Unmanned aerial vehicle bridge inspection method based on radar scanning

Abstract

The invention discloses an unmanned aerial vehicle bridge inspection method based on radar scanning, and the method comprises the steps: obtaining the original data of a bridge through radar scanning, and carrying out the fusion of the original data, and obtaining a first data set; preprocessing the first data set to obtain a second data set; identifying the second data set through a deep learning model to obtain a defect distribution diagram; obtaining a priority processing list according to the defect distribution map, optimizing the inspection path of the unmanned aerial vehicle according to the priority processing list to obtain a path planning result, adjusting the inspection task of the unmanned aerial vehicle according to the path planning result, and obtaining a third data set through radar scanning in real time; and classifying and positioning the third data set through a deep learning model to obtain a final inspection result.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of unmanned aerial vehicle inspection, and particularly relates to a method for unmanned aerial vehicle bridge inspection based on radar scanning. BACKGROUND

[0002] Bridge inspection is an important field for ensuring the safety of transportation infrastructure, and its key lies in timely detection of structural defects to prevent major safety accidents. With the acceleration of urbanization, the number of bridges has increased rapidly. Traditional manual inspection is low in efficiency and high in risk, and it is difficult to meet the demand for large-scale and high-frequency inspection. In recent years, unmanned aerial vehicle technology has been introduced into bridge inspection, combined with optical or laser equipment for data collection. Although some studies have attempted to use multispectral imaging or infrared thermal imaging technology for bridge detection, they are still susceptible to factors such as light and fog under complex weather conditions, resulting in blurred images or misidentification. In addition, visual-based path planning algorithms rely on visibility and are difficult to achieve stable inspection in all-weather and all-time periods. These methods are easily disturbed by light, weather, and other factors in complex environments, and the data reliability is insufficient, making it difficult to fully cover the key parts of the bridge, and the accuracy and stability of the inspection results have obvious limitations. Another method uses single-frequency radar scanning, which has certain penetration ability to overcome the above problems,

[0003] In the research of unmanned aerial vehicle bridge inspection based on radar scanning, although it has certain penetration ability, it is limited by single frequency resolution and weak anti-interference ability. Therefore, the core challenge lies in how to ensure the high reliability and intelligence of the radar scanning system in complex environments. Radar scanning needs to overcome the limitations of single frequency resolution and cope with environmental noise and electromagnetic interference, which requires data processing algorithms to have strong anti-interference ability, otherwise it may lead to a decrease in detection accuracy. Due to the complexity of anti-interference algorithms, the system's computing load increases, which in turn puts higher requirements on the control architecture of the unmanned aerial vehicle, which needs to maintain stable operation under high load. If the control architecture cannot adapt, it may cause system failure and affect the continuity of the inspection task. In addition, existing path planning methods mostly use fixed paths or simple obstacle avoidance strategies, and lack dynamic response mechanisms for defect priority and real-time environmental changes, which cannot effectively improve the coverage rate of key parts. The inspection path planning needs to be adjusted dynamically according to the bridge structure and real-time environment to ensure the coverage rate of key parts, but this depends on efficient intelligent algorithms and real-time data fusion, and the lack of dynamic adjustment capability will lead to low inspection efficiency or the risk of missing detection.

[0004] Therefore, how to design an unmanned aerial vehicle bridge inspection system based on radar scanning, through anti-interference data processing, distributed control architecture, and intelligent path planning, to achieve high reliability and efficient coverage, has become a key problem. SUMMARY

[0005] To solve the above technical problems, the present application provides a kind of unmanned aerial vehicle bridge inspection method based on radar scanning to solve the problems existing in the prior art.

[0006] To achieve the above object, the present application provides a kind of unmanned aerial vehicle bridge inspection method based on radar scanning, comprising:

[0007] The original data of the bridge are acquired by radar scanning, the original data are fused to obtain a first data set, the first data set is preprocessed to obtain a second data set, the second data set is identified by a deep learning model to obtain a defect distribution map, a priority processing list is obtained according to the defect distribution map, the unmanned aerial vehicle inspection path is optimized according to the priority processing list to obtain a path planning result, the inspection task of the unmanned aerial vehicle is adjusted according to the path planning result, and a third data set is acquired in real time by radar scanning, the third data set is classified and located by a deep learning model to obtain a final inspection result.

