An IGA-AdaBoost-SVM model-based power line tower defect identification method and system
By using the improved IGA-AdaBoost-SVM model and UAV multispectral imaging technology, the efficiency and accuracy problems of traditional power line tower defect detection have been solved, realizing automated and intelligent defect identification and trend analysis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- THREE GORGES NEW ENERGY ULANQAB CO LTD
- Filing Date
- 2025-08-26
- Publication Date
- 2026-08-04
AI Technical Summary
Traditional power line tower defect detection relies on manual inspection, which is inefficient. Single SVM models have low recognition accuracy, while CNN models require massive amounts of data and have poor adaptability, making it difficult to adapt to different regions and defect types.
An improved genetic algorithm (IGA) is used to optimize the AdaBoost-SVM model. Combined with UAV multispectral imaging technology, a database of power line tower defects is constructed. Defects are identified through path planning and image preprocessing, and trend analysis and operation and maintenance suggestions are generated.
It achieves efficient and accurate identification and automated detection of defects in power line towers, improves identification accuracy and generalization ability, ensures reliable model performance, and has self-diagnosis and optimization capabilities.
Smart Images
Figure CN121121548B_ABST
Abstract
Description
Technical Field
[0001] This disclosure belongs to the field of power equipment operation and maintenance technology, and in particular relates to a method and system for identifying defects in power line towers based on the IGA-AdaBoost-SVM model. Background Technology
[0002] As the core supporting facilities of the power transmission network, power line towers are exposed to the complex outdoor environment for a long time and are susceptible to various defects caused by natural factors, human factors and equipment aging. These defects can lead to large-scale power outages and may also cause serious consequences such as equipment damage and personal injury. Therefore, timely and accurate identification of tower defects is crucial to ensuring the safe and stable operation of the power grid.
[0003] Traditional power line tower defect detection mainly relies on manual inspection, requiring maintenance personnel to climb towers or patrol the line on foot, a method with significant limitations. With the development of drone technology and artificial intelligence, a "drone photography + image recognition" detection model has begun to be explored. However, most solutions use a single SVM (Support Vector Machine) or CNN (Convolutional Neural Network) model. A single SVM model has weak classification capabilities for multiple defect types and is susceptible to imbalanced samples, resulting in a defect identification accuracy of less than 70% for small samples. While CNN models have strong feature extraction capabilities, they require massive amounts of labeled data, and model parameter optimization relies on human experience, making it difficult to adapt to detection scenarios with different regions and defect types.
[0004] Therefore, there is an urgent need for a method for identifying defects in power line towers that can overcome the above-mentioned shortcomings and improve the accuracy and generalization ability of identification. Summary of the Invention
[0005] To address the aforementioned issues, this disclosure provides a method and system for identifying power line tower defects based on the IGA-AdaBoost-SVM model. It employs an improved genetic algorithm (IGA) to optimize the AdaBoost-SVM model parameters and combines this with UAV multispectral imaging technology to achieve efficient and accurate identification of power line tower defects. Furthermore, it provides trend analysis and maintenance recommendations, effectively improving the accuracy of power line maintenance.
[0006] Firstly, this disclosure provides a method for identifying defects in power line towers based on the IGA-AdaBoost-SVM model, the method comprising: Collect historical defect images of several power line towers and their corresponding historical data, and integrate them into a power line tower defect database. An initial IGA-AdaBoost-SVM model is constructed and validated based on the power line tower defect database. The initial IGA-AdaBoost-SVM model is adjusted based on the validation results to obtain the target IGA-AdaBoost-SVM model, and the power line tower defect database is updated in real time. The target shooting path is planned based on the actual location data of power line towers using drones. Based on the target shooting path, multispectral imaging technology is used to collect real-time tower images and perform preprocessing to obtain actual tower images. Based on the actual tower images and the target IGA-AdaBoost-SVM model, the actual defect type, actual defect location, and actual severity are identified, and a real-time report containing trend analysis and maintenance suggestions is generated.
[0007] Furthermore, Collect historical defect images of several power line towers and their corresponding historical correlation data, specifically including: Determine the types of historical defects corresponding to power line towers; Based on the historical inspection database, determine the inspection defect images and their corresponding inspection association data for each historical defect type. Based on sensor acquisition, defect images and their corresponding associated data for each historical defect type are collected. By comparing the inspection defect image and the collected defect image, images with a similarity greater than a similarity threshold are integrated. The inspection defect images and their corresponding inspection-related data corresponding to each historical defect type, as well as the collected defect images and their corresponding collected-related data, are compiled into the power line tower defect database.
[0008] Furthermore, An initial IGA-AdaBoost-SVM model is constructed based on the aforementioned power line tower defect database, specifically including: The AdaBoost algorithm is used to integrate multiple SVMs. Initially, each defect sample in the power line tower defect database is assigned equal weight. After training the first SVM weak classifier, the sample weights are adjusted according to its classification error. Then, the next SVM weak classifier is trained. This process is repeated until a preset number of weak classifiers are generated. Finally, corresponding weights are assigned according to the classification accuracy of each weak classifier to construct the AdaBoost-SVM model. The key parameters affecting the accuracy of power line tower defect identification in the AdaBoost-SVM model are used as gene segments of a chromosome, and the gene segments are encoded into a chromosome using binary encoding. Based on the actual needs of power line tower defect identification, the population size is set, and the binary code of each individual is randomly generated within a reasonable range of values for each parameter, thus obtaining the chromosome population. The population size is iteratively compared based on the fitness of any two individuals, and crossover and mutation operations are performed based on the comparison results to determine the improved genetic algorithm. The chromosome population is input into the improved genetic algorithm. The fitness value is calculated by measuring the recognition accuracy of the AdaBoost-SVM model corresponding to each chromosome on the validation set of the power line tower defect database. Selection, crossover, and mutation operations are performed based on the fitness value to iteratively update the chromosome population until a preset number of iterations is reached or the fitness value tends to stabilize. At this point, the model parameters determined by the chromosome with the highest fitness are used to construct the initial IGA-AdaBoost-SVM model.
