Electric tower stockbridge damper defect detection method based on multi-scale dynamic artificial bee colony algorithm hyper-parameter optimization

Abstract

The invention discloses an electric tower shockproof hammer defect detection method based on multi-scale dynamic artificial bee colony algorithm hyper-parameter optimization, and the method comprises the steps: designing a dynamic bee colony number adjustment strategy, and adaptively adjusting the proportion of employed bees to scout bees in an iteration process; improving the search range of the employed bees and deleting the following bees to simplify the swarm structure according to the similarity of the optimization strategies of the original employed bees and the following bees; an employed bee candidate mechanism is designed in the employed bee stage, and the number of searching times of employed bees in the current searching range is increased; a last employment bee elimination mechanism is designed, fitness sorting is carried out in the iteration process, and elimination of the last employment bee is converted into scout bees of next iteration; a dynamic probability elite retention mechanism is designed in the reconnaissance bee stage, and a historical optimal solution is used for guiding the search direction; a scout bee multi-scale variation search mechanism is designed, and global exploration and local development capabilities are balanced by adaptively adjusting disturbance scales. According to the method, the bee colony structure is innovatively improved, the dilemma that a traditional artificial bee colony algorithm is caught in a local optimal solution is relieved, the efficiency of the traditional artificial bee colony algorithm for hyper-parameter tuning is improved, and the convergence speed of the artificial bee colony algorithm is remarkably increased.

Description

Technical Field

[0001] This invention relates to the fields of intelligent optimization algorithms, computer vision, power line inspection, and target detection, and particularly to a method for detecting damage to power tower vibration dampers based on hyperparameter optimization using a multi-scale dynamic artificial bee colony algorithm. Background Technology

[0002] With the continuous advancement of smart grid construction in my country, the safe and stable operation of the power system has become a crucial link in ensuring social and economic development and meeting people's livelihood needs. Transmission lines, as an important component of the power system, are constantly exposed to complex and changing natural environments, making them susceptible to external factors such as wind, ice loads, and temperature variations. This can lead to increased line vibration, resulting in safety hazards such as hardware fatigue and conductor strand breakage. Among these, vibration dampers, as the core device for suppressing conductor vibrations in the wind, directly affect the mechanical strength and service life of the transmission line. If a vibration damper shifts, falls off, or is damaged, its vibration damping effect will be significantly weakened, potentially leading to line breakage accidents in severe cases. Therefore, regular, efficient, and accurate condition monitoring of vibration dampers has become a critical task in power line inspection.

[0003] With the development of computer vision and drone technology, power line visual inspection systems based on drones equipped with cameras are gradually becoming the mainstream solution. This system can achieve automated, high-precision image acquisition of transmission lines and their auxiliary equipment, providing a data foundation for subsequent intelligent analysis. However, processing massive amounts of inspection images still faces significant challenges, necessitating efficient image recognition algorithms to achieve automatic detection and condition assessment of vibration dampers.

[0004] The YOLO series of algorithms is widely used in industrial vision inspection scenarios due to its balance between high detection speed and good accuracy. YOLOv8, in particular, has undergone several optimizations in network structure, loss function, and training strategy, further improving detection performance. However, the performance of YOLOv8 is highly dependent on the settings of hyperparameters, such as learning rate, anchor box size, weight decay coefficient, and data augmentation intensity. These parameters are usually tuned empirically or through grid search, which is not only time-consuming and labor-intensive but also difficult to achieve globally optimal configurations in complex scenarios. Therefore, optimization algorithms are needed to optimize these hyperparameters.

[0005] Artificial Bee Colony (ABC) is a metaheuristic optimization algorithm derived from swarm intelligence. First proposed in 2005 by Turkish scholar Dervis Karaboga, it was inspired by the foraging behavior of bee colonies searching for high-quality nectar sources in nature. Since its inception, ABC has been widely used in function optimization, engineering design, and machine learning due to its simple structure, few parameters, and strong robustness. Compared to genetic algorithms (GA) which rely on crossover and mutation operations, or particle swarm optimization (PSO) which is prone to getting trapped in local optima, ABC effectively avoids premature convergence while maintaining population diversity through a mechanism of "neighborhood perturbation + probabilistic selection + random restart." However, the standard ABC algorithm has gradually revealed shortcomings in later development capabilities and slow convergence speed, especially limiting its performance in high-dimensional and complex problems. Summary of the Invention

[0006] To address the shortcomings of the standard ABC algorithm, such as insufficient development capabilities and slow convergence speed, this invention proposes several improvements to the standard ABC algorithm, including: improving the search range of hired bees and removing follower bees to simplify the bee colony structure; designing a hired bee candidate mechanism; designing a hired bee last-place elimination mechanism; designing a dynamic probabilistic elite retention mechanism; designing a multi-scale mutation search mechanism for scout bees; and proposing a dynamic bee colony size adjustment strategy.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows: A method for detecting defects in power tower vibration dampers based on hyperparameter optimization using a multi-scale dynamic artificial bee colony algorithm, comprising the following steps:

[0008] S1. Collect images of vibration dampers on power towers using drones, construct a dataset, and complete data annotation and preprocessing;

[0009] S2. Construct a YOLOV8 target detection network, extract multi-scale features through the backbone network, fuse features of different scales through the neck network, and output the localization and classification results of the shock absorber through the detection head;

[0010] S3. The key hyperparameters of the YOLOV8 network are optimized by adopting an improved artificial bee colony algorithm. The improved artificial bee colony algorithm is as follows: the search range of the hired bees is improved and the follower bees are removed to simplify the colony structure. A hired bee candidate mechanism is designed, a hired bee last-place elimination mechanism is designed, a dynamic probabilistic elite retention mechanism is designed, a scout bee multi-scale mutation search mechanism is designed, and a dynamic colony number adjustment strategy is proposed.

