A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of floating object monitoring technology, and in particular to a method, device, and equipment for identifying floating objects in open spaces of power system substations. Background Technology
[0002] As a critical hub in the power system, the operational safety of substations directly impacts grid stability and power supply reliability. In open substation environments, lightweight floating objects such as plastic film, plastic bags, and kites are easily carried by wind into equipment areas. Once these objects come into contact with high-voltage live parts, they can easily cause safety accidents such as short circuits and discharges. Substations located near farmland are particularly vulnerable to such floating object intrusions. Traditional protection methods relying on manual inspections are no longer sufficient to meet the demands of modern power grids for refined and real-time control.
[0003] Currently, floating object monitoring mainly employs three types of technologies: traditional video surveillance only completes image acquisition and playback, relying on manual inspection for anomalies; drone inspections require manual scheduling for takeoff and landing, completing site inspections according to preset routes; millimeter-wave radar combined with a vision module detects targets, identifying distant targets through radar signals and triggering alarms. All of these methods are primarily passive monitoring or post-event verification, lacking proactive early warning and intelligent analysis capabilities.
[0004] Existing technologies generally have significant limitations: manual and traditional monitoring responses are slow, making it difficult to achieve early warning; drones are limited by weather and scheduling and cannot be continuously monitored; millimeter-wave radar has insufficient sensitivity to identify ultra-thin, low-reflectivity floating objects, resulting in a high false alarm and missed alarm rate. Summary of the Invention
[0005] This invention provides a method, apparatus, and equipment for identifying floating objects in open spaces of power system substations, in order to solve the problems of lagging floating object identification, poor identification accuracy, poor environmental adaptability, and inaccurate early warning in the prior art.
[0006] In a first aspect, embodiments of the present invention provide a method for identifying floating objects in open spaces of power system substations, including: All-weather image acquisition is performed on the open space of the substation to be inspected, and the acquired multimodal visual images are preprocessed to obtain the target image; The target multimodal image is input into a pre-trained floating object detection model to obtain an initial floating object detection result containing multiple initial detection boxes. During the iterative training of the floating object detection model, the parameters of the floating object detection model are optimized and trained with the goal of minimizing the joint loss function. The joint loss function includes at least bounding box localization loss, class confidence loss, and target existence loss to adapt to the detection of flexible and deformable floating objects. The initial floating object detection results are subjected to non-maximum suppression deduplication processing to obtain the final floating object identification results.
[0007] In one possible implementation, the bounding box localization loss is based on the irregular deformation and edge blurring features of the floating object, and the bounding box loss localization of the floating object is performed by jointly calculating CIoU loss and DFL distribution focal loss.
[0008] In one possible implementation, the formula for the CIoU loss is: ; in, This represents the bounding box localization loss value. Indicates the center distance of the bounding box. Indicates the center point of the bounding box. This indicates the center point of the predicted bounding box or the actual floating object's bounding box. This represents the length of the diagonal of the minimum bounding rectangle of the bounding box. This represents the penalty value for center offset. Represents the adaptive coefficient. Indicates the consistency of the aspect ratio of the bounding box; This indicates the overlap ratio between the predicted bounding box and the actual floating object annotation box. Indicates the prediction box. This indicates a real floating object label box. This function represents the overlap ratio calculation function. This indicates the height of the actual floating object label box. This indicates the width of the output prediction box. This indicates the height of the output prediction box.
[0009] In one possible implementation, the formula for the DFL distribution focus loss is: ; in, This represents the focal loss value of the DFL distribution. This represents the left endpoint of the discrete interval of the true label. This represents the right endpoint of the discrete interval of the true label. Indicates the predicted first The predicted probability for each discrete interval. Indicates to Take the natural logarithm.
[0010] In one possible implementation, the category confidence loss is: ; in, This represents the category confidence loss value. This indicates the probability that the current area contains floating objects. Indicates the true category label, This represents a function for calculating classification error.
[0011] In one possible implementation, the target existence loss is: ; in, This represents the loss value indicating the existence of the target. Indicates the target confidence level. Indicates whether the actual target has a label or not.