[0009] Furthermore, The initial IGA-AdaBoost-SVM model was adjusted based on the validation results, specifically including: The power line tower defect database is divided into training set, validation set and test set based on a preset ratio; After training the initial IGA-AdaBoost-SVM model using the training set, the model is validated using the validation set, and the defect identification accuracy, false negative rate, and false positive rate of the model on the validation set are calculated. If the model's recognition accuracy is lower than the preset accuracy threshold, or the false negative rate is higher than the preset false negative rate threshold, or the false positive rate is higher than the preset false positive rate threshold, then analyze the cause of the error. If the parameter settings are unreasonable, return to the improved genetic algorithm to re-optimize the model parameters and adjust the algorithm parameters such as population size and number of iterations; If the database is insufficient or contains biases, supplement the data with relevant defective data, update the database, and then retrain the model. Repeat the above process until the model's performance metrics on the validation set meet the preset requirements, and obtain the target IGA-AdaBoost-SVM model.
[0010] Furthermore, The target shooting path is planned based on the actual location data of power line towers using drones, specifically including: Obtain the actual location data of power line towers; Based on the actual location data and the UAV's flight parameters, a path planning algorithm is used to plan the initial shooting path; The initial shooting path is adjusted in real time based on the real-time environmental conditions to obtain the target shooting path.
[0011] Furthermore, Based on the target shooting path, multispectral imaging technology is used to acquire real-time tower images and perform preprocessing, specifically including: Based on the target shooting path, multispectral imaging sensors are used to acquire multi-band image data of power line towers; Image registration and fusion processing are performed on the acquired image data of multiple bands to generate high-resolution multispectral images; The multispectral image is preprocessed by denoising, enhancement and segmentation to extract the main structural region of the tower and obtain the actual tower image.
[0012] Furthermore, Based on the actual tower images and the target IGA-AdaBoost-SVM model, the actual defect type, actual defect location, and actual severity are identified, specifically including: The actual tower image is input into the target IGA-AdaBoost-SVM model, and the model outputs the probability distribution of defect types. The actual defect type is determined based on the probability distribution, and the actual defect location on the tower is identified using image localization technology. The actual severity is assessed based on the characteristic parameters of the defect type and the confidence level of the model output.
[0013] Furthermore, Generate real-time reports containing trend analysis and operational recommendations, specifically including: Based on the historical defect database and current identification results, time series analysis is conducted to predict defect development trends; Based on the defect type, location, and severity, and combined with the operation and maintenance rule base, targeted operation and maintenance suggestions are generated; The system integrates identification results, trend analysis, and operation and maintenance recommendations to generate a structured report, which is then transmitted to the operation and maintenance platform in real time via the network.
[0014] Furthermore, The power line tower defect database is updated in real time, specifically including: Store actual tower images, identified defect types, locations, severity, and maintenance reports in the database; Regularly obtain feedback data after defect handling from the operation and maintenance platform, and update the tags and associated data of defect samples; The target IGA-AdaBoost-SVM model was retrained based on the new data to maintain the model's recognition accuracy.
[0015] Secondly, based on the same inventive concept, this disclosure provides a power line tower defect identification system based on the IGA-AdaBoost-SVM model, the system comprising: The database construction module is used to collect historical defect images of several power line towers and their corresponding historical related data, and integrate them into a power line tower defect database. The model building module is used to build an initial IGA-AdaBoost-SVM model based on the power line tower defect database and verify it, adjust the initial IGA-AdaBoost-SVM model based on the verification results, obtain the target IGA-AdaBoost-SVM model, and update the power line tower defect database in real time. The data acquisition module is used to plan the target shooting path based on the actual location data of the power line towers combined with the drone. Based on the target shooting path, multispectral imaging technology is used to acquire real-time tower images and perform preprocessing to obtain actual tower images. The report generation module is used to identify the actual defect type, actual defect location and actual severity based on the actual tower image and the target IGA-AdaBoost-SVM model, and generate a real-time report containing trend analysis and operation and maintenance suggestions.
[0016] Compared with the prior art, this disclosure has the following advantages: 1. This disclosure constructs a complete technical closed loop from data preparation, model optimization, on-site collection to intelligent identification and decision-making, deeply integrating artificial intelligence, UAV remote sensing and power operation and maintenance, completely changing the traditional manual operation mode, and realizing the systematization, automation and intelligence of power pole defect identification.
[0017] 2. This disclosure integrates historical inspection data and real-time sensor data, and performs similarity verification to construct a database with a large sample size, high data quality, and high consistency. The introduction of related data (such as material, time, and environmental information) provides the model with multi-dimensional features, making subsequent identification and trend analysis more accurate and reliable.
[0018] 3. This disclosure improves the use of the Integral Genetic Algorithm (IGA) to optimize the parameters of the AdaBoost-SVM ensemble model. IGA's global search capability overcomes the shortcomings of traditional parameter optimization, which is prone to getting trapped in local optima. This results in the final model parameter combination being globally optimal or nearly optimal, thereby greatly improving the model's accuracy in identifying complex defect features and its generalization ability in different scenarios, while reducing the risk of overfitting.