[0011] S4. Organize the optimal solution and update the YOLOV8 hyperparameter file: Update the optimal hyperparameters to the YOLOV8 hyperparameter file. The hyperparameters include: learning rate (lr), weight decay coefficient (weight_decay), and momentum. Train the model and test the results of the hyperparameter optimization on the test set.

[0012] S5. Perform a comparison: Compare the mAP50-95 obtained by the improved YOLOV8 on the dataset with that obtained by other optimization methods.

[0013] Furthermore, in S1, high-resolution images of the power tower's vibration damper are acquired by taking multi-angle surround shots using a drone, and a dedicated dataset is constructed. The preprocessing includes adaptive histogram equalization of the images to enhance defect features, providing a high-quality model input foundation for subsequent hyperparameter optimization.

[0014] Furthermore, in S2, the constructed YOLOV8 network is the evaluation object for hyperparameter optimization; the model performance generated by the combined action of its backbone network and neck network is used as the evaluation basis for the fitness function of the improved artificial bee colony algorithm in the form of average accuracy (mAP); the output of the detection head of this network directly determines the gradient direction of the algorithm when searching the hyperparameter space, and is the core of hyperparameter optimization and final detection performance.

[0015] Furthermore, in S3, improvements are made during the hired bee phase: a dynamic bee colony size adjustment strategy is implemented, the exploration range of hired bees is improved, a hired bee candidate mechanism is constructed, and a last-place hired bee elimination mechanism is established. The aim is to ensure strong exploration and development capabilities throughout the early, middle, and late stages, and to accelerate algorithm convergence.

[0016] The iterative formula for hired bees is:

[0017]

[0018] in To employ bees from the current nectar source, To hire bees to find new nectar sources, The search range for hired bees.

[0019] First, a dynamic bee colony size adjustment strategy is implemented, setting the total number of bees N (value 20) and the initial proportion of hired bees. (Value is 0.1) and the final proportion of employed bees (Value 0.9), calculate the number of hired bees based on the iteration number t (value 10). The number of hired bees in the dynamic bee colony number adjustment strategy is as follows:

[0020]

[0021]

[0022] in Indicates the initial proportion of employed bees. This indicates the final percentage of bees employed. This represents the proportion of hired bees to the total number of bees at the t-th iteration, where N is the total number of bees. This indicates the minimum number of bees required to hire a client. This represents the number of mercenary bees in the t-th iteration. Mercenary bees and scout bees collaborate in global exploration in the early stages; they balance exploration and development in the mid-stages; and in the later stages, scout bees continuously enhance the ability to discover historical optimal solution regions, thus more efficiently approaching the global optimum.

[0023] Then, the hired bees will be used for exploration and development. The hired bees will then explore the current nectar source... A domain search is performed to determine the perturbation range, where the perturbation range of the hired bees before pruning should be 20% of the total range in that dimension:

[0024]

[0025] in This represents the range of bee perturbation before the j-th cut. This represents the coefficients of a uniformly random sample in the interval -0.2 to 0.2. =bounds[j][1], representing the upper bound of the j-th parameter. =bounds[j][0] represents the lower bound of the j-th parameter. The final search range for the hired bees should be:

[0026]

[0027] in This indicates the final search range for hired bees, and `clamp` indicates the pruning operation performed on the parameters after perturbation.

[0028]

[0029] in , indicating that The value is rounded to five decimal places to prevent parameters from exceeding the hyperparameter length limit of YOLOv8.

[0030] Secondly, by utilizing the designed mercenary bee candidate mechanism, within the current mercenary bee search range... Three searches are performed, and the fitness results for each search are obtained using the YOLOv8 network:

[0031]

[0032] Where n is the current search count ( ), m is the current m-th hired bee ( ), This represents the number of hired bees in the t-th iteration. This represents the fitness result of the m-th hired bee in the nth search. This represents the parameter of the m-th hired bee in the nth search. This represents the evaluation function of the YOLOv8 network.

[0033] Then, a fitness comparison is performed to obtain the optimal search value for the mercenary bee within its current range:

[0034]

[0035] in This represents the optimal fitness value for the current hired bee within the current search range, where m is the m-th hired bee.

[0036] Then, the fitness of all mercenaries is assessed and ranked using a mercenary bee elimination mechanism, with the lowest-ranked mercenaries being the least fit. One mercenary bee is eliminated, converted into a scout bee, and abandons the nectar source:

[0037]

[0038]

[0039] in This indicates the number of scout bees at the current stage. Indicates the number of scout bees in the next phase. This indicates the minimum number of scout bees (value is 1).