[0012] In one possible implementation, performing non-maximum suppression deduplication on the initial floating object detection result to obtain the final floating object identification result includes: The bounding boxes are sorted from highest to lowest confidence level to determine the bounding boxes with the highest confidence level. Calculate the overlap ratio between the remaining bounding boxes and the bounding box with the highest confidence level; Bounding boxes with an overlap ratio greater than or equal to a preset suppression threshold are removed, while bounding boxes with an overlap ratio less than the preset suppression threshold are retained.
[0013] Secondly, embodiments of the present invention provide a floating object identification device for open spaces in power system substations, comprising: The acquisition module is used to acquire images of the open space of the substation under all-weather conditions. The preprocessing module is used to preprocess the acquired multimodal visual images to obtain the target image; The recognition module is used to input the target multimodal image into a pre-trained floating object detection model to obtain an initial floating object detection result containing multiple initial detection boxes. During the iterative training of the floating object detection model, the parameters of the floating object detection model are optimized and trained with the goal of minimizing the joint loss function. The joint loss function includes at least bounding box localization loss, class confidence loss, and target existence loss to adapt to the detection of flexible and deformable floating objects. The processing module is used to perform non-maximum suppression deduplication processing on the initial floating object detection results to obtain the final floating object identification results.
[0014] Thirdly, embodiments of the present invention provide an apparatus including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method described in the first aspect or any possible implementation thereof.
[0015] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in the first aspect or any possible implementation thereof.
[0016] Fifthly, embodiments of the present invention provide a computer program product, including a computer program that, when executed by a processor, implements the method described in the first aspect or any possible implementation thereof.
[0017] This invention provides a method, apparatus, and equipment for identifying floating objects in open spaces of power system substations. By acquiring images of the target area in the open space of the substation around the clock, it overcomes the coverage blind spots and time limitations of manual inspections and traditional monitoring. Preprocessing the acquired multimodal visual images yields target images, significantly improving image quality in complex environments such as nighttime, fog, rain, snow, and strong sunlight, providing a stable and reliable data foundation for subsequent high-precision identification and ensuring monitoring effectiveness in harsh scenarios. The target multimodal image is input into a pre-trained floating object detection model to obtain initial floating object detection results containing multiple initial detection boxes. The AI visual model automatically identifies the objects, replacing manual inspection and analysis, greatly improving detection response speed and processing efficiency, and enabling real-time capture of small-sized, long-distance floating objects, fundamentally solving the problems of lag and high false negative rates in manual inspections. During the iterative training of the floating object detection model, minimizing joint... The parameters of the floating object detection model are optimized and trained using a loss function as the objective. The joint loss function includes at least bounding box localization loss, class confidence loss, and target existence loss to adapt to the detection of flexible and deformable floating objects. Through joint constraints of multiple loss terms, the model accurately adapts to the characteristics of targets such as farmland mulch, plastic bags, and kites, which are flexible, shapeless, have blurred edges, and are easily stacked and occluded, significantly improving the localization accuracy and classification accuracy of irregular floating objects. It effectively distinguishes harmful floating objects from harmless targets such as birds, fallen leaves, and ground debris, greatly reducing the false alarm rate and false negative rate. Non-maximum suppression deduplication is performed on the initial floating object detection results to obtain the final floating object recognition results. This solves the problems of multiple bounding boxes for one object, duplicate detection, and redundant labeling caused by the flexible deformation and varying scale of floating objects. It retains the effective detection boxes with the highest confidence and the most accurate location, outputting pure, unique, and reliable recognition results, avoiding duplicate warnings and interference, and improving the accuracy and efficiency of subsequent warnings and handling. Attached Figure Description
[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0019] Figure 1 This is a flowchart illustrating the implementation of the floating object identification method in open spaces of power system substations provided in this embodiment of the invention. Figure 2 This is a schematic diagram of the structure of the floating object identification device in the open space of a power system substation provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the device provided in an embodiment of the present invention. Detailed Implementation
[0020] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0021] See Figure 1 The document illustrates a flowchart of a method for identifying floating objects in open spaces of power system substations, as provided in an embodiment of the present invention. Details are as follows: Step 101: All-weather image acquisition is performed on the open space of the substation to be detected, and the acquired multimodal visual images are preprocessed to obtain the target image.