[0019] 4. This disclosure ensures the reliability of the final deployed model through a rigorous “training-verification-testing” layered verification and bidirectional (parameter / data) optimization mechanism, with all indicators meeting stringent engineering application requirements, enabling the model to have the ability to self-diagnose and optimize.
[0020] Other features and advantages of this disclosure will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the disclosure. The objectives and other advantages of this disclosure may be obtained from the structures pointed out in the description, claims, and drawings. Attached Figure Description
[0021] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0022] Figure 1 A flowchart illustrating a method for identifying defects in power line towers based on an IGA-AdaBoost-SVM model according to an embodiment of this disclosure is shown. Figure 2 A flowchart of an improved genetic algorithm according to an embodiment of this disclosure. Detailed Implementation
[0023] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.
[0024] Figure 1 A flowchart illustrating a method for identifying power line tower defects based on an IGA-AdaBoost-SVM model according to an embodiment of this disclosure is shown, as follows: Figure 1 As shown, the power line tower defect identification method based on the IGA-AdaBoost-SVM model in this disclosure includes, S1. Collect historical defect images of several power line towers and their corresponding historical data, and integrate them into a power line tower defect database. In this embodiment of the disclosure, step S1 specifically includes: S11, determine the historical defect type corresponding to the power line tower; S12, Based on the historical inspection database, determine the inspection defect images and their corresponding inspection association data for each historical defect type. S13, based on the sensor, collect the defect images corresponding to each historical defect type and their corresponding associated data; S14, compare the inspection defect image and the collected defect image, and integrate the images with a similarity greater than the similarity threshold; S15, statistically analyze the inspection defect images and their corresponding inspection association data corresponding to each historical defect type, as well as the collected defect images and their corresponding collected association data, to form the power line tower defect database.
[0025] In this disclosure, historical defect types may include loose bolts, missing tower material, worn conductors, misaligned vibration dampers, and interference from external objects (such as bird nests).
[0026] In this embodiment of the disclosure, for the loosening of bolts, photos of different locations, various specifications and different degrees of loosening are collected, along with related data such as bolt material information and installation time; Regarding the missing tower materials, relevant photos were collected, and the model of the tower and the structural characteristics of the surrounding area of the missing tower materials were recorded. If the loss was caused by theft, information such as the time of theft cases in the nearby area and whether other towers in the vicinity had similar situations was also collected. Regarding conductor wear, relevant photos were collected, and information such as conductor tension data and contact methods with tower hardware was added. For the vibration damper misalignment, collect relevant photos, as well as information such as the vibration damper model, installation method, and cable spacing; For bird nests, collect corresponding photos and record information such as the species of bird, the time period of nest construction, and whether there are chicks.
[0027] In this embodiment of the disclosure, sensor technology is used to assist in data collection. Vibration sensors are installed at key parts of the tower to monitor vibration data related to bolt loosening and anti-vibration hammer displacement. Image sensors are installed to collect local images of the tower in real time or periodically to supplement and verify the data.
[0028] S2, construct an initial IGA-AdaBoost-SVM model based on the power line tower defect database and verify it, adjust the initial IGA-AdaBoost-SVM model based on the verification results, obtain the target IGA-AdaBoost-SVM model, and update the power line tower defect database in real time; In this embodiment of the disclosure, step S2 specifically includes: S21. The AdaBoost algorithm is used to integrate multiple SVMs. Initially, each defect sample in the power line tower defect database is assigned equal weight. After training the first SVM weak classifier, the sample weights are adjusted according to its classification error. Then, the next SVM weak classifier is trained. This process is repeated until a preset number of weak classifiers are generated. Finally, the corresponding weights are assigned according to the classification accuracy of each weak classifier to construct the AdaBoost-SVM model. S22, the key parameters affecting the accuracy of power line tower defect identification in the AdaBoost-SVM model are used as gene segments of a chromosome, and the gene segments are encoded into a chromosome using binary encoding. S23. Based on the actual needs of power line tower defect identification, the population size is set, and the binary code of each individual is randomly generated within a reasonable range of values for each parameter, thus obtaining the chromosome population. S24: Perform fitness iterative comparison of the population size based on any two individuals, and perform crossover and mutation operations based on the comparison results to determine the improved genetic algorithm; S25, the chromosome population is input into the improved genetic algorithm. The recognition accuracy of the AdaBoost-SVM model corresponding to each chromosome on the power line tower defect database validation set is calculated as the fitness value. Based on the fitness value, selection, crossover, and mutation operations are performed to iteratively update the chromosome population until a preset number of iterations is reached or the fitness value tends to stabilize. At this time, the model parameters determined by the chromosome with the highest fitness are used to construct the initial IGA-AdaBoost-SVM model.
[0029] Figure 2 For a flowchart of the improved genetic algorithm according to embodiments of this disclosure, see [link to flowchart]. Figure 2 The examples include practical problems: clearly defining the real-world task to be solved (such as function optimization, path planning, etc.), which is the starting point for algorithm application.
[0030] Determine the chromosome that represents the solution to the problem: use chromosomes (usually encoded strings, arrays, etc.) to correspond to the potential solutions to the problem. For example, in the Traveling Salesman Problem, chromosomes can represent the order in which cities are visited.
[0031] Initialize the chromosome population: Randomly generate a certain number of initial chromosomes to form a "population" as the starting point for the search for solutions.
[0032] Calculate the fitness of each chromosome: Use the fitness function to evaluate the quality of the chromosome (solution). The higher the fitness, the closer the solution is to the "optimal".
[0033] Calculate crossover and mutation probabilities: Based on the algorithm strategy or population state, determine the probability of chromosome "crossover" (gene exchange) and "mutation" (random gene change) to control the evolutionary process.
