A trend-aware weighted iterative power grid bird detection method and system

Abstract

The present application relates to the technical field of power grid bird detection, in particular to a trend perception weighted iterative power grid bird detection method and system. It comprises: deploying a lightweight target detection model on an edge device for real-time detection and outputting candidate images; transmitting the candidate images to the cloud in real time, and periodically extracting and transmitting images that do not detect birds; the cloud multi-modal large model identifies bird targets and key power equipment, calculates the Euclidean distance normalized relative to the equipment scale, and calculates the spatial risk index combined with the vertical orientation relationship and semantic confidence; determine the early warning and mark the bad examples according to the spatial risk index; input the bad example features into the gradient boosting tree model to predict the training weight; use the training weight and the spatial risk index to weight the loss function for incremental training and redeployment. The present application realizes high-precision detection, low false alarm rate and model continuous evolution ability through edge-cloud cooperation, spatial risk quantification and trend perception weighted iteration.

Description

Technical Field

[0001] This invention relates to the field of power grid bird detection technology, specifically a trend-aware weighted iterative method and system for power grid bird detection. Background Technology

[0002] With the expansion of power grid construction and the improvement of the ecological environment, bird activity poses a serious threat to power grid safety. According to statistics, bird-related faults account for 10.4% of the 110kV and above lines of the power grid company and are on the rise. The cost of a single repair can reach tens of thousands of yuan, making efficient and intelligent detection solutions urgently needed.

[0003] Existing technologies typically deploy a single lightweight target detection model on front-end devices, which has significant drawbacks: First, due to limitations in edge computing power, the model suffers from severe false negatives and false negatives in complex backgrounds, with small targets, or in adverse weather conditions; second, it cannot distinguish between distant birds and background birds, resulting in a large number of invalid warnings; and finally, the model's performance becomes fixed after deployment, making it unable to learn from new samples in the field, lacking continuous optimization capabilities, and having low reliability in long-term operation. Summary of the Invention

[0004] This invention provides a trend-aware weighted iterative method and system for detecting birds in power grids, which solves the problems of insufficient accuracy, high false alarm rate and rigid model in existing power grid bird detection methods, and achieves high-precision and continuously evolving intelligent detection.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] This invention discloses a trend-aware weighted iterative method for detecting birds in power grids, comprising:

[0007] S100: Deploys a lightweight target detection model on an edge computing device to perform real-time detection on surveillance images and output candidate images containing birds;

[0008] S200: Transmits candidate images output from the edge to the cloud in real time, and periodically extracts and transmits images of birds that were not detected to the cloud;

[0009] S300: The cloud-based multimodal large model performs global detection on the received images to identify bird targets and key power equipment; when bird targets and key power equipment are detected simultaneously, it extracts the spatial location information and semantic confidence of bird targets and key power equipment, calculates the normalized Euclidean distance between bird targets and key power equipment relative to the equipment scale, and calculates the spatial risk index by combining the vertical orientation relationship and semantic confidence.

[0010] S400: Determine whether to issue an early warning based on the spatial risk index, and mark missed, false, or low-confidence samples as bad cases;

[0011] S500: Records the spatial risk index, error type, and trend characteristics of bad examples; inputs these features into the gradient boosting tree model to automatically predict the training weights for each bad example;

[0012] S600: The loss function of bad examples is weighted using training weights and spatial risk index to incrementally train the lightweight object detection model, and the trained lightweight object detection model is redeployed to edge devices.

[0013] As a preferred embodiment of the present invention, the training process of the lightweight target detection model includes:

[0014] Construct a power grid bird image dataset containing pole ID metadata, and group it by pole ID;

[0015] The resulting tower groups are randomly divided into K disjoint subsets;

[0016] Perform K rounds of cross-validation training, selecting one subset as the validation set and the remaining subsets as the training set in each round;

[0017] The model with the best performance in K rounds of cross-validation is saved as the lightweight target detection model.

[0018] As a preferred embodiment of the present invention, the identification of bird targets and key power equipment includes:

[0019] Perform a global scan of the image to detect whether it contains bird features and equipment features;

[0020] Filter the image if it contains only one type of feature;

[0021] When both types of features are included, extract the bird target set and the key power equipment set.

[0022] As a preferred embodiment of the present invention, each bird target in the bird target set includes a detection box, geometric center point coordinates, and semantic confidence; each critical power equipment in the critical power equipment set includes a detection box, geometric center point coordinates, and semantic confidence.

[0023] As a preferred embodiment of the present invention, the calculation of the spatial risk index by combining vertical orientation relationship and semantic confidence includes:

[0024] Determine the vertical orientation of bird targets relative to critical power equipment, and set a vertical penalty factor when bird targets are above critical power equipment;

[0025] Construct a distance decay function based on normalized Euclidean distance;

[0026] The spatial risk index is calculated by combining semantic confidence, distance decay function and vertical penalty factor.

[0027] As a preferred embodiment of the present invention, marking missed, false, or low-confidence samples as bad examples includes:

[0028] For images of undetected birds that are periodically extracted and transmitted to the cloud, when the cloud detects both bird targets and critical power equipment and the spatial risk index is higher than the high-risk threshold, they are marked as missed bad cases.