[0040] Finally, the retained honey sources enter the honey source replacement stage. Through optimal replacement, computational resources are continuously focused on promising search regions, improving the depth and accuracy of local searches. If the retained honey source is better than the original, the historical best solution and the best solution group are updated, and the current honey source replaces the original honey source as the source honey source for the next range search, effectively preventing solution quality degradation caused by random perturbations. If the searched honey source is not as good as the original, the current honey source remains unchanged. Incremental optimization ensures that the search always proceeds in the direction of performance improvement, while maintaining the global optimum, historical optimum, and current optimum. It also allows computational resources to automatically focus on promising search regions. While effectively filtering evaluation noise to prevent solution quality degradation and ensure reliability, optimal replacement accelerates the convergence by guiding towards high-quality solution regions.

[0041] Furthermore, in S3, after the mercenary bee phase is completed, the scout bee phase begins. Improvements include: constructing a dynamic probabilistic elite retention mechanism and a multi-scale mutation search mechanism for scout bees. The objectives are: to accelerate convergence speed, improve solution quality, and ensure strong exploration and development capabilities in the mid-to-late stages.

[0042] The formula for scout bees' current nectar source:

[0043]

[0044] in, As the current honey source, For the last honey source, This is the historically optimal solution. This represents the search range when the historical best solution is not chosen as the base solution. The value indicates whether to select the historical best solution (1 for selection, 0 for non-selection).

[0045] The iterative formula for the scout bee is:

[0046]

[0047] in, This indicates that the scout bees are looking for new nectar sources. This is the scale for mutation search.

[0048] First, calculate the number of scout bees. Among them, the number of scout bees :

[0049]

[0050] in, This represents the number of scout bees in the t-th iteration, where N is the total number of bees. This represents the number of hired bees in the t-th iteration. This indicates the minimum number of scout bees (value is 1).

[0051] Then, in the dynamic probability selection mechanism, the probability value of the selection is used. and probability value The decision to use the historical best solution as the current honey source is made through comparison. The dynamic probabilistic elite retention mechanism must calculate probabilities based on the number of iterations, and the probability should increasingly favor the historical best solution as the number of iterations increases. The probability formula for this dynamic elite retention mechanism is:

[0052] .

[0053] in This represents the probability of choosing the historical optimal solution as the base solution in the t-th iteration, where t represents the current iteration number (starting from 0 and ending at T-1), and T represents the maximum number of iterations. The iteration number cannot start from 1 to prevent the first iteration from having a probability other than the minimum 50%. Furthermore, the minimum probability should not be too large to prevent getting stuck in local optima in the early stages. The rules for determining the probability of choosing the historical optimal solution are as follows:

[0054]

[0055] Where r represents the uniformly random sample number in the interval [0, 1), if r is less than the probability of selecting the historical best solution as the base solution. ,but It is 1 if it is 1, otherwise it is 0. If the value is 1, select the historically optimal solution and replace the original honey source. When the value is 0, then select the total range of that dimension. As the perturbation range, a new honey source is randomly selected as the base solution:

[0056]

[0057]

[0058] in As the current honey source, For the last honey source, This represents the search range when the historical best solution is not chosen as the base solution. This represents the coefficients of a uniformly random sample in the interval -0.5 to 0.5. =bounds[j][1], representing the upper bound of the j-th parameter. =bounds[j][0] represents the lower bound of the j-th parameter.

[0059] Finally, the reconnaissance bee multi-scale search mechanism was utilized at different scales. The algorithm makes a maximum of k=4 attempts, comparing the search results with the original results. If the results are better than the original, it enters the update phase, updates the historical best solution and the best solution set, and proceeds to the next step. If the fitness of the searched honey source is not as good as the original, it directly proceeds to the next step.

[0060] The constructed reconnaissance bee multi-scale mutation search mechanism employs a four-step decreasing search range method, with the following search scales:

[0061]

[0062] Where k represents the number of attempts, with values ​​ranging from 0 to 3. Let t represent the mutation scale of the k-th attempt, and t represent the current iteration number, starting from 0 to 1. Finish, This indicates the maximum number of iterations. As the number of iterations increases... As it gets larger, with the same number of attempts The value of increases, but as the number of attempts k increases, The number of iterations decreases further with each iteration, while the `max` function ensures the minimum value will not be less than 0.1, guaranteeing that the scout bee's exploration range will not be less than that of the mercenary bee, thus ensuring strong overall exploration capabilities. The number of search attempts is limited to four to prevent excessively long attempts due to an excessively large scale, and to prevent an insufficiently refined search due to an excessively small scale. The scout bee's multi-scale search mechanism avoids both the lack of fine-grained search capabilities at a single large scale and the lack of exploration capabilities at a single small scale. Through multi-scale mutation search, it covers high-quality regions near the current solution with a high probability, ensuring both strong exploration capabilities and good depth development capabilities.

[0063] Furthermore, S4 includes a parameter update and model reconstruction process: the optimal hyperparameter combination obtained from the improved artificial bee colony algorithm is updated to the YOLOV8 model's configuration file, and a new round of training is performed. After training is completed, the system automatically evaluates the performance metrics of the final model on the reserved test set, thereby verifying the effectiveness of the hyperparameter optimization algorithm.