[0034] Termination condition: Determine whether to stop the algorithm. Conditions are usually "reaching the number of iterations" or "the fitness is good enough". If "yes", output the optimal solution; if "no", proceed to the next evolutionary step.
[0035] Selection: Based on fitness, superior chromosomes are selected from the population to enter the next generation. Common strategies include "roulette wheel" and "elite retention" to give high-quality solutions more opportunities to be inherited.
[0036] Comparing fitness: (Optional step; the logic in the diagram may be to select chromosomes to participate in evolution) Comparing chromosome fitness assists in selection, crossover, and other operations.
[0037] Crossover: Selected chromosomes exchange some genes according to probability to generate new chromosomes, simulating biological hybridization and expanding the search range of solutions.
[0038] Mutation: Allow chromosomes to randomly change genes with probability to avoid the algorithm getting stuck in local optima and maintain population diversity.
[0039] Calculate chromosome fitness after crossover mutation: assess the quality of the new chromosome after evolution and prepare for selection of the next generation.
[0040] Calculate crossover and mutation probabilities: (This may involve dynamically adjusting the probabilities or repeating the initial probability calculation logic) Update the probabilities of evolutionary operations and continue iterating.
[0041] Is the number of iterations less than the set number?: Determine if the number of iterations has reached the target. If "yes", return to the "selection" step to continue evolution; if "no", end the iteration and output the optimal solution.
[0042] Output the optimal solution: The algorithm ends and obtains the chromosome with the highest fitness in the current population as the approximate optimal solution to the problem.
[0043] In this embodiment of the disclosure, the preset number of weak classifiers is 10 to 20.
[0044] In this embodiment, the key parameters include the SVM penalty factor C (range 1-100), the kernel function parameter σ (range 0.1-10), and the number of iterations T of the weak classifier in the AdaBoost algorithm (iterative algorithm) (range 5-30). Each parameter is encoded using an 8-12 bit binary encoding method, and then the encoded sequences of each parameter are spliced together to form a complete chromosome.
[0045] In this embodiment of the disclosure, determining the improved genetic algorithm includes: randomly selecting two individuals from the population each time, calculating the recognition accuracy of their respective AdaBoost-SVM models on the validation set as fitness values, retaining individuals with high fitness values, and performing crossover and mutation operations on individuals with low fitness values; the crossover operation sets the crossover point at the boundary of each parameter gene segment (such as the junction of the penalty factor C encoding sequence and the kernel function parameter σ encoding sequence) to avoid destroying superior genes; the mutation operation adaptively adjusts the probability according to the individual fitness value, and the improved genetic algorithm is determined based on the above rules.
[0046] In this embodiment, the chromosome population is input into the improved genetic algorithm. The fitness value is calculated by measuring the recognition accuracy of the AdaBoost-SVM model corresponding to each chromosome on the validation set of the power line tower defect database. Based on the fitness value, selection (retaining the top 30% to 50% of individuals), crossover, and mutation operations are performed to iteratively update the chromosome population until a preset number of iterations (50 to 100 times) is reached or the fitness value fluctuation is less than 1% for 5 consecutive iterations. At this point, the model parameters (penalty factor C, kernel function parameter σ, iteration number T) determined by the chromosome with the highest fitness are used to construct the initial IGA-AdaBoost-SVM model. The formulas for calculating the adaptive crossover probability Pc and the adaptive mutation probability Pm are shown below:
[0047]
[0048] In the formula, Pc is the adaptive crossover probability, and Pm is the adaptive mutation probability f. max f is the fitness value of the chromosome with the highest fitness in the population. avg f is the average fitness of all chromosomes in the population. j f is the fitness value of the chromosome with lower fitness among the two chromosomes to be crossed. b For the fitness values of the individuals to be mutated, this patent uses K1=K2=0.8 and K3=K4=0.06.
[0049] In this embodiment of the disclosure, step S2 further includes: The power line tower defect database is divided into training set, validation set and test set based on a preset ratio; After training the initial IGA-AdaBoost-SVM model using the training set, the model is validated using the validation set, and the defect identification accuracy, false negative rate, and false positive rate of the model on the validation set are calculated. If the model's recognition accuracy is lower than the preset accuracy threshold, or the false negative rate is higher than the preset false negative rate threshold, or the false positive rate is higher than the preset false positive rate threshold, then analyze the cause of the error. If the parameter settings are unreasonable, return to the improved genetic algorithm to re-optimize the model parameters and adjust the algorithm parameters such as population size and number of iterations; If the database is insufficient or contains biases, supplement the data with relevant defective data, update the database, and then retrain the model. Repeat the above process until the model's performance metrics on the validation set meet the preset requirements, and obtain the target IGA-AdaBoost-SVM model.
[0050] In this embodiment of the disclosure, the preset ratio is training set: validation set: test set = 7:1:2.
[0051] In this embodiment of the disclosure, the defect identification accuracy of the model on the validation set = number of correctly identified defect samples / total number of defect samples in the validation set × 100%; False negative rate = (Number of unidentified real defect samples / Total number of defect samples in the validation set) × 100%; False positive rate = (Number of non-defect samples falsely identified as defects / Number of non-defect samples in the validation set) × 100%; In this embodiment of the disclosure, the preset accuracy threshold is 90% to 95%, the preset false negative threshold is 3% to 5%, and the preset false positive threshold is 3% to 5%.