[0029] For candidate images output from the edge, if the cloud simultaneously detects bird targets and critical power equipment and the spatial risk index is below the safety threshold, they are marked as false alarms or bad examples.

[0030] For candidate images output from the edge, when the cloud simultaneously detects bird targets and critical power equipment and the spatial risk index is between the safety threshold and the high-risk threshold, they are marked as low-confidence bad examples.

[0031] As a preferred technical solution of the present invention, the trend characteristics include the cumulative number of bad cases, the distribution of bad case types, and the trend of spatial risk index changes of the tower within the historical time window.

[0032] As a preferred embodiment of the present invention, the step of weighting the loss function of bad example samples using training weights and spatial risk index includes:

[0033] Calculate the classification loss and regression loss for each bad example;

[0034] The classification loss and regression loss are added together and then multiplied by the training weight and spatial risk index of the corresponding sample.

[0035] The weighted loss of all bad examples is summed to obtain the total loss for incremental training.

[0036] This invention also proposes a trend-aware weighted iterative power grid bird detection system, comprising:

[0037] The edge detection module is used to deploy a lightweight target detection model on edge computing devices to perform real-time detection on surveillance images and output candidate images containing birds;

[0038] The collaborative transmission module is used to transmit candidate images output from the edge to the cloud in real time, and periodically extract and transmit images of birds that were not detected to the cloud.

[0039] The risk calculation module is used by a cloud-based multimodal large model to perform global detection on received images, identify bird targets and critical power equipment; when bird targets and critical power equipment are detected simultaneously, the module extracts the spatial location information and semantic confidence of bird targets and critical power equipment, calculates the normalized Euclidean distance between bird targets and critical power equipment relative to the equipment scale, and calculates the spatial risk index by combining the vertical orientation relationship and semantic confidence.

[0040] The bad case marking module is used to determine whether to issue an early warning based on the spatial risk index, and to mark missed, false, or low-confidence samples as bad cases.

[0041] The weight prediction module records the spatial risk index, error type, and trend characteristics of bad examples; these features are input into the gradient boosting tree model to automatically predict the training weights for each bad example.

[0042] The iterative optimization module is used to weight the loss function of bad examples using training weights and spatial risk index, incrementally train the lightweight object detection model, and redeploy the trained lightweight object detection model to edge devices.

[0043] The beneficial effects of this invention are:

[0044] 1. This invention calculates the spatial risk index by introducing a distance decay function and a vertical potential energy rule, thereby achieving accurate quantification of the physical threat relationship between birds and power equipment. It effectively filters out interference from distant and background birds, and solves the problem of high false alarms caused by traditional methods that rely solely on visual classification.

[0045] 2. This invention utilizes gradient boosting trees to automatically predict training weights based on spatial risk index, error type, and historical trends, enabling the marginal model to prioritize high-risk missed samples during incremental learning. This achieves rapid adaptation to new scenarios while avoiding catastrophic forgetting, forming an intelligent closed loop.

[0046] 3. The lightweight target detection model deployed at the edge of this invention performs millisecond-level screening, while the large model in the cloud is responsible for global inspection and risk calculation in complex scenarios. This not only overcomes the problem of missed detection caused by insufficient edge computing power, but also avoids the bandwidth pressure of uploading all data, thus achieving a dual improvement in detection accuracy and response speed. Attached Figure Description

[0047] 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:

[0048] Figure 1 This is a flowchart illustrating a trend-aware weighted iterative method for detecting birds in a power grid according to the present invention.

[0049] Figure 2 This is a schematic diagram of the framework of a trend-aware weighted iterative method for detecting birds in a power grid according to the present invention;

[0050] Figure 3 This is a flowchart of the multidimensional space-semantic coupling discrimination process of this invention;

[0051] Figure 4 It is an adaptive weighted framework diagram based on trend perception;

[0052] Figure 5 This is a schematic diagram of the structure of a trend-aware weighted iterative power grid bird detection system according to the present invention. Detailed Implementation

[0053] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0054] This invention provides a trend-aware weighted iterative method and system for detecting birds in power grids, which solves the problems of insufficient accuracy, high false alarm rate and rigid model in existing power grid bird detection methods, and achieves high-precision and continuously evolving intelligent detection.

[0055] The "trend perception" described in this invention refers to the system dynamically identifying the problem evolution patterns of different tower locations by analyzing the cumulative number of bad cases, the distribution of bad case types, and the changing trends of the spatial risk index of the towers to which the bad case samples belong within a historical time window. Specifically, for tower locations where bad cases frequently occur, the system determines that there is a persistent problem and automatically increases the training weight of newly added bad cases at that location in subsequent incremental training; for tower locations where the problem has been largely resolved, the system determines that this type of problem has converged and automatically decreases the baseline weight of new samples of this type. This trend perception mechanism runs through the weight prediction process in step S500, mapping historical trend features to training weights through a gradient boosting tree model, enabling the model to adaptively optimize according to historical evolution patterns.

[0056] Example 1: As Figure 1 and Figure 2 As shown, the present invention provides a trend-aware weighted iterative method for detecting birds in power grids, comprising:

[0057] S100: Deploys a lightweight target detection model on an edge computing device to perform real-time detection on surveillance images and output candidate images containing birds;

[0058] The purpose of this step is to perform preliminary screening of surveillance images at the edge, filtering out most of the pure background images that do not contain bird targets, thereby reducing the amount of data processed in the cloud later.