[0064] Furthermore, in S5, the performance comparison is conducted under identical experimental conditions to ensure fairness. The comparison objects include optimized YOLOv8 models using methods such as grid search, genetic algorithm, and standard artificial bee colony algorithm. The mean average accuracy (mAP50-95) is used as the core performance indicator to quantitatively compare the optimization results of different methods.

[0065] Compared with the prior art, the advantages of the present invention are as follows:

[0066] This invention effectively solves the problems of insufficient late-stage development, premature convergence, and slow search speed in the standard artificial bee colony algorithm, significantly improving algorithm performance. Its main advantages include: 1. Employing a dynamic bee colony size adjustment strategy to adaptively adjust the ratio of exploration to development in different iteration stages, ensuring both early-stage global exploration and enhanced local discovery capabilities in later stages; 2. Introducing a dynamic probabilistic elite retention mechanism to accelerate convergence while maintaining population diversity, effectively preventing the loss of high-quality solutions and premature convergence; 3. Designing a multi-scale mutation search mechanism for scout bees, which balances the ability to escape local optima and finely approximate the optimal solution through scale changes from large to small; 4. Improving the behavior of hired bees and simplifying the bee colony structure, thereby enhancing overall search efficiency. These mechanisms collectively improve the algorithm's convergence speed and optimization accuracy.

[0067] The algorithm proposed in this paper is evaluated on the anti-vibration hammer defect dataset using the mAP50-95 index. The mAP50-95 reaches 0.81612 with 7 iterations, while the traditional standard ABC algorithm has a mAP50-95 value of 0.78813 with 10 iterations. Compared with the traditional standard ABC algorithm, the algorithm proposed in this paper has a faster training convergence speed and better model performance. Attached Figure Description

[0068] Figure 1 is an algorithm flowchart of a method for detecting defects in power tower vibration dampers based on hyperparameter optimization of multi-scale dynamic artificial bee colony algorithm in an embodiment of the present invention.

[0069] Figure 2 is a flowchart of the reconnaissance bee dynamic probability elite retention mechanism of a power tower anti-vibration hammer defect detection method based on hyperparameter optimization of multi-scale dynamic artificial bee colony algorithm in an embodiment of the present invention.

[0070] Figure 3 is a flowchart of the multi-scale mutation search mechanism of the reconnaissance bee in a method for detecting damage to power tower vibration dampers based on hyperparameter optimization of a multi-scale dynamic artificial bee colony algorithm in an embodiment of the present invention. Detailed Implementation

[0071] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments described in this application and the specific technical features in each embodiment are illustrative of the technical solution of this application, and not a limitation on the scope of protection of the technical solution of this application. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0072] In the description of this invention, the term "several" means one or more, and "multiple" means two or more. The terms "greater than," "less than," and "exceeding" are understood to exclude the number itself; while "above," "below," and "within" are understood to include the number. If ordinal numbers such as "first" and "second" appear in the description, their purpose is solely to distinguish different technical features and should not be construed as implying or indicating a relative importance, a limitation on quantity, or a limitation on the order of execution of the features they define.

[0073] In this specification, the use of terms such as "one embodiment," "some embodiments," "example," or "specific example," etc., indicates that the specific structure, method, or feature described in connection with that embodiment or example is included in at least one embodiment or implementation of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to completely identical embodiments or examples. Furthermore, the specific structures, methods, or features described in different embodiments or examples can be combined in any non-exclusive way to form other implementations that can be understood by those skilled in the art.

[0074] Example 1

[0075] like Figure 1 As shown, a method for detecting defects in power tower vibration dampers based on hyperparameter optimization using a multi-scale dynamic artificial bee colony algorithm includes the following steps:

[0076] S1. Collect images of vibration dampers on power towers using drones, construct a dataset, and complete data annotation and preprocessing;

[0077] In S1, high-resolution images of the power tower's vibration damper are acquired by taking multi-angle surround shots using a drone, and a dedicated dataset is constructed. The preprocessing includes adaptive histogram equalization of the images to enhance defect features, providing a high-quality model input foundation for subsequent hyperparameter optimization.

[0078] S2. Construct a YOLOV8 target detection network, extract multi-scale features through the backbone network, fuse features of different scales through the neck network, and output the localization and classification results of the shock absorber through the detection head;

[0079] In S2, the constructed YOLOV8 network is the evaluation object for hyperparameter optimization; the model performance generated by the combined action of its backbone network and neck network is used as the evaluation basis for the fitness function of the improved artificial bee colony algorithm in the form of average accuracy (mAP); the output of the detection head of this network directly determines the gradient direction of the algorithm when searching the hyperparameter space, and is the core of hyperparameter optimization and final detection performance.

[0080] S3. The key hyperparameters of the YOLOV8 network are optimized by adopting an improved artificial bee colony algorithm. The improved artificial bee colony algorithm is as follows: the search range of the hired bees is improved and the follower bees are removed to simplify the colony structure. A hired bee candidate mechanism is designed, a hired bee last-place elimination mechanism is designed, a dynamic probabilistic elite retention mechanism is designed, a scout bee multi-scale mutation search mechanism is designed, and a dynamic colony number adjustment strategy is proposed.