[0052] In this embodiment of the disclosure, if the model is overfitted due to unreasonable parameter settings (such as the penalty factor C being too large, or the model being underfitted due to insufficient iterations T), the improved genetic algorithm is returned to re-optimize the model parameters, and the algorithm parameters such as population size (e.g., the original size of 50 is adjusted to 70) and iteration count (e.g., the original count of 50 is adjusted to 80) are adjusted.
[0053] In this embodiment of the disclosure, if the database data is insufficient or there are deviations (such as fewer than 50 samples of a certain type of defect or missing data related to some samples), relevant defect data is collected (such as adding inspection images and sensor vibration data of this type of defect) and the database is updated before the model is retrained.
[0054] S3: Based on the actual location data of power line towers combined with the UAV, the target shooting path is planned. Based on the target shooting path, multispectral imaging technology is used to collect real-time tower images and perform preprocessing to obtain actual tower images. In this embodiment of the disclosure, step S3 specifically includes: S31, Obtain the actual location data of power line towers; S32, combining the actual location data and the UAV's flight parameters, a path planning algorithm is used to plan the initial shooting path; S33 adjusts the initial shooting path in real time based on the real-time environmental conditions to obtain the target shooting path.
[0055] In this embodiment of the disclosure, the flight parameters of the UAV may include, for example, a maximum flight altitude of 100-150m, a flight time of 20-30min, and a shooting radius of 5-10m / tower.
[0056] In this embodiment of the disclosure, adjusting the initial shooting path in real time based on real-time environmental conditions may include, for example, increasing the flight altitude by 10-20m when the wind speed is greater than level 5, adjusting the shooting angle by 15-30° when vegetation obstructs the view, prioritizing shooting during periods of sufficient sunlight (9:00-11:00 AM, 2:00-4:00 PM, and in rain-free weather), and avoiding image reflection problems caused by direct sunlight at noon (12:00-1:00 PM) to obtain the target shooting path.
[0057] In this embodiment, the UAV's path planning is not static but is adjusted in real time by combining static geographic information and dynamic environmental factors. This ensures that image acquisition tasks can be completed in the most efficient and safest manner under various complex environments, guaranteeing timely and high-quality data acquisition.
[0058] In this embodiment of the disclosure, step S3 further includes: S34, based on the target shooting path, uses a multispectral imaging sensor to acquire multiple band image data of power line towers; S35 performs image registration and fusion processing on the acquired image data of multiple bands to generate a high-resolution multispectral image; S36 performs denoising, enhancement, and segmentation preprocessing on the multispectral image to extract the main structural region of the tower and obtain the actual tower image.
[0059] In this embodiment of the disclosure, the multiple bands include the visible light band (400-760nm, used to identify missing tower materials and bird nests), the near-infrared band (760-1100nm, used to identify wire wear), and the thermal infrared band (8-14μm, used to assist in judging poor contact and heat generation caused by loose bolts).
[0060] In this embodiment of the disclosure, image registration and fusion processing are performed on the acquired image data of multiple bands: a registration algorithm based on feature point matching is used to align images of different bands (registration error ≤ 1 pixel), and a high-resolution multispectral image is generated by a weighted fusion algorithm (weights are set according to visible light: near infrared: thermal infrared = 4:3:3).
[0061] In this embodiment of the disclosure, the multispectral image is preprocessed by denoising, enhancement and segmentation: a Gaussian filtering algorithm with a kernel size of 3×3 is used to remove image noise, a histogram equalization algorithm is used to enhance the contrast of defect areas, a semantic segmentation algorithm (such as the U-Net model) is used to extract the main structure area of the tower, and background areas such as sky and trees are removed to obtain the actual tower image.
[0062] In this embodiment, multispectral imaging technology is employed to capture defect features (such as thermal anomalies and early corrosion) that are imperceptible to the human eye, providing the model with richer criteria for discrimination. Professional registration, fusion, and preprocessing procedures effectively highlight the main body of the tower, eliminate background interference, and significantly improve the efficiency and accuracy of subsequent model recognition and processing.
[0063] S4. Based on the actual tower images and the target IGA-AdaBoost-SVM model, identify the actual defect type, actual defect location, and actual severity, and generate a real-time report containing trend analysis and maintenance suggestions.
[0064] In this embodiment of the disclosure, step S4 specifically includes: S41, Input the actual tower image into the target IGA-AdaBoost-SVM model, and the model outputs the probability distribution of defect types; S42, determine the actual defect type based on probability distribution, and use image positioning technology to identify the actual defect location on the tower; S43, assess the actual severity based on the characteristic parameters of the defect type and the confidence level of the model output.
[0065] In this embodiment of the disclosure, the model first extracts multi-dimensional features of the actual tower image, including texture features (using the LBP local binary mode algorithm to capture the texture changes of defects such as loose bolts and worn wires), shape features (extracting morphological parameters such as missing tower materials, area, perimeter, and aspect ratio of bird nests through edge detection and contour analysis), and spectral features (based on the band gray value differences of multispectral images, such as the reflectance difference between the worn area and the normal area of the wire in the near-infrared band, and the temperature anomaly caused by loose bolts and poor contact in the thermal infrared band).
[0066] In this embodiment of the disclosure, the extracted features are compared with the historical defect feature library learned during the model training phase, and the similarity between the current features and various defect features is calculated using an SVM classifier.
[0067] By leveraging the weight allocation mechanism of the AdaBoost ensemble algorithm and combining the classification results of multiple SVM weak classifiers, the matching similarity of various defects is weighted and fused (weak classifiers with higher classification accuracy have a higher weight ratio, such as a weight range of 0.1-0.3).