[0059] Furthermore, the training process of the lightweight target detection model includes:

[0060] Construct a power grid bird image dataset containing pole ID metadata, and group it by pole ID;

[0061] The resulting tower groups are randomly divided into K disjoint subsets;

[0062] Perform K rounds of cross-validation training, selecting one subset as the validation set and the remaining subsets as the training set in each round;

[0063] The model with the best performance in K rounds of cross-validation is saved as the lightweight target detection model.

[0064] Specifically, a dataset of images depicting bird damage to power transmission lines was constructed. While annotating the image content, each image was precisely recorded with its key metadata—a "tower ID"—to identify the specific tower location corresponding to the image. This is crucial for subsequent trend feature extraction. Subsequently, the constructed dataset was divided into a training set according to a predetermined ratio. and the final test set The training set is used for model training and validation, while the final test set is completely archived and used only for a one-time evaluation of the model's baseline performance.

[0065] To avoid data leakage and improve the model's generalization ability, this invention adopts a K-fold cross-validation strategy based on pole ID grouping:

[0066] training set Grouping based on "tower ID" metadata ensures that all image samples from the same tower are treated as an indivisible unit, preventing images from the same tower from appearing in both the training and validation sets simultaneously.

[0067] The resulting tower groups are randomly and uniformly divided into K disjoint subsets.

[0068] Perform K rounds of independent training and validation. In each iteration, select one subset (all data from several tower groups) as the validation set, and use the remaining K-1 subsets as the training set for that round. The value of K is usually 5-10.

[0069] In each training round, a transfer learning strategy is adopted: a target detection model (such as YOLOv8) pre-trained on a large-scale general target detection dataset (such as the MSCOCO dataset) is selected as the initial model. This pre-trained model has learned general feature extraction capabilities. The backbone network weights of the pre-trained model are frozen, and only the detection head part is fine-tuned using the training set of this round, thereby transferring the general target detection capability to the specific domain of power grid bird detection. At the same time, the validation set of this round is used to monitor and evaluate the model performance in real time.

[0070] After K rounds of cross-validation, the model with the best overall performance in K rounds is saved as a lightweight target detection model for deployment at the edge.

[0071] After the model training is complete, the saved optimal model is applied to the final test set to evaluate the model's baseline performance. The performance metrics are recorded as the baseline performance of the model.

[0072] Subsequently, the optimal model was deployed on an edge computing device in a real-world operating environment. The deployed lightweight model then began processing online-acquired surveillance images, performing real-time detection on each frame. When the model detected a bird target in an image, it output a candidate image containing the bird and its corresponding detection box information; when the model did not detect a bird target, the image was initially determined to be a background image without the target. In this way, the edge computing device completed the first-level screening of the surveillance images.

[0073] S200: Transmits candidate images output from the edge to the cloud in real time, and periodically extracts and transmits images of birds that were not detected to the cloud;

[0074] This step establishes a data transmission mechanism for edge-cloud collaboration, and initiates two independent collaborative processing paths based on the detection results from the edge.

[0075] Specifically, when the lightweight target detection model detects a bird target and outputs a candidate bounding box, the system determines it to be in a "bounding box state." At this point, the image containing the candidate bounding box and its detection box information are transmitted to the cloud server in real time, entering the cloud-based collaborative discrimination process. This path ensures a timely response to potential threats.

[0076] When the lightweight object detection model fails to detect any bird targets and does not output candidate boxes, the system classifies it as a "frameless state". Considering that the lightweight object detection model may miss detections in complex backgrounds, with small targets, or under adverse weather conditions, the system adopts a periodic extraction mechanism: images are extracted from frameless images and transmitted to the cloud at preset time intervals (e.g., every 30-300 seconds) or according to the number of images (e.g., every 50-200 frameless images). The extraction strategy can be adjusted according to the frequency of bird activity and bandwidth resources.

[0077] Through these two paths, the system ensures real-time processing of detected targets and compensates for the missed detection defects of the edge model through a periodic inspection mechanism, providing a data foundation for subsequent cloud-based collaborative discrimination and continuous model optimization.

[0078] S300: The cloud-based multimodal large model performs global detection on the received images to identify bird targets and key power equipment; when bird targets and key power equipment are detected simultaneously, it extracts the spatial location information and semantic confidence of bird targets and key power equipment, calculates the normalized Euclidean distance between bird targets and key power equipment relative to the equipment scale, and calculates the spatial risk index by combining the vertical orientation relationship and semantic confidence.

[0079] This step achieves the leap from "target identification" to "threat assessment" by establishing a "multi-dimensional space-semantic coupling discrimination mechanism." Specifically, as follows... Figure 3 As shown:

[0080] Furthermore, the bird-identifying targets and key electrical equipment include:

[0081] Perform a global scan of the image to detect whether it contains bird features and equipment features;

[0082] Filter the image if it contains only one type of feature;

[0083] When both types of features are included, extract the bird target set and the key power equipment set.