[0022] Multimodal visual acquisition modules are deployed at pre-set locations around the substation perimeter and inside the station. These modules use a combination of three high-definition visible light cameras with motorized zoom. Each unit has a coverage radius of no less than 200 meters and achieves blind-spot-free monitoring of open spaces through three cameras with different focal lengths. The cameras have an IP67 protection rating and can work stably in a wide temperature range of -40℃ to +70℃. They can withstand rain, snow, salt spray, and strong electromagnetic interference, and are suitable for various complex outdoor scenarios, thus enabling 24 / 7 uninterrupted image acquisition of the area to be monitored.
[0023] The acquired multimodal visual images are preprocessed and optimized in multiple dimensions, including denoising, image enhancement, and distortion correction. This effectively improves the imaging quality of images under complex lighting and adverse weather conditions such as nighttime, backlight, fog, haze, rain, and snow, and eliminates image distortion, noise interference, and edge distortion to obtain clear and usable target multimodal images, providing high-quality input data for subsequent intelligent identification of floating objects.
[0024] Step 102: Input the target multimodal image into the pre-trained floating object detection model to obtain the initial floating object detection result containing multiple initial detection boxes; wherein, during the iterative training of the floating object detection model, the parameters of the floating object detection model are optimized and trained with the goal of minimizing the joint loss function. The joint loss function includes at least the bounding box localization loss, the class confidence loss and the target existence loss, in order to adapt to the detection of flexible deformable floating objects.
[0025] The target multimodal image obtained from the preprocessing in step 101 is input into the pre-trained floating object detection model, and the model inference outputs the initial floating object detection result containing multiple initial detection boxes.
[0026] Optionally, the floating object detection model is built using an improved YOLOv8 algorithm, incorporating a multi-scale feature fusion module (MSCNN) and a cross-attention mechanism (CAFM). It is trained on a dedicated dataset of 12 typical floating objects, including farmland mulch, kites, and plastic bags, which contains labeled samples. This model can accurately detect floating objects with a size ≥5cm and a distance ≤500 meters.
[0027] The training process for the improved YOLOv8 algorithm may include: I. Training data preparation.
[0028] We constructed a dedicated dataset of floating objects, including various types of targets that can easily cause power safety hazards, such as farmland mulch, ultra-thin plastic bags, kites, and strips of cloth, covering complex scenarios such as sunny days, cloudy days, nighttime, fog, haze, rain, and snow.
[0029] The samples are uniformly labeled, with the minimum size of the labeled target being greater than or equal to 5cm and the maximum detection distance being less than or equal to 500 meters, taking into account difficult samples such as small targets, long distances, and transparent / wrinkled / stacked samples.
[0030] Perform data augmentation, such as random scaling, cropping, rotating, brightness adjustment, and noise superposition of labeled samples, to improve the model's generalization ability and adaptability to harsh environments.
[0031] II. Initialization of the basic structure of the model.
[0032] The native Anchor-Free decoupled detection head of YOLOv8 is adopted, and the classification, bounding box regression and target confidence are split into three independent parallel branches. Each branch uses a dedicated convolution path, does not share parameters and calculates loss independently, so as to eliminate the mutual interference between classification and localization tasks and improve the accuracy of bounding box localization and the reliability of confidence.
[0033] Classification branch: Independently performs target category discrimination and outputs the predicted probability of the corresponding category to determine the specific type of floating object.
[0034] Bounding box regression branch: Independently completes target location localization and outputs the coordinate offset of the predicted box relative to the ground truth box, which is used to accurately determine the spatial position of the floating object.
[0035] Target confidence branch: Independently determines whether there is a valid target within the feature grid and outputs the foreground target confidence score, which is used to characterize the degree of credibility of the target's existence.
[0036] By embedding a multi-scale convolutional neural network (MSCNN) and a cross-attention fusion module (CAFM) into the backbone network, the feature extraction capabilities for small targets, weak textures, and transparent floating objects are enhanced.