[0068] In this embodiment of the disclosure, the weighted fusion similarity is converted into probability values (range 0-1, retained to 4 decimal places) to form a probability distribution of defect types. For example, the output results may be "loose bolts: 0.8921, missing tower material: 0.0356, worn wires: 0.0512, vibration damper offset: 0.0187, bird nest: 0.0024", while the overall recognition confidence of the model is also output (a normalized evaluation value that combines the probabilities of various defects, such as 0.91).
[0069] In this embodiment, a probability threshold is set (preset to 0.8 in this embodiment), and the defect type with the highest probability value in the probability distribution that is greater than the threshold is selected as the actual defect type. If the probability values of all defect types are less than 0.8, or the difference between the highest probability value and the second highest probability value is less than 0.1 (e.g., loose bolt: 0.78, worn wire: 0.75), the identification result is marked as "awaiting manual review" to avoid misjudgment due to high feature similarity.
[0070] If multiple defects occur simultaneously (such as a single tower having both loose bolts and a bird's nest), then the types with a probability value greater than 0.8 in each defect category are selected and all are determined as actual defect types, resulting in a multi-defect labeling result.
[0071] In this embodiment of the disclosure, the image coordinates are marked by determining the range of pixel coordinates of the defect in the actual tower image through an image segmentation algorithm (such as the U-Net semantic segmentation model), for example, "the coordinates of the upper left corner of the bolt loosening defect area are (285, 412) and the coordinates of the lower right corner are (320, 448)", and then marking them on the image with a red rectangle.
[0072] Physical location transformation: Combining the parameters taken by the drone (stored in the image metadata), including the shooting focal length (e.g., 16mm), shooting angle (the tilt angle relative to the vertical direction of the tower, e.g., 15°), shooting height (e.g., 120m), and the actual height of the tower (obtained from the power GIS system, e.g., 30m), the pixel coordinates are converted into the actual physical location using the perspective projection transformation formula.
[0073] Standardized location descriptions: In accordance with the power industry's operation and maintenance standards, physical locations are converted into structured descriptions, such as "Tower No. #123, bolt group at 18.5m above ground on phase A crossarm" and "Tower No. #456, conductor mid-section at 8m from tower body," ensuring that operation and maintenance personnel can quickly locate the specific location of defects.
[0074] In this embodiment of the disclosure, key feature parameters corresponding to different defect types are extracted as evaluation criteria: Loose bolts: The rotation angle of the bolt is calculated using image measurement tools (based on the relative position change between the bolt head and the tower connection, with an accuracy of ±1°), combined with vibration frequency data collected by vibration sensors (the vibration frequency of loose bolts is usually 2 to 5 Hz higher than that of normal bolts). Missing tower materials: The actual missing area is calculated by converting the area of images into pixels (based on the scale of drone photography, such as 1 pixel corresponding to 0.001m). 2 , accuracy ±0.01m 2 Record whether the missing part is a critical structural component of the tower (such as the main pole or crossarm load-bearing part). Wire wear: The wear depth is calculated by the difference in gray values of near-infrared images (establishing a calibration model between gray values and actual wear depth, such as an increase of 0.1 in gray value difference corresponding to an increase of 0.1 mm in wear depth), combined with real-time tension data collected by wire tension sensors (wear leads to a reduction in wire cross-section, with abnormal tension fluctuations within ±5%). Vibration damper offset: Measure the straight-line distance between the actual position of the vibration damper and the designed installation position (based on the conversion relationship between image coordinates and physical position, with an accuracy of ±0.05m), combined with the line span parameters (obtained from the power line design file, such as 200m). Bird's nest: The volume of the bird's nest was calculated through image contour analysis (using an approximate formula for the volume of a sphere or cylinder, with an accuracy of ±0.02m). 3 Record the minimum distance between the bird's nest and the wire (to determine if there is a risk of discharge).
[0075] In this embodiment of the disclosure, a grading rule is formulated by combining feature parameters and model recognition confidence (confidence ≥ 0.9 is high confidence, 0.7-0.9 is medium confidence, and < 0.7 is low confidence), as shown below: When the defect type is loose bolts, the following conditions apply: mild (loosening angle < 15°, vibration frequency fluctuation < 2Hz, confidence level ≥ 0.8), moderate (15° ≤ loosening angle ≤ 30°, 2Hz ≤ vibration frequency fluctuation ≤ 5Hz, confidence level ≥ 0.7), and severe (loosening angle > 30°, vibration frequency fluctuation > 5Hz, confidence level ≥ 0.7). When the defect type is missing tower material, a minor defect is defined as a missing area < 0.1m². 2 For non-load-bearing parts, the confidence level is ≥0.8, and the moderate confidence level is 0.1m. 2 ≤Missing area≤0.5m 2 For non-critical load-bearing areas, with a confidence level ≥ 0.7, the height is defined as the missing area > 0.5m. 2 Key stress-bearing parts, confidence level ≥ 0.7; When the defect type is wire wear, mild is defined as wear depth < 10% of wire diameter, tension fluctuation < 2%, and confidence level ≥ 0.8; moderate is defined as 10% ≤ wear depth ≤ 30%, 2% ≤ tension fluctuation ≤ 5%, and confidence level ≥ 0.7; and severe is defined as wear depth > 30%, tension fluctuation > 5%, and confidence level ≥ 0.7. When the defect type is vibration damper misalignment, the following criteria apply: Mild: Misalignment distance < 0.5m, span ratio < 0.25%, confidence level ≥ 0.8; Moderate: Misalignment distance ≤ 1m, span ratio ≤ 0.25%, confidence level ≥ 0.7; High: Misalignment distance > 1m, span ratio > 0.5%, confidence level ≥ 0.7. When the defect type is bird's nest, a minor defect is defined as a volume <0.2m³. 3 The distance from the conductor is >1m, with a confidence level ≥0.8, and a moderate confidence level of 0.2m. 3 ≤Volume≤0.5m³ 3 0.5m ≤ distance from conductor ≤ 1m, confidence level ≥ 0.7, height is volume > 0.5m 3 The distance from the conductor is <0.5m, and the confidence level is ≥0.7. If the model identifies a low confidence level (<0.7), the severity level is reduced by one level based on the feature parameter classification (e.g., if the feature parameter is judged as moderate, the final assessment is mild) to ensure the reliability of the assessment results.