[0084] Furthermore, each bird target in the bird target set includes a detection box, geometric center point coordinates, and semantic confidence; each key power equipment in the key power equipment set includes a detection box, geometric center point coordinates, and semantic confidence.

[0085] Specifically, the cloud-based multimodal large model performs a global scan of the received images to detect whether they contain bird and equipment features. This cloud-based multimodal large model can employ a multimodal large language model with visual understanding capabilities, such as commercial models like GPT-4V, Gemini Pro Vision, and Claude, or a custom detection model built on vision-language pre-training architectures like CLIP and BLIP. The model receives image data and text prompts (such as "Please detect bird targets and electrical equipment in the image and label their locations"), and outputs structured detection results including bounding box coordinates, category labels, and confidence scores. If only one type of feature is detected (only birds or only equipment), the system determines that the image does not pose a threat, directly filters the image, and does not proceed to the subsequent risk assessment process.

[0086] Only when both bird and device features are detected simultaneously, the cloud-based large model decouples the two target classes from the environmental background, extracting two independent feature sets:

[0087] Bird target set Each bird target include: : No. A detection box for a bird target; : No. The coordinates of the geometric center point of each bird target; : No. Semantic confidence of bird targets.

[0088] Key power equipment assembly Each device target include: : No. Inspection framework for key power equipment (insulators, equipotential rings, conductor hanging points, etc.); : No. Coordinates of the geometric center point of a key power equipment; : No. Semantic confidence of key power equipment.

[0089] in, The number of bird targets detected. This represents the number of critical power devices detected. The geometric center point coordinates are the pixel coordinates of the center position of the detection box, and the semantic confidence score is the confidence score of the large model in judging the target category.

[0090] Furthermore, the calculation of the spatial risk index by combining vertical orientation relationship and semantic confidence includes:

[0091] Determine the vertical orientation of bird targets relative to critical power equipment, and set a vertical penalty factor when bird targets are above critical power equipment;

[0092] Construct a distance decay function based on normalized Euclidean distance;

[0093] The spatial risk index is calculated by combining semantic confidence, distance decay function and vertical penalty factor.

[0094] Specifically, to eliminate errors caused by image scale and shooting distance, this step constructs a relative spatial field for normalization:

[0095] For each pair of bird targets With key power equipment The diagonal length of the inspection frame for key power equipment is selected as the physical dimensional benchmark.

[0096] ;

[0097] in, and The first Width and height of the detection frame for each key power equipment.

[0098] Calculate bird targets With key power equipment Normalized Euclidean distance between :

[0099] ;

[0100] in, The Euclidean distance between the coordinates of two center points is calculated as follows:

[0101] ;

[0102] in, and The first The x and y coordinates of the center point of each bird target and The first The x and y coordinates of the center point of each key power equipment. Divide by... Normalization eliminates the influence of shooting distance and image resolution.

[0103] Establish vertical orientation determination rules : Set the image coordinate system so that the y-axis points downwards, and the ordinate of the bird's center point is... Less than the ordinate of the equipment center point (i.e., the bird is located above the device), then it is determined that... ;otherwise This rule is used to identify high-risk spatial configurations where birds are positioned above insulators, as bird droppings are more likely to contaminate the insulator surface, leading to a flashover risk.

[0104] Based on the extracted spatial geometric parameters and semantic information, the system constructs a formula for calculating the Spatial Risk Index (SRI). For each pair of coupled targets... Its risk value is calculated as follows:

[0105] ;

[0106] Among them, the first item : Represents semantic certainty, which is the product of two confidence levels, and its value ranges from [0,1]. This term ensures that the risk assessment is based on reliable target detection results. When the confidence level of any target is low, the overall risk value is reduced accordingly.

[0107] Second item : A distance attenuation function in Gaussian distribution, simulating the nonlinear decreasing trend of electric field risk with distance. Parameters The radius constant controls the decay rate. The value can be set according to the safety distance standard in the power safety regulations. In the preferred embodiment, Possible values ​​for the normalized distance When the distance equals the safety distance threshold (e.g., 2-3 times the equipment size), the attenuation function value drops to a certain level (e.g., 0.1-0.3). Specifically, let the safety distance threshold be... The expected decay value is Then, we can solve the equation of the attenuation function to obtain... .when When the bird coincides with the center point of the equipment, the second value is 1, reaching the maximum threat; as... As the value increases, it decreases along a Gaussian curve, reflecting the physical law that the threat decreases rapidly with distance.

[0108] Third item : This is the vertical potential energy gain term, with parameters... This is a vertical penalty coefficient used to reinforce the risk weighting when birds are located above equipment. According to power system operation statistics, the failure risk is typically 50%-200% higher when birds are above equipment than in other locations. Typically, a value of 0.5-2.0 is used, but the specific value can be determined based on historical fault statistics. When the bird is at the top, this value is... Applying a spatial risk index Magnification by a factor of 1; when When the birds are not above, this item has a value of 1, and no additional penalty is incurred. For example, when At that time, birds located above the equipment will double the risk value.

[0109] By multiplying the above three terms, A quantitative risk assessment of a single bird-equipment coupling was formed by comprehensively considering the reliability of target identification (semantic confidence), spatial proximity (distance attenuation), and vertical positional threat (potential gain).