[0008] Optionally, the acquisition process of the first data set comprises:

[0009] The three-dimensional point cloud data of the bridge are acquired by a multi-band radar carried by the unmanned aerial vehicle, the three-dimensional point cloud data of the multi-band are fused, and the spatial domain features of the fused data are extracted to obtain the first data set.

[0010] Optionally, the acquisition process of the second data set comprises:

[0011] The first data set is judged, and when the first data set has noise or interference, the first data set is preprocessed by an anti-interference method and an adaptive filter to obtain the second data set.

[0012] Optionally, the acquisition process of the defect distribution map comprises:

[0013] The second data set is identified by a deep learning model to obtain the category and position information of the bridge surface defects, the spatial coordinates of the bridge label defects are extracted according to the category and position information of the bridge surface defects, the defect distribution map is visualized according to the spatial coordinates, when the accuracy of the defect distribution map is lower than a threshold value, the second data set is enhanced and expanded, and the deep learning model is iteratively optimized, the defect distribution map is reacquired and updated according to the optimized deep learning model.

[0014] Optionally, the acquisition process of the priority processing list comprises:

[0015] The defect distribution map is calculated by a distributed control architecture, and the calculation of the distributed tasks is performed by a multi-node calculation method to obtain the defect density and repair priority of the key parts of the bridge, and the priority processing list is obtained.

[0016] Optionally, the path planning result acquisition process comprises:

[0017] According to the spatial coordinates of the defect density of the bridge key part in the priority processing list, the spatial coordinates are taken as reference points, the unmanned aerial vehicle inspection path is optimized through an optimization algorithm, and the optimized unmanned aerial vehicle inspection path is adjusted through real-time environmental data to obtain a path planning result, wherein when the key part coverage rate of the path planning result is lower than a threshold value, the parameters of the optimization algorithm are adjusted through a reinforcement learning method to re-generate the path planning result.

[0018] Optionally, the final inspection result acquisition process comprises:

[0019] The defect distribution map is updated through the radar scanning result, and new defect data is extracted from the updated defect distribution map to obtain a third data set, the new defect data in the third data set is classified and positioned through a convolutional neural network, and according to the defect and positioning classification result, the defect distribution map and the priority list are updated to obtain a final inspection result.

[0020] Optionally, the path planning result re-generation process comprises:

[0021] The key part coverage rate of the path planning result is judged, when it is lower than a threshold value, the parameters of the optimization algorithm are adjusted through reinforcement learning to obtain an optimization parameter set, the optimization parameter set is used to drive the optimization algorithm to generate a new path set to obtain a candidate path planning, the path coverage rate of the key part of the candidate path planning is calculated to obtain coverage rate data, and the candidate path planning is re-iterated according to the coverage rate data until the coverage rate comparison of the candidate path planning and the historical path planning exceeds the threshold value to obtain a path planning result.

[0022] Compared with the prior art, the present application has the following advantages and technical effects:

[0023] The application discloses a bridge structure defect detection and unmanned aerial vehicle inspection system based on radar scanning and artificial intelligence. The system obtains bridge three-dimensional point cloud data through multi-frequency radar fusion, processes and analyzes the data by using an anti-interference algorithm and a convolutional neural network, and realizes accurate positioning of surface defects such as cracks and corrosion. The application also optimizes the unmanned aerial vehicle inspection path by using a distributed computing architecture and a genetic algorithm, and dynamically adjusts the flight trajectory through reinforcement learning to ensure comprehensive coverage of key parts. During the inspection process, the system can update the defect distribution map in real time and identify new defects, thereby providing continuously optimized detection results. This intelligent bridge detection method significantly improves the accuracy and efficiency of defect identification, and provides reliable technical support for bridge maintenance and safety management.

[0024] The core innovation of the present application is to propose an intelligent bridge inspection system that integrates multi-band radar data acquisition and deep learning defect identification. In the present application, high-precision defect detection in complex environments is achieved based on multi-band radar fusion and anti-interference processing; the problem of high computational load is solved by using a distributed control architecture; and the inspection path is optimized in real time according to the defect priority and environmental changes by combining a dynamic path planning algorithm based on reinforcement learning, thereby significantly improving the coverage rate and efficiency of key parts. The three work together to form a closed-loop intelligent inspection system, solving the problems of environmental interference, computational bottleneck and dynamic response. In the above content, by introducing a high-frequency and low-frequency radar cooperative scanning mechanism, three-dimensional point cloud modeling of the bridge structure and synchronous detection of internal and external defects are realized, making up for the limitations of single-band radar. The present application first applies distributed computing architecture and multi-node parallel processing to bridge defect density analysis, solving the computational bottleneck problem in large-scale radar data processing and significantly improving the efficiency of defect identification and priority sorting. In terms of path planning, the present application innovatively combines genetic algorithm and reinforcement learning optimization mechanism to realize adaptive path generation based on defect distribution map and support dynamic adjustment of flight trajectory according to environmental changes, thereby ensuring efficient coverage of high-priority areas. In addition, the present application constructs a closed-loop feedback mechanism, updates the defect distribution map through real-time radar scanning, classifies and locates the newly added defects using a convolutional neural network, and forms continuously optimized inspection result output, thereby improving the intelligent level of bridge health monitoring. BRIEF DESCRIPTION OF DRAWINGS