[0076] In this embodiment of the disclosure, step S4 further includes: S44, based on the historical defect database and current identification results, performs time series analysis to predict defect development trends; S45 generates targeted operation and maintenance suggestions based on the defect type, location, and severity, combined with the operation and maintenance rule base; S46 integrates identification results, trend analysis, and operation and maintenance suggestions to generate structured reports, which are then transmitted to the operation and maintenance platform in real time via the network.
[0077] In this embodiment of the disclosure, based on the constructed time series model, the defect development status is predicted for the next 3-6 months, and a "severity change curve" and a "risk level warning" are output: If the prediction results show that the rate of increase in defect severity is <0.5 levels / month (e.g., it takes more than 2 months to go from mild to moderate), and it is still in the mild or moderate range after 6 months, it is marked as "low-risk trend"; If the growth rate is ≥0.5 levels / month, or if it will reach the severe standard in 6 months (e.g., the wear depth of the wire will exceed 30% of the diameter), it will be marked as a "high-risk trend" with key time points noted (e.g., "It is expected to enter a severe state in 3 months, and early intervention is required"). In this embodiment of the disclosure, real-time updating of the power line tower defect database specifically includes: Store actual tower images, identified defect types, locations, severity, and maintenance reports in the database; Regularly obtain feedback data after defect handling from the operation and maintenance platform, and update the tags and associated data of defect samples; The target IGA-AdaBoost-SVM model was retrained based on the new data to maintain the model's recognition accuracy.
[0078] Based on the above method, this disclosure also provides a power line tower defect identification system based on the IGA-AdaBoost-SVM model corresponding to the above method, the system comprising: The database construction module is used to collect historical defect images of several power line towers and their corresponding historical related data, and integrate them into a power line tower defect database. The model building module is used to build an initial IGA-AdaBoost-SVM model based on the power line tower defect database and verify it, adjust the initial IGA-AdaBoost-SVM model based on the verification results, obtain the target IGA-AdaBoost-SVM model, and update the power line tower defect database in real time. The data acquisition module is used to plan the target shooting path based on the actual location data of the power line towers combined with the drone. Based on the target shooting path, multispectral imaging technology is used to acquire real-time tower images and perform preprocessing to obtain actual tower images. The report generation module is used to identify the actual defect type, actual defect location and actual severity based on the actual tower image and the target IGA-AdaBoost-SVM model, and generate a real-time report containing trend analysis and operation and maintenance suggestions.
[0079] Based on the same inventive concept as the above disclosure, this disclosure also provides an electronic device. The electronic device of this disclosure includes at least one processor and at least one memory electrically connected to the processor. The memory is electrically connected to the processor, wherein the memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the method described above.
[0080] It should be noted that the electrical connection between the above-mentioned units does not necessarily mean the connection between lines. The indirect connection method can be applied to the embodiments of this disclosure as long as it achieves the purpose of this disclosure.
[0081] Based on the same inventive concept, this disclosure also provides a computer storage medium storing a computer program, which, when executed by a processor, implements the steps of the above method.
[0082] Although the present disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure.
Claims
1. A method for identifying defects in power line towers based on the IGA-AdaBoost-SVM model, characterized in that, The method includes, Collect historical defect images of several power line towers and their corresponding historical data, and integrate them into a power line tower defect database. Collect historical defect images of several power line towers and their corresponding historical correlation data, specifically including: Determine the types of historical defects corresponding to power line towers; Based on the historical inspection database, determine the inspection defect images and their corresponding inspection association data for each historical defect type. Based on sensor acquisition, defect images and their corresponding associated data for each historical defect type are collected. By comparing the inspection defect image and the collected defect image, images with a similarity greater than a similarity threshold are integrated. The inspection defect images and their corresponding inspection-related data corresponding to each historical defect type are statistically analyzed, as well as the collected defect images and their corresponding collected-related data, to form the power line tower defect database. An initial IGA-AdaBoost-SVM model is constructed and validated based on the power line tower defect database. The initial IGA-AdaBoost-SVM model is adjusted based on the validation results to obtain the target IGA-AdaBoost-SVM model, and the power line tower defect database is updated in real time. The target shooting path is planned based on the actual location data of power line towers using drones. Based on the target shooting path, multispectral imaging technology is used to collect real-time tower images and perform preprocessing to obtain actual tower images. Based on the actual tower images and the target IGA-AdaBoost-SVM model, the actual defect type, actual defect location, and actual severity are identified, and a real-time report containing trend analysis and maintenance suggestions is generated.