[0110] When an image contains multiple bird targets and multiple critical power devices, the system needs to calculate the risk value of all possible bird-device pairs and take the global maximum value as the final spatial risk index for the image.

[0111] ;

[0112] This involves iterating through M×N pairs and selecting the pair with the highest risk value as the threat level for the image. This maximum value strategy ensures that even if there is only one pair of high-risk coupled targets in the image, the system can accurately identify the threat and will not be diluted by the presence of other low-risk pairs.

[0113] S400: Determine whether to issue an early warning based on the spatial risk index, and mark missed, false, or low-confidence samples as bad cases;

[0114] This step is based on the spatial risk index calculated in step S300. Based on preset safety and high-risk thresholds, the image is assessed for threat level and marked as a bad example. This step executes different judgment logic for the two collaborative processing paths in S200.

[0115] Furthermore, marking missed, false positive, or low-confidence samples as bad examples includes:

[0116] For images of undetected birds that are periodically extracted and transmitted to the cloud, when the cloud detects both bird targets and critical power equipment and the spatial risk index is higher than the high-risk threshold, they are marked as missed bad cases.

[0117] For candidate images output from the edge, if the cloud simultaneously detects bird targets and critical power equipment and the spatial risk index is below the safety threshold, they are marked as false alarms or bad examples.

[0118] For candidate images output from the edge, when the cloud simultaneously detects bird targets and critical power equipment and the spatial risk index is between the safety threshold and the high-risk threshold, they are marked as low-confidence bad examples.

[0119] Specifically, when candidate bounding boxes are output at the edge and the image is transmitted to the cloud in real time, after the cloud-based global detection in step S300: if the cloud only detects a single category feature, the image is filtered and no warning is issued. If the cloud simultaneously detects bird targets and critical power equipment, and calculates the spatial risk index... Then the following judgment is made:

[0120] when Greater than the preset high-risk threshold When the system confirms a high-risk threat, it immediately triggers an alert and marks the image as a "positive sample." Between the safety threshold High-risk threshold between( When a detection result is deemed insufficiently confident, the system does not issue a warning but marks the image as a "low-confidence bad case." These types of samples typically correspond to boundary situations with ambiguous risk levels, requiring further training to improve the model's discriminative ability.

[0121] when Less than the safety threshold If a bird is identified as a distant target or background bird and does not pose an actual threat, no warning is issued, and the image is marked as a "false positive." This type of sample corresponds to a false detection in the edge model, where candidate boxes are output at the edge but there is actually no threat.

[0122] When no candidate bounding boxes are output at the edge, and images transmitted to the cloud are periodically extracted, after the global inspection in step S300: if the cloud detects only a single category feature or no feature at all, the edge judgment is confirmed to be correct, and no further operations are performed. Greater than the high-risk threshold When the system detects a "high-risk missed detection" at the edge, it immediately triggers an alert and marks the image as a "missed bad instance." These types of samples are key areas for model optimization, representing high-risk vulnerabilities that the edge model failed to identify. Not greater than the high-risk threshold When the system determines that the missed detection at the edge will not cause serious consequences, it will not trigger an early warning and will operate normally.

[0123] The above judgment logic involves two key thresholds: a safety threshold to distinguish between low-risk scenarios and scenarios requiring attention, determined by statistically analyzing the spatial risk index distribution of distant and background birds and taking their upper quantile; and a high-risk threshold to identify high-risk scenarios requiring immediate warning, determined by statistically analyzing the spatial risk index distribution of historical fault cases and taking their lower quantile, or by inversely calculating based on the safety distance requirements in power safety regulations. Distant birds refer to bird targets whose detection box area to total image area ratio is below a preset threshold, as they are too far away and too small to pose a real threat. Background birds refer to bird targets whose detection box center point is located in the image edge region or whose detection box is cropped by the image boundary, making it difficult to accurately determine their threat level due to their remote location. These two types of birds are identified by calculating the geometric features of the detection box (area ratio, boundary position relationship) and combining them with their spatial risk index.

[0124] Under normal circumstances, The interval between the two thresholds These are ambiguous regions, corresponding to low-confidence samples. Following the above judgment process, any cases where the cloud-based large model has corrected the edge detection results are marked as "bad examples" of the corresponding type.

[0125] Unreported bad cases: Situations where high-risk targets are not detected at the edge but are detected in the cloud;

[0126] False alarm: A situation where a problem is detected at the edge but is determined to be low-risk by the cloud.

[0127] Low-confidence bad cases: Cases where the risk value is detected at the edge but determined to be in an ambiguous range by the cloud.

[0128] The system will list these bad examples and their corresponding... Values, error types, and tower ID information are automatically recorded and stored, providing a data foundation for the subsequent weighted training of S500.

[0129] S500: Records the spatial risk index, error type, and trend characteristics of bad examples; inputs these features into the gradient boosting tree model to automatically predict the training weights for each bad example;

[0130] This step constructs a "trend-aware adaptive weighting mechanism" as follows: Figure 4 As shown, the optimal training weights for bad examples are automatically learned through the Gradient Boosting Tree (GBDT) model, thus achieving intelligent incremental training of the model.

[0131] The system records and stores all bad sample cases marked in step S400 and their related information. For each bad sample... Record its image data and the spatial risk index calculated in step S300. Error type determined in step S400 (Misreporting, false alarms, or low confidence) and the corresponding tower identifier for that sample. .