[0025] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application and are incorporated in and constitute a part of this application. The embodiments of this application, and of which the principles are explained, are illustrated in the drawings, wherein:

[0026] Figure 1 A flowchart of the unmanned aerial vehicle bridge inspection method based on radar scanning according to an embodiment of the present application;

[0027] Figure 2 A schematic diagram of the unmanned aerial vehicle bridge inspection system based on radar scanning according to an embodiment of the present application. DETAILED DESCRIPTION

[0028] It should be noted that the embodiments and features in the present application can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0029] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in an order different from that shown herein.

[0030] AsFigure 1 As shown, the embodiment of the radar scanning-based unmanned aerial vehicle bridge inspection method can specifically include:

[0031] S101, obtain three-dimensional point cloud data and surface defect signals of the bridge structure through radar scanning technology, and adopt a multi-frequency radar fusion method to extract feature values from signals of different frequency bands to obtain a first data set.

[0032] The three-dimensional point cloud data and the surface defect signals are obtained by scanning the bridge structure through the multi-frequency radar, and the original data set is generated. The signal fusion method is adopted to perform weighted processing on the different frequency band signals in the original data set, extract the feature values, and generate the first data set.

[0033] For example, the three-dimensional point cloud data and the surface defect signals are obtained by scanning the bridge structure through the multi-frequency radar, and the original data set is generated, which involves the combination of radar scanning technology and data acquisition. A concrete bridge is scanned by combining a high-frequency radar (such as 24 GHz) and a low-frequency radar (such as 1 GHz). The high-frequency radar is good at capturing fine defects such as surface cracks, and the low-frequency radar can penetrate the internal structure to detect deep cavities or steel corrosion. During scanning, the radar equipment moves along the longitudinal direction of the bridge at a fixed speed to generate point cloud data containing spatial coordinates (x, y, z) and signal intensity information. A 100-meter-long bridge may generate about 5 million point cloud data points, while recording the surface crack width (such as 0.5 mm) or internal cavity position signal.

[0034] It should be noted that the point cloud data reflects the geometric shape of the bridge, and the defect signal indicates the material anomaly. The benefit of this step is to provide comprehensive structural information and lay the foundation for subsequent analysis.

[0035] In one possible implementation, the original data set is subjected to signal fusion, and a weighted processing method is adopted to integrate different frequency band signals. Specifically, the high-frequency signal is assigned a higher weight (such as 0.7) due to its high resolution, and the low-frequency signal is assigned a lower weight (such as 0.3) due to its deep detection capability. For example, for the surface crack signal of a certain bridge, the high-frequency radar detects a crack with a width of 0.3 mm, and the low-frequency radar confirms that the depth is 5 cm. Through weighted fusion, a comprehensive signal intensity value is generated to reflect the severity of the crack. Preferably, the fusion process can be completed through time domain or frequency domain analysis, such as aligning the signals in time sequence and removing noise. The corresponding spatial domain features are extracted, and the corresponding point cloud depth and pixel value are obtained as feature values according to their spatial domain positions. The benefit of this method is to improve data accuracy and reduce the risk of misjudgment of a single frequency band.

[0036] For this step, the single frequency band limitation can be broken through, the high-frequency radar can accurately capture the surface micro-cracks, and the low-frequency radar can penetrate the structure to detect internal cavities; the signal-to-noise ratio of different radar signals is improved through weighted fusion, and the different radar signals are effectively fused, so that accurate data basis is obtained through the detection mode of multiple radars and multiple frequency bands.

[0037] In S102, if the first data set is affected by environmental noise or electromagnetic interference, an anti-interference algorithm is used to process the first data set, noise is suppressed and interference signals are separated through an adaptive filter to obtain a second data set.