[0037] III. Iterative Training and Parameter Optimization The floating object detection model aims to minimize the joint loss function. It achieves end-to-end training by continuously optimizing and updating the network parameters through backpropagation. The joint loss function includes at least bounding box localization loss, class confidence loss, and object presence loss. 1. The joint loss function is: ; Optional, This represents the value of the joint loss function. , , These represent the weight coefficients corresponding to the bounding box localization loss, class confidence loss, and target existence loss, respectively. This represents the bounding box localization loss, responsible for accurately outlining irregular floating objects. This represents the category confidence loss, responsible for distinguishing floating objects from the background. Indicates the existence of a target, used to determine whether the foreground is a real floating object.
[0038] Optionally, the default value for the weighting coefficients can be set to: =7.5, =0.5, =1.0.
[0039] 2. Training the bounding box localization loss.
[0040] After the backbone network extracts features, the detector head identifies the bounding boxes of floating objects using the captured images. The bounding boxes are represented by four parameters, namely... ,in, Describes the top-left pixel coordinates of the bounding box. This represents the pixel coordinates of the top-right corner of the bounding box.
[0041] The bounding box localization loss is based on the irregular deformation and edge blurring characteristics of floating objects. A joint calculation method using CIoU loss and DFL distribution focal point loss is employed for the bounding box localization of floating objects. , By using four constraints—overlap, center position, aspect ratio, and edge precision—floating objects of any shape can be precisely framed.
[0042] Optionally, the bounding box localization loss is fused with the IoU (Intersection over Union) loss, CIoU loss, and DFL (Distribution of Focus) focus loss, where: IoU is used to measure the overlap ratio between the predicted bounding box and the actual labeled bounding box, and it is suitable for flexible floating objects of any shape, such as wrinkles, stretching, and curling. CIoU achieves precise positioning of flexible deformation targets by using center distance, the diagonal of the minimum bounding rectangle, and adaptive aspect ratio penalty, thus avoiding positioning deviation of rapidly moving floating objects. The DFL distributed focus loss model the floating object boundary as a discrete probability distribution, achieving sub-pixel accuracy regression and solving the problem of blurred edges on transparent and wrinkled targets.
[0043] In one embodiment, the formula for CIoU loss is: ; in, This represents the bounding box localization loss value. Indicates the center distance of the bounding box. Indicates the center point of the bounding box. This indicates the center point of the predicted bounding box or the actual floating object's bounding box. This represents the length of the diagonal of the minimum bounding rectangle of the bounding box. This represents the penalty value for center offset. Represents the adaptive coefficient. Indicates the consistency of the aspect ratio of the bounding box; This indicates the overlap ratio between the predicted bounding box and the actual floating object annotation box. Indicates the prediction box. This indicates a real floating object label box. This function represents the overlap ratio calculation function. This indicates the height of the actual floating object label box. This indicates the width of the output prediction box. This indicates the height of the output prediction box; The formula for the focal loss of the DFL distribution is: ; in, This represents the focal loss value of the DFL distribution. This represents the left endpoint of the discrete interval of the true label. This represents the right endpoint of the discrete interval of the true label. Indicates the predicted first The predicted probability for each discrete interval. Indicates to Take the natural logarithm.
[0044] Optional, calculate first It measures the overlap between the predicted bounding box and the ground truth bounding box, adapting to arbitrary deformations of floating objects such as wrinkles, stretching, and curling; and calculates... The penalty center offset and aspect ratio difference are considered. Adaptive flexible targets do not have a fixed aspect ratio, thus avoiding positioning deviation for rapidly moving targets. Calculation By modeling the target boundary (e.g., top, bottom, left, and right) as 16 discrete probability distributions, sub-pixel refinement is achieved, resolving the issue of unclear boundaries caused by the transparency of floating objects and the blurring of wrinkled edges, thus significantly improving the bounding box accuracy of irregular floating objects. 3. Training with category confidence loss.
[0045] Using a single-class binary classification loss, the output image region is the probability of "floating object or background". That is, first calculate whether there is something in the current region. If there is something in the current region, then calculate what it is. Multiply the two probabilities. When the product is greater than a preset threshold, the recognition is determined to be correct. The preset threshold filters out interference such as transparent plastic being mistaken for ground or water surface, and wrinkled debris being mistaken for garbage, thereby reducing false alarms.
[0046] In one embodiment, the category confidence loss is: ; in, This represents the category confidence loss value. This indicates the probability that the current area contains floating objects. Indicates the true category label, This represents a function for calculating classification error.