[0081] In S3, improvements during the hired bee phase include: constructing a dynamic bee colony size adjustment strategy, improving the exploration range of hired bees, constructing a hired bee candidate mechanism, and constructing a hired bee last-place elimination mechanism. The aim is to ensure strong exploration and development capabilities in the early, middle, and late stages, and to accelerate the algorithm's convergence speed.

[0082] The iterative formula for hired bees is:

[0083]

[0084] in To employ bees from the current nectar source, To hire bees to find new nectar sources, The search range for hired bees.

[0085] First, a dynamic bee colony size adjustment strategy is implemented, setting the total number of bees N (value 20) and the initial proportion of hired bees. (Value is 0.1) and the final proportion of employed bees (Value 0.9), calculate the number of hired bees based on the iteration number t (value 10). The number of hired bees in the dynamic bee colony number adjustment strategy is as follows:

[0086]

[0087]

[0088] in Indicates the initial proportion of employed bees. This indicates the final percentage of bees employed. This represents the proportion of hired bees to the total number of bees at the t-th iteration, where N is the total number of bees. This indicates the minimum number of bees required to hire a client. This represents the number of mercenary bees in the t-th iteration. Mercenary bees and scout bees collaborate in global exploration in the early stages; they balance exploration and development in the mid-stages; and in the later stages, scout bees continuously enhance the ability to discover historical optimal solution regions, thus more efficiently approaching the global optimum.

[0089] Then, the hired bees will be used for exploration and development. The hired bees will then explore the current nectar source... A domain search is performed to determine the perturbation range, where the perturbation range of the hired bees before pruning should be 20% of the total range in that dimension:

[0090]

[0091] in This represents the range of bee perturbation before the j-th cut. This represents the coefficients of a uniformly random sample in the interval -0.2 to 0.2. =bounds[j][1], representing the upper bound of the j-th parameter. =bounds[j][0] represents the lower bound of the j-th parameter. The final search range for the hired bees should be:

[0092]

[0093] in This indicates the final search range for hired bees, and `clamp` indicates the pruning operation performed on the parameters after perturbation.

[0094]

[0095] in , indicating that The value is rounded to five decimal places to prevent parameters from exceeding the hyperparameter length limit of YOLOv8.

[0096] Secondly, by utilizing the designed mercenary bee candidate mechanism, within the current mercenary bee search range... Three searches are performed, and the fitness results for each search are obtained using the YOLOv8 network:

[0097]

[0098] Where n is the current search count ( ), m is the current m-th hired bee ( ), This represents the number of hired bees in the t-th iteration. This represents the fitness result of the m-th hired bee in the nth search. This represents the parameter of the m-th hired bee in the nth search. This represents the evaluation function of the YOLOv8 network.

[0099] Then, a fitness comparison is performed to obtain the optimal search value for the mercenary bee within its current range:

[0100]

[0101] in This represents the optimal fitness value for the current hired bee within the current search range, where m is the m-th hired bee.

[0102] Then, the fitness of all mercenaries is assessed and ranked using a mercenary bee elimination mechanism, with the lowest-ranked mercenaries being the least fit. One mercenary bee is eliminated, converted into a scout bee, and abandons the nectar source:

[0103]

[0104]

[0105] in This indicates the number of scout bees at the current stage. Indicates the number of scout bees in the next phase. This indicates the minimum number of scout bees (value is 1).

[0106] Finally, the retained honey sources enter the honey source replacement stage. Through optimal replacement, computational resources are continuously focused on promising search regions, improving the depth and accuracy of local searches. If the retained honey source is better than the original, the historical best solution and the best solution group are updated, and the current honey source replaces the original honey source as the source honey source for the next range search, effectively preventing solution quality degradation caused by random perturbations. If the searched honey source is not as good as the original, the current honey source remains unchanged. Incremental optimization ensures that the search always proceeds in the direction of performance improvement, while maintaining the global optimum, historical optimum, and current optimum. It also allows computational resources to automatically focus on promising search regions. While effectively filtering evaluation noise to prevent solution quality degradation and ensure reliability, optimal replacement accelerates the convergence by guiding towards high-quality solution regions.

[0107] In S3, after the mercenary bee phase is completed, the scout bee phase begins. Improvements include: constructing a dynamic probabilistic elite retention mechanism and a multi-scale mutation search mechanism for scout bees. The objectives are: to accelerate convergence speed, improve solution quality, and ensure strong exploration and development capabilities in the mid-to-late stages.

[0108] The formula for scout bees' current nectar source:

[0109]

[0110] in, As the current honey source, For the last honey source, This is the historically optimal solution. This represents the search range when the historical best solution is not chosen as the base solution. The value indicates whether to select the historical best solution (1 for selection, 0 for non-selection).

[0111] The iterative formula for the scout bee is:

[0112]

[0113] in, This indicates that the scout bees are looking for new nectar sources. This is the scale for mutation search.

[0114] First, calculate the number of scout bees. Among them, the number of scout bees :

[0115]

[0116] in, This represents the number of scout bees in the t-th iteration, where N is the total number of bees. This represents the number of hired bees in the t-th iteration. This indicates the minimum number of scout bees (value is 1).