[0038] The first data set is extracted and identified for environmental noise or electromagnetic interference, and whether the first data set is affected by environmental noise and electromagnetic interference is determined according to the identification result to obtain a judgment result. If the first data set is affected by noise and interference, the first data set is processed through an anti-interference algorithm, and the first data set is preliminarily processed through an adaptive filter to suppress environmental noise and separate electromagnetic interference signals to obtain a second data set.

[0039] Exemplarily, the extraction and identification of environmental noise and electromagnetic interference for the first data set can be realized through a spectrum analysis method. Specifically, environmental noise usually appears as low-frequency random fluctuations, and electromagnetic interference may appear as high-frequency peak signals. In one possible implementation, the time domain signal of the first data set is converted into a frequency domain using fast Fourier transform, and whether there is an abnormal peak value in the frequency spectrum is observed. For example, in a bridge radar scanning scene, assuming that the first data set contains signals from multiple frequency band radars, spectrum analysis may show 0.5Hz low-frequency noise, which is derived from nearby wind vibration, and 2.4GHz high-frequency interference, which is derived from wireless communication equipment. By setting a spectrum threshold, such as an amplitude exceeding 3 times the average value of normal signals, it is determined whether the data set is affected by interference. It should be noted that this method is convenient for quickly locating the source of interference and provides a basis for subsequent processing.

[0040] In one embodiment, according to the judgment result, if the first data set is affected by noise and electromagnetic interference, an anti-interference algorithm can be used for processing. For example, the anti-interference algorithm based on wavelet transform can decompose the signal into different scales, retain the effective feature signals of the bridge structure, and filter out the noise. Wavelet transform separates low-frequency noise and high-frequency interference through multi-resolution analysis. For example, when processing a data set containing bridge surface defect signals, a Daubechies wavelet basis function is selected, and the signal is decomposed to 3 layers, and the middle frequency band signal is retained to protect the defect features. This method is simple to operate and can effectively protect the details of the original signal.

[0041] Preferably, the first data set is preliminarily processed by combining an adaptive filter to suppress environmental noise and separate electromagnetic interference signals. The adaptive filter tracks the interference components in the signal by adjusting the filter parameters in real time. For example, in a bridge radar scan, the adaptive filter can dynamically adjust the weights based on the least mean square error algorithm to suppress the 0.5 Hz environmental noise. One possible implementation is to set the initial step size of the filter to 0.01, and after 10 iterations, the noise amplitude can be reduced to 20% of the original. For electromagnetic interference, the filter can lock the 2.4 GHz spike signal and separate it by notch filtering. It should be noted that the advantage of the adaptive filter is that it does not need to know the exact frequency of the interference in advance, and it is suitable for real-time processing in complex environments. After the above processing generates the second data set, the data quality is significantly improved. For example, in bridge detection, the second data set can more clearly reflect the depth and position information of surface cracks, facilitating subsequent analysis.

[0042] To address the limitations of environmental noise and electromagnetic interference, the above-mentioned combination of various anti-interference methods forms a complete processing chain from spectral analysis to anti-interference algorithm to adaptive filtering, with layer-by-layer progression to ensure the purity of the signal. Wavelet transform separates signal levels, and adaptive filter dynamically suppresses environmental noise and electromagnetic interference. For example, spectral analysis provides the basis for interference identification, wavelet transform removes noise specifically, and adaptive filter further optimizes the signal. This multi-faceted collaborative solution not only has a logical structure, but also adapts to different interference scenarios, ensuring the reliability of the bridge structure data, and further purifying and optimizing the above-mentioned comprehensive data in multiple frequency bands, providing further effective data basis for subsequent content.

[0043] S103, according to the characteristic value distribution of the second data set, using a convolutional neural network algorithm to classify and locate the bridge surface defects, and extracting the spatial coordinates of the cracks, corrosion and other defects to obtain a first defect distribution map.