[0047] 4. Target Existence Loss Training.
[0048] To determine whether there are real floating objects within the image grid, enhance the detection capabilities for stacked occlusion and multi-target overlapping scenes, and reduce missed detections.
[0049] In one embodiment, the target existence loss is: ; in, This represents the loss value indicating the existence of the target. Indicates the target confidence level. Indicates whether the actual target has a label or not.
[0050] Through the optimization of the joint loss function, the floating object detection model can accurately adapt to the floating object detection needs in the open space of substations, which are characterized by flexible deformation, random scale, blurred edges, and easy stacking and obstruction, thus providing reliable detection results for subsequent trajectory prediction and early warning.
[0051] Step 103: Perform non-maximum suppression deduplication on the initial floating object detection results to obtain the final floating object identification results.
[0052] For the initial floating object detection results output in step 102, which contain multiple initial detection boxes, a non-maximum suppression (NMS) algorithm is used for deduplication post-processing to remove redundant and duplicate boxes, resulting in a precise and unique final floating object identification result. This step also corresponds to the solidification of the post-processing strategy during the floating object detection model training phase.
[0053] Because the floating objects detected in this embodiment, such as farmland mulch, plastic bags, and kites, have characteristics such as no fixed shape, random flexible deformation, and large scale span, relaxing the constraints on the shape and size of the target during the bounding box localization process can easily lead to the problem of the same floating object being repeatedly identified by multiple detection boxes. Therefore, a non-maximum suppression algorithm is used to retain the effective detection box with the highest confidence and remove all duplicate detection boxes with an overlap greater than a preset threshold with the highest confidence box. This solves the problems of multiple target stacking, multiple boxes for one object caused by flexible deformation, duplicate annotation, and redundant detection, ensuring that the output is a unique and accurate target box.
[0054] In one embodiment, non-maximum suppression deduplication is performed on the initial floating object detection results to obtain the final floating object identification results, including: The bounding boxes are sorted from highest to lowest confidence level to determine the bounding boxes with the highest confidence level. Calculate the overlap ratio between the remaining bounding boxes and the bounding box with the highest confidence level; Remove bounding boxes with an overlap ratio greater than or equal to a preset suppression threshold, and retain bounding boxes with an overlap ratio less than the preset suppression threshold.
[0055] After nonmaximum suppression and deduplication, the final floating object identification result, which includes the precise coordinates, category, and confidence level of the floating object, is output, providing reliable and clean detection data support for subsequent dynamic trajectory prediction and hierarchical early warning.
[0056] V. Training Convergence and Model Output.
[0057] 1. Stop iterating when the total loss tends to stabilize, the recognition accuracy is greater than or equal to the preset recognition accuracy, and the false alarm rate is less than or equal to the preset false alarm rate. Optionally, the preset recognition accuracy and preset false alarm rate can be set according to requirements. For example, the preset recognition accuracy can be set to 98.5%, and the preset false alarm rate can be set to 3%.
[0058] 2. Export a lightweight training model, i.e., a trained floating object detection model that supports edge computing module deployment and has an inference latency of less than or equal to 50 milliseconds; 3. The floating object detection model supports incremental self-learning, can be continuously updated with new samples, and continuously optimizes the recognition and positioning accuracy.