[0117] Then, in the dynamic probability selection mechanism, the probability value of the selection is used. and probability value The decision to use the historical best solution as the current honey source is made through comparison. The dynamic probabilistic elite retention mechanism must calculate probabilities based on the number of iterations, and the probability should increasingly favor the historical best solution as the number of iterations increases. The probability formula for this dynamic elite retention mechanism is:

[0118] .

[0119] in This represents the probability of choosing the historical optimal solution as the base solution in the t-th iteration, where t represents the current iteration number (starting from 0 and ending at T-1), and T represents the maximum number of iterations. The iteration number cannot start from 1 to prevent the first iteration from having a probability other than the minimum 50%. Furthermore, the minimum probability should not be too large to prevent getting stuck in local optima in the early stages. The rules for determining the probability of choosing the historical optimal solution are as follows:

[0120]

[0121] Where r represents the uniformly random sample number in the interval [0, 1), if r is less than the probability of selecting the historical best solution as the base solution. ,but It is 1 if it is 1, otherwise it is 0. If the value is 1, select the historically optimal solution and replace the original honey source. When the value is 0, then select the total range of that dimension. As the perturbation range, a new honey source is randomly selected as the base solution:

[0122]

[0123]

[0124] in As the current honey source, For the last honey source, This represents the search range when the historical best solution is not chosen as the base solution. This represents the coefficients of a uniformly random sample in the interval -0.5 to 0.5. =bounds[j][1], representing the upper bound of the j-th parameter. =bounds[j][0] represents the lower bound of the j-th parameter.

[0125] Finally, the reconnaissance bee multi-scale search mechanism was utilized at different scales. The algorithm makes a maximum of k=4 attempts, comparing the search results with the original results. If the results are better than the original, it enters the update phase, updates the historical best solution and the best solution set, and proceeds to the next step. If the fitness of the searched honey source is not as good as the original, it directly proceeds to the next step.

[0126] The constructed reconnaissance bee multi-scale mutation search mechanism employs a four-step decreasing search range method, with the following search scales:

[0127]

[0128] Where k represents the number of attempts, with values ​​ranging from 0 to 3. Let t represent the mutation scale of the k-th attempt, and t represent the current iteration number, starting from 0 to 1. Finish, This indicates the maximum number of iterations. As the number of iterations increases... As it gets larger, with the same number of attempts The value of increases, but as the number of attempts k increases, The number of iterations decreases further with each iteration, while the `max` function ensures the minimum value will not be less than 0.1, guaranteeing that the scout bee's exploration range will not be less than that of the mercenary bee, thus ensuring strong overall exploration capabilities. The number of search attempts is limited to four to prevent excessively long attempts due to an excessively large scale, and to prevent an insufficiently refined search due to an excessively small scale. The scout bee's multi-scale search mechanism avoids both the lack of fine-grained search capabilities at a single large scale and the lack of exploration capabilities at a single small scale. Through multi-scale mutation search, it covers high-quality regions near the current solution with a high probability, ensuring both strong exploration capabilities and good depth development capabilities.

[0129] S4. Organize the optimal solution and update the YOLOV8 hyperparameter file: Update the optimal hyperparameters to the YOLOV8 hyperparameter file. The hyperparameters include: learning rate (lr), weight decay coefficient (weight_decay), and momentum. Train the model and test the results of the hyperparameter optimization on the test set.

[0130] S4 includes parameter update and model reconstruction processes: the optimal hyperparameter combination obtained by the improved artificial bee colony algorithm is updated to the YOLOV8 model configuration file, and a new round of training is carried out. After training is completed, the system automatically evaluates the performance index of the final model on the reserved test set, thereby verifying the effectiveness of the hyperparameter optimization algorithm.

[0131] S5. Perform a comparison: Compare the mAP50-95 obtained by the improved YOLOV8 on the dataset with that obtained by other optimization methods;

[0132] In S5, the performance comparison is conducted under identical experimental conditions to ensure fairness. The comparison objects include optimized YOLOv8 models using methods such as grid search, genetic algorithm, and standard artificial bee colony algorithm. The mean average accuracy (mAP50-95) is used as the core performance indicator to quantitatively compare the optimization results of different methods.

[0133] Example 2

[0134] This embodiment provides a method and device for detecting defects in power tower vibration dampers based on hyperparameter optimization using a multi-scale dynamic artificial bee colony algorithm, including a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the method according to Embodiment 1.

[0135] Example 3

[0136] This embodiment provides a storage medium on which a computer program is stored. When executed by a processor, the program implements the steps of the method described in Embodiment 1. The program instructions can also be loaded into a programmable device to achieve the functions specified in the flowchart and / or block diagram through a series of executable operations.

[0137] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features.