[0044] The second data set is standardized by using a preprocessing algorithm to obtain a feature distribution matrix. If the resolution of the feature distribution matrix is lower than a preset threshold, the resolution is enhanced by using an interpolation algorithm to obtain a standardized feature matrix. The standardized feature matrix is subjected to a convolution operation by using a convolutional neural network to extract deep features of the bridge surface defects to obtain a first feature map. According to the first feature map, a classification layer of the convolutional neural network is used to judge the defect type. If the classification probability is higher than a preset threshold, the crack or corrosion defect type is determined to obtain a defect classification result. The spatial coordinates of the defect are extracted from the first feature map by using a positioning layer of the convolutional neural network. If the confidence of the spatial coordinates is higher than a preset threshold, the crack coordinates and the corrosion coordinates are determined to obtain a defect positioning result. According to the defect classification result and the defect positioning result, the crack coordinates and the corrosion coordinates are fused to generate a first defect coordinate set. The first defect coordinate set is subjected to mapping processing by using a visualization algorithm to generate a first defect distribution map. The first defect distribution map is subjected to optimization processing by using an image enhancement algorithm. If the contrast of the distribution map is lower than a preset threshold, the brightness and the contrast are adjusted to obtain an optimized first defect distribution map.

[0045] Exemplarily, in the field of bridge surface defect detection, the standardization processing of the second data set by using the preprocessing algorithm is crucial. The standardization aims to unify the data format, eliminate the dimensional differences, and ensure the accuracy of subsequent analysis. For example, in a possible implementation manner, the pixel values of the second data set can be converted into a distribution with a mean value of 0 and a standard deviation of 1 by using a Z-score standardization method. Specifically, assuming that the second data set includes a gray point cloud image of a bridge surface, the pixel value range is 50 to 200, and after the standardization processing, the data distribution is more suitable for neural network input. It should be noted that the feature distribution matrix after the standardization reflects the texture and gray change of the image, and lays a foundation for subsequent feature extraction.

[0046] In an embodiment, if the resolution of the feature distribution matrix is lower than a preset threshold, for example, lower than 512x512 pixels, a bilinear interpolation algorithm can be used to enhance the resolution. For example, the original 256x256 matrix is interpolated to 512x512 to generate the standardized feature matrix. This method estimates the weighted average value of adjacent pixels, smooths the transition between pixels, and preserves the image details. Preferably, the interpolated matrix can more clearly present the fine cracks or corrosion traces on the bridge surface.

[0047] Specifically, convolutional neural networks perform convolution operations on the standardized feature matrix to extract deep features. For example, using a 3×3 convolution kernel with a stride of 1, a first feature map is extracted to capture the edge features of cracks or the texture changes of eroded areas. It can be understood that convolutional operations, through multiple layers of convolution and pooling, can progressively abstract from low-level edge features to high-level semantic features. For example, cracks might appear as thin, high-contrast lines, while eroded areas present as irregular patchy features. For instance, the classification layer determines the defect type based on the first feature map. Assuming a preset classification probability threshold of 0.8, if the predicted probability of a crack is 0.9, it is determined to be a crack defect. It should be noted that the classification layer typically uses the softmax function to transform the feature map into a probability distribution. For example, the network might output a crack probability of 0.9, an erosion probability of 0.05, and a normal surface probability of 0.05, thus clearly defining the defect type.

[0048] In one possible implementation, the localization layer extracts the spatial coordinates of the defects from the first feature map. For example, the bounding box of the crack is predicted by a regression branch, outputting the center point coordinates such as (200, 300) and a confidence score of 0.85. If the confidence score is higher than the threshold of 0.8, the coordinates are confirmed as the location of the crack. Preferably, this localization method can be accurate to the pixel level, facilitating subsequent maintenance and localization. Exemplarily, the crack and corrosion coordinates are fused to generate a first set of defect coordinates. For example, the crack coordinates (200, 300) and the corrosion coordinates (250, 320) are merged into a coordinate set that reflects the spatial distribution of the defects. Specifically, the coordinate set is grouped using a clustering algorithm to ensure that the coordinates of defects of the same type are clearly classified.

[0049] In one embodiment, the visualization algorithm maps the coordinate set to a first defect distribution map. For example, a heatmap is used to mark defect locations on a bridge surface image, with cracks displayed in red and corrosion in yellow. It should be noted that if the distribution map contrast is below a preset threshold, such as a contrast value less than 0.5, brightness and contrast can be adjusted through histogram equalization. For example, increasing the pixel values ​​in dark areas by 20% enhances the visibility of defect areas, generating an optimized distribution map that allows engineers to intuitively assess the bridge's condition.

[0050] Based on the aforementioned effective data, the deep learning model achieves pixel-level localization and supports simultaneous identification of multiple defects such as cracks and corrosion. By combining the aforementioned multi-frequency band data fusion and optimization, the effectiveness and generalization of the model for defect identification are further improved.