[0059] This invention provides a method for identifying floating objects in open spaces of power system substations. By acquiring images of the target area in the open space of the substation around the clock, it overcomes the coverage blind spots and time limitations of manual inspections and traditional monitoring. The acquired multimodal visual images are preprocessed to obtain target images, significantly improving image quality in complex environments such as nighttime, fog, rain, snow, and strong sunlight, providing a stable and reliable data foundation for subsequent high-precision identification and ensuring monitoring effectiveness in harsh scenarios. The target multimodal image is input into a pre-trained floating object detection model to obtain initial floating object detection results containing multiple initial detection boxes. AI visual models are used for automatic identification, replacing manual inspection and analysis, greatly improving detection response speed and processing efficiency. This enables real-time capture of small-sized (≥5cm) and long-distance (≤500m) floating objects, fundamentally solving the problems of lag and high false negative rates in manual inspections. During the iterative training of the floating object detection model, a minimum... The parameters of the floating object detection model are optimized and trained using a joint loss function as the objective. The joint loss function includes at least bounding box localization loss, class confidence loss, and target existence loss to adapt to the detection of flexible and deformable floating objects. Through joint constraints of multiple loss terms, the model accurately adapts to the characteristics of targets such as farmland mulch, plastic bags, and kites, which are flexible, irregular in shape, have blurred edges, and are easily stacked and obscured, significantly improving the localization accuracy and classification accuracy of irregular floating objects. It effectively distinguishes harmful floating objects from harmless targets such as birds, fallen leaves, and ground debris, greatly reducing the false alarm rate and false negative rate. Non-maximum suppression deduplication is performed on the initial floating object detection results to obtain the final floating object recognition results. This solves the problems of multiple bounding boxes for one object, duplicate detection, and redundant labeling caused by the flexible deformation and varying scale of floating objects. It retains the effective detection boxes with the highest confidence and the most accurate location, outputting pure, unique, and reliable recognition results, avoiding duplicate warnings and interference, and improving the accuracy and efficiency of subsequent warnings and handling.
[0060] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.
[0061] The following are device embodiments of the present invention. For details not described in detail, please refer to the corresponding method embodiments described above.
[0062] Figure 2The diagram shows a schematic representation of a floating object identification device for open spaces in a power system substation, according to an embodiment of the present invention. For ease of explanation, only the parts relevant to the embodiment of the present invention are shown, and are described in detail below: like Figure 2 As shown, the floating object identification device in the open space of a power system substation includes: a data acquisition module 21, a preprocessing module 22, an identification module 23, and a processing module 24.
[0063] The acquisition module 21 is used to acquire images of the area to be detected in the open space of the substation around the clock. Preprocessing module 22 is used to preprocess the acquired multimodal visual images to obtain the target image; The recognition module 23 is used to input the target multimodal image into the pre-trained floating object detection model to obtain the initial floating object detection result containing multiple initial detection boxes. During the iterative training of the floating object detection model, the parameters of the floating object detection model are optimized and trained with the goal of minimizing the joint loss function. The joint loss function includes at least the bounding box localization loss, the class confidence loss, and the target existence loss to adapt to the detection of flexible and deformable floating objects. The processing module 24 is used to perform non-maximum suppression deduplication processing on the initial floating object detection results to obtain the final floating object recognition results.
[0064] In one possible implementation, the bounding box localization loss is based on the irregular deformation and edge fuzziness features of the floating object, and the bounding box localization of the floating object is performed by jointly calculating CIoU loss and DFL distribution focal loss.
[0065] In one possible implementation, the formula for CIoU loss is: ; in, This represents the bounding box localization loss value. Indicates the center distance of the bounding box. Indicates the center point of the bounding box. This indicates the center point of the predicted bounding box or the actual floating object's bounding box. This represents the length of the diagonal of the minimum bounding rectangle of the bounding box. This represents the penalty value for center offset. Represents the adaptive coefficient. Indicates the consistency of the aspect ratio of the bounding box; This indicates the overlap ratio between the predicted bounding box and the actual floating object annotation box. Indicates the prediction box. This indicates a real floating object label box. This function represents the overlap ratio calculation function. This indicates the height of the actual floating object label box. This indicates the width of the output prediction box. This indicates the height of the output prediction box.
[0066] In one possible implementation, the formula for the focal loss of the DFL distribution is: ; in, This represents the focal loss value of the DFL distribution. This represents the left endpoint of the discrete interval of the true label. This represents the right endpoint of the discrete interval of the true label. Indicates the predicted first The predicted probability for each discrete interval. Indicates to Take the natural logarithm.
[0067] In one possible implementation, the category confidence loss is: ; in, This represents the category confidence loss value. This indicates the probability that the current area contains floating objects. Indicates the true category label, This represents a function for calculating classification error.
[0068] In one possible implementation, the loss on the existence of the target is: ; in, This represents the loss value indicating the existence of the target. Indicates the target confidence level. Indicates whether the actual target has a label or not.