Claims

1. A method for detecting defects in power tower vibration dampers based on hyperparameter optimization using a multi-scale dynamic artificial bee colony algorithm, characterized in that, Includes the following steps: S1. Collect images of vibration dampers on power towers using drones, construct a dataset, and complete data annotation and preprocessing; S2. Construct a YOLOV8 target detection network, extract multi-scale features through the backbone network, fuse features of different scales through the neck network, and output the localization and classification results of the shock absorber through the detection head; S3. The key hyperparameters of the YOLOV8 network are optimized by adopting an improved artificial bee colony algorithm, the search range of the hired bees is improved and the follower bees are removed to simplify the colony structure, a hired bee candidate mechanism is designed, a hired bee last-place elimination mechanism is designed, a dynamic probability elite retention mechanism is designed, a scout bee multi-scale mutation search mechanism is designed, and a dynamic colony number adjustment strategy is proposed. S4. Organize the optimal solution and update the YOLOV8 hyperparameter file: Update the optimal hyperparameters to the YOLOV8 hyperparameter file. The hyperparameters include: learning rate (lr), weight decay coefficient (weight_decay), and momentum. Train the model and test the results of the hyperparameter optimization on the test set. S5. Perform a comparison: Compare the mAP50-95 obtained by the improved YOLOV8 on the dataset with the results obtained by other optimization methods.

2. The method for detecting damage to power tower vibration dampers based on hyperparameter optimization using a multi-scale dynamic artificial bee colony algorithm according to claim 1, characterized in that, In S1, high-resolution images of the power tower's vibration damper are acquired by taking multi-angle surround shots using a drone, and a dedicated dataset is constructed. The preprocessing includes adaptive histogram equalization of the images to enhance defect features, providing a high-quality model input foundation for subsequent hyperparameter optimization.

3. The method for detecting damage to power tower vibration dampers based on multi-scale dynamic artificial bee colony algorithm hyperparameter optimization according to claim 1, characterized in that, In S2, the constructed YOLOV8 network is the evaluation object for hyperparameter optimization; the model performance generated by the combined effect of its backbone network and neck network is used as the evaluation basis for the fitness function of the improved artificial bee colony algorithm in the form of average accuracy (mAP).

4. The method for detecting damage to power tower vibration dampers based on hyperparameter optimization using a multi-scale dynamic artificial bee colony algorithm according to claim 1, characterized in that, In S3, during the mercenary bee phase, a dynamic bee colony size adjustment strategy is implemented, the exploration range of mercenary bees is improved, a mercenary bee candidate mechanism is constructed, and a mercenary bee last-place elimination mechanism is established. The iterative formula for hired bees is: in To employ bees from the current nectar source, To hire bees to find new nectar sources, To define the search area for hired bees, First, a dynamic bee colony size adjustment strategy is implemented, setting the total number of bees N (value 20) and the initial proportion of hired bees. (Value is 0.1) and the final proportion of employed bees (Value 0.9), calculate the number of hired bees based on the iteration number t (value 10). The number of hired bees in the dynamic bee colony number adjustment strategy is as follows: in Indicates the initial proportion of employed bees. This indicates the final percentage of bees employed. This represents the proportion of hired bees to the total number of bees at the t-th iteration, where N is the total number of bees. This indicates the minimum number of bees required to hire a client. This represents the number of hired bees in the t-th iteration. Hired bees and scout bees collaborate in the early stages of global exploration; in the mid-stage, they balance exploration and development; and in the later stages, they leverage the scout bees to continuously enhance their ability to discover historical optimal solution regions, thereby more efficiently approaching the global optimum. Then, the hired bees are used for exploration and development, and they explore the current nectar source using... A domain search is performed to determine the perturbation range, where the perturbation range of the hired bees before pruning should be 20% of the total range in that dimension: in This represents the range of bee perturbation before the j-th cut. This represents the coefficients of a uniformly random sample in the interval -0.2 to 0.

2. =bounds[j][1], representing the upper bound of the j-th parameter. =bounds[j][0] represents the lower bound of the j-th parameter. The final search range for the hired bee should be: in This indicates the final search range for hired bees, and `clamp` indicates the pruning operation performed on the parameters after perturbation. in , indicating that The value is rounded to five decimal places to prevent parameters from exceeding the hyperparameter length limit of YOLOv8. Secondly, by utilizing the designed mercenary bee candidate mechanism, within the current mercenary bee search range... Three searches are performed, and the fitness results for each search are obtained using the YOLOv8 network: Where n is the current search count ( ), m is the current m-th hired bee ( ), This represents the number of hired bees in the t-th iteration. This represents the fitness result of the m-th hired bee in the nth search. This represents the parameter of the m-th hired bee in the nth search. This represents the evaluation function of the YOLOv8 network. Then, a fitness comparison is performed to obtain the optimal search value for the mercenary bee within its current range: in This represents the optimal fitness value for the current hired bee within the current search range, where m is the m-th hired bee. Then, the fitness of all mercenaries is assessed and ranked using a mercenary bee elimination mechanism, with the lowest-ranked mercenaries being the least fit. One mercenary bee is eliminated, converted into a scout bee, and abandons the nectar source: in This indicates the number of scout bees at the current stage. Indicates the number of scout bees in the next phase. This indicates the minimum number of scout bees (value is 1). Finally, the retained honey sources enter the honey source replacement stage. Through optimal replacement, computing resources are continuously focused on promising search regions, improving the depth and accuracy of local searches. If the retained searched honey source is better than the original, the historical best solution and the best solution group are updated, and the current honey source replaces the original honey source as the source honey source for the next range search, effectively preventing solution quality degradation caused by random perturbations. If the searched honey source is not as good as the original, the current honey source is not changed. Progressive optimization ensures that the search always proceeds in the direction of performance improvement, while maintaining the global optimum, historical optimum, and current optimum. This also allows computing resources to automatically focus on promising search regions. While effectively filtering evaluation noise to prevent solution quality degradation and ensure reliability, optimal replacement accelerates the convergence by guiding towards high-quality solution regions.