[0132] These bad examples are continuously recorded and accumulated to form a bad example database, providing a data foundation for subsequent trend analysis.

[0133] Furthermore, the trend characteristics include the cumulative number of defective cases, the distribution of defective case types, and the trend of spatial risk index changes of the tower within the historical time window.

[0134] Specifically, the trend characteristics include the cumulative number of defective cases, the distribution of defective case types, and the trend of spatial risk index changes for the corresponding tower within a historical time window. For defective case samples whose weights are to be calculated... The system extracts its tower identification. Then, retrieve all bad case records for the tower within the historical time window from the bad case database and calculate the following trend characteristics:

[0135] The first item is the cumulative number of bad cases, which counts the total number of bad cases accumulated for this tower within the time window. This characteristic reflects the frequency of problems at this tower location.

[0136] The second item is the distribution of bad case types, which counts the number or percentage of missed alarms, false alarms, and low-confidence bad cases for this tower, denoted as ( This feature reflects the main type of problem with the tower.

[0137] The third item is the trend of spatial risk index changes. Based on the historical bad examples of this tower Value sequence, calculate average Maximum value Or statistical measures such as the rate of change in a time series, this feature reflects the evolution of the threat level of the tower.

[0138] The length of the historical time window can be set according to the actual operation and maintenance cycle. In the preferred implementation, a certain number of bad case records in the recent period (such as the most recent 50-200 bad cases) or records within a certain period of time (such as the most recent 1-3 months) can be taken.

[0139] It should be noted that the instantaneous characteristics of a single bad sample are ( , , The historical aggregation characteristics of the towers to which the bad examples belong are ( , , , , The former reflects the individual attributes of the current sample, while the latter reflects the historical trend of the tower's location. Only by combining the two can the value of bad examples and training priorities be fully characterized.

[0140] The system uses high-dimensional class features of bad cases. , , Using [] as the query key, extract the corresponding trend data from the "Bad Case Calculation Section". , , , , The query values ​​are combined to construct the complete input feature vector. .

[0141] The system uses a gradient boosting tree (GBDT) model to learn feature vectors. With optimal training weights The complex nonlinear mapping relationship between them, i.e. .

[0142] The GBDT model is trained based on empirical data accumulated during historical iterations. During model iteration, the system records the training effects of different feature combinations under different weight configurations, including metrics such as convergence speed and accuracy improvement. This empirical data constitutes the supervision signal for the GBDT model. The feature vectors of historical bad examples are used as input, and the corresponding effective weight values ​​are used as output labels to train the GBDT model. GBDT constructs multiple decision trees, with each tree fitting the residuals of the previous tree, gradually approximating the mapping relationship between features and weights.

[0143] Weight prediction follows these principles. Those with higher... Samples with high values ​​are given greater weight because they represent more serious threats. For poles with frequent bad examples (…),… For poles with relatively large numbers of new bad cases, the weight of newly added bad cases is appropriately increased because there is a persistent problem at this location. For bad cases of the "missed detection" type, a higher weight is given compared to false positives because the harm caused by missed detections is greater. For poles where a certain type of problem has been largely resolved, for example, if the accuracy rate of identifying a certain type of error on that pole has remained above 95% for three consecutive iterations, the system determines that this type of problem has converged, automatically and smoothly lowers the baseline weight of new samples of this type, and transfers computational resources to other problems that need optimization.

[0144] Through automatic prediction using the GBDT model, the system identifies each bad case sample. Generate personalized training weights These weight values ​​will be applied to the weighted loss function in step S600.

[0145] S600: The loss function of bad examples is weighted using training weights and spatial risk index to incrementally train the lightweight object detection model, and the trained lightweight object detection model is redeployed to edge devices.

[0146] This step completes the incremental training and redeployment of the model, forming a continuously evolving closed-loop system. By combining the training weights predicted in step S500 with the spatial risk index, a weighted loss function is constructed, enabling the edge model to prioritize high-risk targets during backpropagation, thereby gradually learning the implicit features that distinguish between "high-risk hazards" and "safe backgrounds".

[0147] Specifically, the system continuously records the bad sample marked in step S400 and its corresponding training weights. Space risk index When a preset number of bad examples are accumulated, the system automatically triggers an incremental training process. This preset number can be set according to the data scale and training resources of the actual operation and maintenance scenario. In a preferred embodiment, it can be set to trigger training once every 50-200 bad examples.

[0148] The accumulated bad examples constitute the incremental training set, denoted as . ,in For the first Image data of a bad example, The corresponding annotation information (including the location of the detection box and the category). The training weights predicted in step S500, The spatial risk index calculated in step S300. This is the total number of bad examples accumulated during training.

[0149] Furthermore, the weighting of the loss function for bad examples using training weights and a spatial risk index includes:

[0150] Calculate the classification loss and regression loss for each bad example;

[0151] The classification loss and regression loss are added together and then multiplied by the training weight and spatial risk index of the corresponding sample.

[0152] The weighted loss of all bad examples is summed to obtain the total loss for incremental training.