[0069] In one possible implementation, when the identification module 23 performs non-maximum suppression deduplication processing on the initial floating object detection results to obtain the final floating object identification results, it is used for: The bounding boxes are sorted from highest to lowest confidence level to determine the bounding boxes with the highest confidence level. Calculate the overlap ratio between the remaining bounding boxes and the bounding box with the highest confidence level; Remove bounding boxes with an overlap ratio greater than or equal to a preset suppression threshold, and retain bounding boxes with an overlap ratio less than the preset suppression threshold.
[0070] The above embodiments provide a floating object identification device for open spaces in power system substations. The acquisition module performs all-weather image acquisition of the target area in the open space of the substation, overcoming the coverage blind spots and time limitations of manual inspections and traditional monitoring. The preprocessing module preprocesses the acquired multimodal visual images to obtain the target image, significantly improving image quality in complex environments such as nighttime, fog, rain, snow, and strong light, providing a stable and reliable data foundation for subsequent high-precision identification and ensuring monitoring effectiveness in harsh scenarios. The identification module inputs the target multimodal image into a pre-trained floating object detection model to obtain initial floating object detection results containing multiple initial detection boxes. The AI visual model automatically identifies the floating objects, replacing manual inspection and analysis, greatly improving detection response speed and processing efficiency. This enables real-time capture of small-sized (≥5cm) and long-distance (≤500m) floating objects, fundamentally solving the problems of lag and high false negative rates in manual inspections. During the iterative training of the floating object detection model… The parameters of the floating object detection model are optimized and trained with the goal of minimizing the joint loss function. The joint loss function includes at least bounding box localization loss, class confidence loss, and target existence loss to adapt to the detection of flexible and deformable floating objects. Through joint constraints of multiple loss terms, the model can accurately adapt to the characteristics of targets such as farmland mulch, plastic bags, and kites, which are flexible, irregular in shape, have blurred edges, and are easily stacked and obscured, significantly improving the localization accuracy and classification accuracy of irregular floating objects. It can effectively distinguish harmful floating objects from harmless targets such as birds, fallen leaves, and ground debris, greatly reducing the false alarm rate and false negative rate. The processing module performs non-maximum suppression deduplication on the initial floating object detection results to obtain the final floating object recognition results. This solves the problems of multiple bounding boxes for one object, duplicate detection, and redundant annotation caused by the flexible deformation and scale variation of floating objects. It retains the effective detection boxes with the highest confidence and the most accurate location, and outputs pure, unique, and reliable recognition results, avoiding duplicate warnings and interference, and improving the accuracy and efficiency of subsequent warnings and handling.
[0071] Figure 3 This is a schematic diagram of the device provided in an embodiment of the present invention. Figure 3 As shown, the device 3 in this embodiment includes a processor 30 and a memory 31. The memory 31 stores a computer program 32. When the processor 30 executes the computer program 32, it implements the steps in the various method embodiments described above. Alternatively, when the processor 30 executes the computer program 32, it implements the functions of each module / unit in the various device embodiments described above.
[0072] For example, computer program 32 may be divided into one or more modules / units, which are stored in memory 31 and executed by processor 30 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 32 in device 3.
[0073] Device 3 may include, but is not limited to, processor 30 and memory 31. Those skilled in the art will understand that... Figure 3 This is merely an example of device 3 and does not constitute a limitation on device 3. It may include more or fewer components than shown, or combine certain components, or different components. For example, device 3 may also include input / output devices, network access devices, buses, etc.
[0074] The processor 30 can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.
[0075] The memory 31 can be an internal storage unit of the device 3, such as a hard disk or RAM of the device 3. The memory 31 can also be an external storage device of the device 3, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the device 3. Furthermore, the memory 31 can include both internal and external storage units of the device 3. The memory 31 is used to store the computer program 32 and other programs and data required by the device 3. The memory 31 can also be used to temporarily store data that has been output or will be output.
[0076] For the sake of simplicity and clarity, only the above-described functional modules / units are used as examples. In practical applications, the functions described above can be assigned to different functional modules / units as needed. These modules / units can be implemented in hardware, software, or a combination of both.
[0077] This invention also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the methods described in the above-described method embodiments.
[0078] This invention also provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the methods described in the above-described method embodiments.