5. The method for detecting damage to power tower vibration dampers based on multi-scale dynamic artificial bee colony algorithm hyperparameter optimization according to claim 1, characterized in that, In S3, after the mercenary bee phase is completed, the scout bee phase begins. A dynamic probabilistic elite retention mechanism and a multi-scale mutation search mechanism for scout bees are constructed. The current nectar source formula for scout bees is as follows: in, As the current honey source, For the last honey source, This is the historically optimal solution. This represents the search range when the historical best solution is not chosen as the base solution. The value indicates whether to select the historical best solution (1 for selection, 0 for non-selection). The iterative formula for the scout bee is: in, This indicates that the scout bees are looking for new nectar sources. As the scale for mutation search, First, calculate the number of scout bees. Among them, the number of scout bees : in, This represents the number of scout bees in the t-th iteration, where N is the total number of bees. This represents the number of hired bees in the t-th iteration. This indicates the minimum number of scout bees (value is 1). Then, in the dynamic probability selection mechanism, the probability value of the selection is used. and probability value The comparison determines whether to use the historical best solution as the current honey source. The dynamic probabilistic elite retention mechanism must calculate probabilities based on the number of iterations, and the probability should increasingly favor the historical best solution as the number of iterations increases. The probability formula for this dynamic elite retention mechanism is: ， in This represents the probability of choosing the historical optimal solution as the base solution in the t-th iteration, where t represents the current iteration number, starting from 0 and ending at T-1. T represents the maximum number of iterations. The iteration number cannot start from 1 to prevent the probability from not being the minimum 50% in the first iteration. Furthermore, the minimum probability should not be too large to prevent getting stuck in local optima in the early stages. The rules for determining the probability of choosing the historical optimal solution are as follows: Where r represents the uniformly random sample number in the interval [0, 1), if r is less than the probability of selecting the historical best solution as the base solution. ,but It is 1 if it is 1, otherwise it is 0. If the value is 1, select the historically optimal solution and replace the original honey source. When the value is 0, then select the total range of that dimension. As the perturbation range, a new honey source is randomly selected as the base solution: in As the current honey source, For the last honey source, This represents the search range when the historical best solution is not chosen as the base solution. This represents the coefficients of a uniformly random sample in the interval -0.5 to 0.

5. =bounds[j][1], representing the upper bound of the j-th parameter. =bounds[j][0], which represents the lower bound of the j-th parameter. Finally, the reconnaissance bee multi-scale search mechanism was utilized at different scales. The algorithm performs a maximum of k=4 attempts, comparing the search results with the original results. If the search result is better than the original, it proceeds to the update phase, updating the historical best solution and the best solution set, and then proceeds to the next step. If the fitness of the searched honey source is not as good as the original, it directly proceeds to the next step. The constructed reconnaissance bee multi-scale mutation search mechanism employs a four-step decreasing search range method, with the following search scales: Where k represents the number of attempts, with values ​​ranging from 0 to 3. Let t represent the mutation scale of the k-th attempt, and t represent the current iteration number, starting from 0 to 1. Finish, This represents the maximum number of iterations; as the number of iterations increases... As it gets larger, with the same number of attempts The value of increases, but as the number of attempts k increases, The number of iterations decreases further with each iteration, while the max function ensures that the minimum value will not be less than 0.1, ensuring that the scout bee's exploration range will not be less than that of the mercenary bee, thus guaranteeing strong overall exploration capabilities. The number of search attempts is limited to 4 to prevent excessively large attempts from taking too long and excessively small attempts from being too imprecise. The scout bee's multi-scale search mechanism avoids both the problem of lacking fine search capabilities at a single large scale and the problem of lacking exploration capabilities at a single small scale. Through multi-scale mutation search, it covers high-quality areas near the current solution with a high probability, ensuring good depth development capabilities while possessing high exploration capabilities.

6. The method for detecting damage to power tower vibration dampers based on hyperparameter optimization using a multi-scale dynamic artificial bee colony algorithm according to claim 1, characterized in that, S4 includes parameter update and model reconstruction processes: the optimal hyperparameter combination obtained by the improved artificial bee colony algorithm is updated to the YOLOV8 model configuration file, and a new round of training is carried out. After training is completed, the system automatically evaluates the performance index of the final model on the reserved test set, thereby verifying the effectiveness of the hyperparameter optimization algorithm. The method for detecting damage to power tower vibration dampers based on hyperparameter optimization using a multi-scale dynamic artificial bee colony algorithm as described in claim 1 is characterized in that, In S5, the performance comparison is conducted under the same experimental environment to ensure the fairness of the comparison. The comparison objects include optimized YOLOv8 models such as grid search, genetic algorithm, and standard artificial bee colony algorithm. The mean accuracy (mAP50-95) is used as the core performance indicator to quantitatively compare the optimization results of different methods.

Images