[0153] Specifically, firstly, for each bad example in the incremental training set... Calculate its classification loss in the object detection task. and regression loss The classification loss measures the accuracy of the model in determining the target category, and is usually achieved using cross-entropy loss; the regression loss measures the accuracy of the model in predicting the location of the detection box, and is usually achieved using smoothed L1 loss or IoU loss. Next, the classification loss and regression loss of the sample are added together to obtain the base loss for that sample. Then, the base loss is double-weighted and multiplied by the training weights corresponding to that sample. Space risk index The weighted loss for this sample is obtained as follows: .

[0154] Finally, for all of the incremental training sets The weighted sum of the losses for each bad example is used to obtain the total loss function for incremental training:

[0155] ;

[0156] The training weights in the weighted loss function The spatial risk index is automatically predicted by the GBDT model based on the individual and trend characteristics of the sample, reflecting the value of the sample in model optimization; The physical threat level of the corresponding scenario for the sample is quantified by calculations from a large cloud-based model. The product of these two factors ensures that the system prioritizes learning bad examples that pose both high physical risk and high value for model improvement.

[0157] Use the constructed weighted loss function The lightweight object detection model deployed in step S100 is then incrementally trained. The incremental training employs a fine-tuning strategy, using the currently deployed edge model parameters as initialization on the incremental training set. The model is trained for several rounds, and the backpropagation algorithm is used to update the model parameters and minimize the weighted loss function. .

[0158] During training, because the loss from high-risk, high-value samples is significantly amplified, the model parameters will be adjusted more towards correctly identifying these high-risk samples during gradient backpropagation. This allows the model to gradually learn the implicit features that distinguish between "high-risk hazards" and "safe backgrounds." Hyperparameters such as the learning rate and batch size can be adjusted according to the incremental training set size and model convergence.

[0159] After incremental training is completed, the system evaluates the performance of the updated model on the validation set, paying particular attention to the changes in the accuracy of identifying various bad cases, and confirming that the model performance has improved compared to before training.

[0160] After confirming the improved model performance, the updated lightweight object detection model is redeployed to the edge computing device, replacing the original model. The updated model continues to perform the real-time detection task in step S100 and re-enters the collaborative inference and iterative optimization loop in steps S200 to S600.

[0161] Through the aforementioned periodic incremental training and model updates, the system forms an intelligent closed loop capable of continuously learning from actual operational errors and self-optimizing, thus achieving continuous evolution of the edge model. With each iteration, the edge model's ability to capture high-risk vulnerabilities continuously improves, while the false negative and false positive rates steadily decrease, ultimately reaching a stable and reliable operating state.

[0162] Example 2: A power grid company adopted the trend-aware weighted iterative power grid bird detection system of this invention for bird damage monitoring scenarios on transmission lines, such as... Figure 5 As shown, it includes an edge detection module, a collaborative transmission module, a risk calculation module, a bad case marking module, a weight prediction module, and an iterative optimization module. The specific implementation of each module is described in the corresponding steps of Example 1.

[0163] The power grid company deployed edge computing devices on dozens of towers along a key transmission corridor, with each device running a lightweight target detection model. In the initial stages of system deployment, there were numerous false alarms for distant and background birds, causing difficulties for maintenance personnel.

[0164] After the system activates the edge-cloud collaboration mechanism, the cloud-based multimodal large model performs secondary discrimination on the candidate images output from the edge. This is achieved by calculating a spatial risk index. The system successfully filtered out most of the interference from distant and background birds. For example, in a monitoring image of a certain tower, the edge detection detected birds on the treetops in the distance and output candidate boxes, but the results calculated by the cloud... The value was low, so the system determined it to be a distant target and did not issue a warning, effectively reducing the number of invalid warnings. Meanwhile, the cloud-based periodic sampling mechanism successfully recovered high-risk missed alerts from the edge. For example, the system randomly selected a bird from a frameless image of a tower and found it perched directly above an insulator, calculating a high value... If the value is detected, an alert is immediately triggered and the case is marked as a missed bad case, and maintenance personnel rush to the scene to handle it in a timely manner.

[0165] The system continuously accumulates bad case samples during operation. Due to frequent bird activity in the surrounding area, a certain pole accumulated a significant number of missed bad cases in the first iteration. The GBDT model automatically assigned higher training weights to newly added bad cases at this pole based on the accumulated number of bad cases, the percentage of missed cases, and the average spatial risk index. After the first round of incremental training, the edge model significantly improved its detection rate at this pole location.

[0166] For another tower where the problem has been largely resolved, the accuracy of its false alarm rate has remained stable after multiple iterations. The system automatically reduces the weight of new samples for this tower, allocating computational resources to other locations requiring optimization. After multiple rounds of incremental training iterations, the system performance has been significantly improved, as shown in Table 1.

[0167] Table 1 System Performance Comparison Table

[0168]

[0169] After the system was deployed and put into operation, the number of bird-related faults in the power transmission corridor decreased significantly, and the frequency of deployment and workload of maintenance personnel were significantly reduced. The system requires no additional dedicated sensors or modification of tower equipment; it achieves intelligent monitoring of bird damage using only existing monitoring equipment, demonstrating high engineering practicality and widespread application value.