2. The method according to claim 1, characterized in that, An initial IGA-AdaBoost-SVM model is constructed based on the aforementioned power line tower defect database, specifically including: The AdaBoost algorithm is used to integrate multiple SVMs. Initially, each defect sample in the power line tower defect database is assigned equal weight. After training the first SVM weak classifier, the sample weights are adjusted according to its classification error. Then, the next SVM weak classifier is trained. This process is repeated until a preset number of weak classifiers are generated. Finally, corresponding weights are assigned according to the classification accuracy of each weak classifier to construct the AdaBoost-SVM model. The key parameters affecting the accuracy of power line tower defect identification in the AdaBoost-SVM model are used as gene segments of a chromosome, and the gene segments are encoded into a chromosome using binary encoding. Based on the actual needs of power line tower defect identification, the population size is set, and the binary code of each individual is randomly generated within a reasonable range of values for each parameter, thus obtaining the chromosome population. The population size is iteratively compared based on the fitness of any two individuals, and crossover and mutation operations are performed based on the comparison results to determine the improved genetic algorithm. The chromosome population is input into the improved genetic algorithm. The fitness value is calculated by measuring the recognition accuracy of the AdaBoost-SVM model corresponding to each chromosome on the validation set of the power line tower defect database. Selection, crossover, and mutation operations are performed based on the fitness value to iteratively update the chromosome population until a preset number of iterations is reached or the fitness value tends to stabilize. At this point, the model parameters determined by the chromosome with the highest fitness are used to construct the initial IGA-AdaBoost-SVM model.
3. The method according to claim 2, characterized in that, The initial IGA-AdaBoost-SVM model was adjusted based on the validation results, specifically including: The power line tower defect database is divided into training set, validation set and test set based on a preset ratio; After training the initial IGA-AdaBoost-SVM model using the training set, the model is validated using the validation set, and the defect identification accuracy, false negative rate, and false positive rate of the model on the validation set are calculated. If the model's recognition accuracy is lower than the preset accuracy threshold, or the false negative rate is higher than the preset false negative rate threshold, or the false positive rate is higher than the preset false positive rate threshold, then analyze the cause of the error. If the parameter settings are unreasonable, return to the improved genetic algorithm to re-optimize the model parameters and adjust the algorithm parameters such as population size and number of iterations; If the database is insufficient or contains biases, supplement the data with relevant defective data, update the database, and then retrain the model. Repeat the above process until the model's performance metrics on the validation set meet the preset requirements, and obtain the target IGA-AdaBoost-SVM model.
4. The method according to claim 3, characterized in that, The target shooting path is planned based on the actual location data of power line towers using drones, specifically including: Obtain the actual location data of power line towers; Based on the actual location data and the UAV's flight parameters, a path planning algorithm is used to plan the initial shooting path; The initial shooting path is adjusted in real time based on the real-time environmental conditions to obtain the target shooting path.
5. The method according to claim 4, characterized in that, Based on the target shooting path, multispectral imaging technology is used to acquire real-time tower images and perform preprocessing, specifically including: Based on the target shooting path, multispectral imaging sensors are used to acquire multi-band image data of power line towers; Image registration and fusion processing are performed on the acquired image data of multiple bands to generate high-resolution multispectral images; The multispectral image is preprocessed by denoising, enhancement and segmentation to extract the main structural region of the tower and obtain the actual tower image.
6. The method according to claim 5, characterized in that, Based on the actual tower images and the target IGA-AdaBoost-SVM model, the actual defect type, actual defect location, and actual severity are identified, specifically including: The actual tower image is input into the target IGA-AdaBoost-SVM model, and the model outputs the probability distribution of defect types. The actual defect type is determined based on the probability distribution, and the actual defect location on the tower is identified using image localization technology. The actual severity is assessed based on the characteristic parameters of the defect type and the confidence level of the model output.
7. The method according to claim 6, characterized in that, Generate real-time reports containing trend analysis and operational recommendations, specifically including: Based on the historical defect database and current identification results, time series analysis is conducted to predict defect development trends; Based on the defect type, location, and severity, and combined with the operation and maintenance rule base, targeted operation and maintenance suggestions are generated; The system integrates identification results, trend analysis, and operation and maintenance recommendations to generate a structured report, which is then transmitted to the operation and maintenance platform in real time via the network.
8. The method according to claim 7, characterized in that, The power line tower defect database is updated in real time, specifically including: Store actual tower images, identified defect types, locations, severity, and maintenance reports in the database; Regularly obtain feedback data after defect handling from the operation and maintenance platform, and update the tags and associated data of defect samples; The target IGA-AdaBoost-SVM model was retrained based on the new data to maintain the model's recognition accuracy.
9. A power line tower defect identification system based on the IGA-AdaBoost-SVM model, characterized in that, The system includes: The database construction module is used to collect historical defect images of several power line towers and their corresponding historical related data, and integrate them into a power line tower defect database. The model building module is used to build an initial IGA-AdaBoost-SVM model based on the power line tower defect database and verify it, adjust the initial IGA-AdaBoost-SVM model based on the verification results, obtain the target IGA-AdaBoost-SVM model, and update the power line tower defect database in real time. Collect historical defect images of several power line towers and their corresponding historical correlation data, specifically including: Determine the types of historical defects corresponding to power line towers; Based on the historical inspection database, determine the inspection defect images and their corresponding inspection association data for each historical defect type. Based on sensor acquisition, defect images and their corresponding associated data for each historical defect type are collected. By comparing the inspection defect image and the collected defect image, images with a similarity greater than a similarity threshold are integrated. The inspection defect images and their corresponding inspection-related data corresponding to each historical defect type are statistically analyzed, as well as the collected defect images and their corresponding collected-related data, to form the power line tower defect database. The data acquisition module is used to plan the target shooting path based on the actual location data of the power line towers combined with the drone. Based on the target shooting path, multispectral imaging technology is used to acquire real-time tower images and perform preprocessing to obtain actual tower images. The report generation module is used to identify the actual defect type, actual defect location and actual severity based on the actual tower image and the target IGA-AdaBoost-SVM model, and generate a real-time report containing trend analysis and operation and maintenance suggestions.