[0079] Computer programs include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. Computer-readable media can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.
[0080] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Unless otherwise specified or in conflict with logic, the terminology and / or descriptions between different embodiments are consistent and can be referenced interchangeably. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.
[0081] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.
Claims
1. A method for identifying floating objects in open spaces of power system substations, characterized in that, include: All-weather image acquisition is performed on the open space of the substation to be inspected, and the acquired multimodal visual images are preprocessed to obtain the target image; The target multimodal image is input into a pre-trained floating object detection model to obtain an initial floating object detection result containing multiple initial detection boxes. During the iterative training of the floating object detection model, the parameters of the floating object detection model are optimized and trained with the goal of minimizing the joint loss function. The joint loss function includes at least bounding box localization loss, class confidence loss, and target existence loss to adapt to the detection of flexible and deformable floating objects. The initial floating object detection results are subjected to non-maximum suppression deduplication processing to obtain the final floating object identification results.
2. The method for identifying floating objects in open spaces of power system substations according to claim 1, characterized in that, The bounding box localization loss is based on the irregular deformation and edge ambiguity features of the floating object, and the bounding box localization of the floating object is performed by jointly calculating CIoU loss and DFL distribution focal loss.
3. The method for identifying floating objects in open spaces of power system substations according to claim 2, characterized in that, The formula for the CIoU loss is: ； in, This represents the bounding box localization loss value. Indicates the center distance of the bounding box. Indicates the center point of the bounding box. This indicates the center point of the predicted bounding box or the actual floating object's bounding box. This represents the length of the diagonal of the smallest bounding rectangle of the bounding box. This represents the penalty value for center offset. Represents the adaptive coefficient. Indicates the consistency of the aspect ratio of the bounding box; This indicates the overlap ratio between the predicted bounding box and the actual floating object bounding box. Indicates the prediction box. This indicates a real floating object label box. This function represents the overlap ratio calculation function. This indicates the height of the actual floating object label box. This indicates the width of the output prediction box. This indicates the height of the output prediction box.
4. The method for identifying floating objects in open spaces of power system substations according to claim 3, characterized in that, The formula for the focal loss of the DFL distribution is: ； in, This represents the focal loss value of the DFL distribution. This represents the left endpoint of the discrete interval of the true label. This represents the right endpoint of the discrete interval of the true label. Indicates the predicted first The predicted probability for each discrete interval. Indicates to Take the natural logarithm.
5. The method for identifying floating objects in open spaces of power system substations according to claim 1, characterized in that, The category confidence loss is: ； in, This represents the category confidence loss value. This indicates the probability that the current area contains floating objects. Indicates the true category label, This represents a function for calculating classification error.
6. The method for identifying floating objects in open spaces of power system substations according to claim 1, characterized in that, The loss regarding the existence of the target is: ； in, This represents the loss value indicating the existence of the target. Indicates the target confidence level. Indicates whether the actual target has a label or not.
7. The method according to claim 1, characterized in that, The process of performing non-maximum suppression deduplication on the initial floating object detection results to obtain the final floating object identification results includes: The bounding boxes are sorted from highest to lowest confidence level to determine the bounding boxes with the highest confidence level. Calculate the overlap ratio between the remaining bounding boxes and the bounding box with the highest confidence level; Bounding boxes with an overlap ratio greater than or equal to a preset suppression threshold are removed, while bounding boxes with an overlap ratio less than the preset suppression threshold are retained.
8. A floating object identification device for open spaces in power system substations, characterized in that, include: The acquisition module is used to acquire images of the open space of the substation under all-weather conditions. The preprocessing module is used to preprocess the acquired multimodal visual images to obtain the target image; The recognition module is used to input the target multimodal image into a pre-trained floating object detection model to obtain an initial floating object detection result containing multiple initial detection boxes; wherein, during the iterative training of the floating object detection model, the parameters of the floating object detection model are optimized and trained with the goal of minimizing the joint loss function, and the joint loss function includes at least bounding box localization loss, class confidence loss and target existence loss, in order to adapt to the detection of flexible and deformable floating objects; The processing module is used to perform non-maximum suppression deduplication processing on the initial floating object detection results to obtain the final floating object identification results.
9. A device, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method as described in any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.