[0170] 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 trend-aware weighted iterative method for detecting birds in power grids, characterized in that, include: S100: Deploys a lightweight target detection model on an edge computing device to perform real-time detection on surveillance images and output candidate images containing birds; S200: Transmits candidate images output from the edge to the cloud in real time, and periodically extracts and transmits images of birds that were not detected to the cloud; S300: The cloud-based multimodal large model performs global detection on the received images to identify bird targets and key power equipment; when bird targets and key power equipment are detected simultaneously, it extracts the spatial location information and semantic confidence of bird targets and key power equipment, calculates the normalized Euclidean distance between bird targets and key power equipment relative to the equipment scale, and calculates the spatial risk index by combining the vertical orientation relationship and semantic confidence. S400: Determine whether to issue an early warning based on the spatial risk index, and mark missed, false, or low-confidence samples as bad cases; S500: Records the spatial risk index, error type, and trend characteristics of bad sample data; These features are input into a gradient boosting tree model to automatically predict the training weights for each bad case; S600: The loss function of bad examples is weighted using training weights and spatial risk index to incrementally train the lightweight object detection model, and the trained lightweight object detection model is redeployed to edge devices.

2. The trend-aware weighted iterative method for detecting birds in power grids according to claim 1, characterized in that, The training process of the lightweight object detection model includes: Construct a power grid bird image dataset containing pole ID metadata, and group it by pole ID; The resulting tower groups are randomly divided into K disjoint subsets; Perform K rounds of cross-validation training, selecting one subset as the validation set and the remaining subsets as the training set in each round; The model with the best performance in K rounds of cross-validation is saved as the lightweight target detection model.

3. The trend-aware weighted iterative method for detecting birds in power grids according to claim 1, characterized in that, The bird-identifying targets and key power equipment include: Perform a global scan of the image to detect whether it contains bird features and equipment features; Filter the image if it contains only one type of feature; When both types of features are included, extract the bird target set and the key power equipment set.

4. The trend-aware weighted iterative method for detecting birds in power grids according to claim 3, characterized in that, Each bird target in the bird target set includes a bounding box, geometric center point coordinates, and semantic confidence; each critical power equipment in the critical power equipment set includes a bounding box, geometric center point coordinates, and semantic confidence.

5. The trend-aware weighted iterative method for detecting birds in power grids according to claim 1, characterized in that, The calculation of the spatial risk index by combining vertical orientation relationship and semantic confidence includes: Determine the vertical orientation of bird targets relative to critical power equipment, and set a vertical penalty factor when bird targets are above critical power equipment; Construct a distance decay function based on normalized Euclidean distance; The spatial risk index is calculated by combining semantic confidence, distance decay function and vertical penalty factor.

6. The trend-aware weighted iterative method for detecting birds in power grids according to claim 1, characterized in that, The step of marking missed, false, or low-confidence samples as bad examples includes: For images of undetected birds that are periodically extracted and transmitted to the cloud, when the cloud detects both bird targets and critical power equipment and the spatial risk index is higher than the high-risk threshold, they are marked as missed bad cases. For candidate images output from the edge, if the cloud simultaneously detects bird targets and critical power equipment and the spatial risk index is below the safety threshold, they are marked as false alarms or bad examples. For candidate images output from the edge, when the cloud simultaneously detects bird targets and critical power equipment and the spatial risk index is between the safety threshold and the high-risk threshold, they are marked as low-confidence bad examples.

7. The trend-aware weighted iterative method for detecting birds in power grids according to claim 1, characterized in that, The trend characteristics include the cumulative number of bad cases, the distribution of bad case types, and the trend of spatial risk index changes of the tower within the historical time window.

8. The trend-aware weighted iterative method for detecting birds in power grids according to claim 1, characterized in that, The weighting of the loss function for bad examples using training weights and a spatial risk index includes: Calculate the classification loss and regression loss for each bad example; The classification loss and regression loss are added together and then multiplied by the training weight and spatial risk index of the corresponding sample. The weighted loss of all bad examples is summed to obtain the total loss for incremental training.

9. A trend-aware weighted iterative power grid bird detection system, characterized in that, include: The edge detection module is used to deploy a lightweight target detection model on edge computing devices to perform real-time detection on surveillance images and output candidate images containing birds; The collaborative transmission module is used to transmit candidate images output from the edge to the cloud in real time, and periodically extract and transmit images of birds that were not detected to the cloud. The risk calculation module is used by a cloud-based multimodal large model to perform global detection on received images, identify bird targets and critical power equipment; when bird targets and critical power equipment are detected simultaneously, the module extracts the spatial location information and semantic confidence of bird targets and critical power equipment, calculates the normalized Euclidean distance between bird targets and critical power equipment relative to the equipment scale, and calculates the spatial risk index by combining the vertical orientation relationship and semantic confidence. The bad case marking module is used to determine whether to issue an early warning based on the spatial risk index, and to mark missed, false, or low-confidence samples as bad cases. The weight prediction module is used to record the spatial risk index, error type, and trend characteristics of bad sample data. These features are input into a gradient boosting tree model to automatically predict the training weights for each bad case; The iterative optimization module is used to weight the loss function of bad examples using training weights and spatial risk index, incrementally train the lightweight object detection model, and redeploy the trained lightweight object detection model to edge devices.

Images