[0051] S104. If the detection accuracy of the first defect distribution map is lower than the preset threshold, the second dataset is expanded by data augmentation method and re-input into the convolutional neural network algorithm for iterative optimization to obtain the second defect distribution map.

[0052] If the detection accuracy of the first defect distribution map is lower than the preset threshold, the second data set is data augmented by an image transformation method to obtain an augmented data set. The augmented data set is input into a convolutional neural network for training, and a stochastic gradient descent method is used for iterative optimization to obtain an optimized model. The test data is predicted according to the optimized model to generate a second defect distribution map.

[0053] Exemplarily, if the detection accuracy of the first defect distribution map is lower than the preset threshold, the second data set needs to be data augmented to improve the model performance. Data augmentation is realized by an image transformation method, and common methods include geometric transformation and color transformation. In a possible implementation, geometric transformation can adopt random rotation, translation or flipping. For example, the bridge surface point cloud image is randomly rotated by ±15 degrees, or the image is horizontally flipped to simulate the defect morphology under different shooting angles. Color transformation can adjust the image brightness or contrast, such as randomly adjusting the brightness by ±20% to enhance the robustness of the model to light changes. These transformations generate diversified samples to expand the data set scale.

[0054] For example, the original data set contains 1000 bridge point cloud images, and after rotation, flipping and brightness adjustment, 3000 augmented images can be generated, thereby enriching the training data and reducing the risk of model overfitting.

[0055] Specifically, when the augmented data set is input into the convolutional neural network for training, a reasonable network structure and training strategy need to be designed. The convolutional neural network usually contains multiple convolutional layers and pooling layers for extracting spatial features of defects.

[0056] In an embodiment, a three-layer convolutional structure can be used, with 32 3x3 convolutional kernels in each layer, cooperating with a max pooling layer to compress the feature map size.

[0057] For example, the input image resolution is 256x256, and after convolution and pooling, the feature map size is reduced to 32x32, retaining the key defect information. During training, the batch size can be set to 64, and the training rounds are 50 rounds to balance the calculation efficiency and model convergence. This structure can effectively capture the texture features of cracks or corrosion, providing a basis for subsequent classification and positioning.

[0058] It should be noted that when using the random gradient descent method for iterative optimization, the selection of the learning rate is crucial. In an embodiment, the initial learning rate can be set to 0.001, and a learning rate decay strategy can be combined, such as reducing the learning rate by 10% every 10 rounds, to ensure that the model converges stably in the later training period. For example, in the early stage of training, the model quickly learns the rough features of the defects; while in the later stage, a smaller learning rate helps the model refine the parameters and optimize the detection ability for small cracks. In addition, the momentum method can be introduced, with a momentum coefficient of 0.9, to accelerate the gradient descent convergence. This optimization strategy can improve the adaptability of the model to complex defect morphology. Preferably, when the optimized model makes predictions on test data, it is necessary to ensure that the distribution of the test data is consistent with that of the training data. In an embodiment, the test set can include 200 high-resolution bridge point cloud images that have not been seen before, with a resolution of 512x512. The optimized model infers each image and outputs a probability distribution map of cracks and corrosion.

[0059] For example, the model may identify the crack probability of a certain region in an image as 0.85 and the corrosion probability as 0.12. Based on a probability threshold of 0.8, it can be determined that the region is a crack defect, and its spatial coordinates are extracted. The coordinate information is further mapped to the image to generate a second defect distribution map. This distribution map directly shows the location of the defect, making it easy for engineers to quickly locate the problem area. It can be understood that the generation of the second defect distribution map depends on the generalization ability of the model. In a possible implementation, the performance of the model can be evaluated through cross-validation.

[0060] For example, the data set is divided into 5 folds, 4 folds are used for training and 1 fold is used for validation, repeated 5 times, and the average detection accuracy is calculated. If the accuracy is above 90%, the model can be used for actual bridge detection. This method ensures the stability of the model on different data subsets, providing a guarantee for generating high-quality defect distribution maps.

[0061] For example, the combination of data augmentation, model training and optimized prediction can significantly improve the robustness and accuracy of defect detection. In practical applications, engineers can use the second defect distribution map to quickly locate high-risk areas on the surface of the bridge and prioritize maintenance work. This method supports bridge safety management through technical means and has high practical value. In order to further improve the recognition accuracy of the previous step, in this content, in order to further supplement the deep learning model, enhance the guidance of the spatial coordinates of the deep learning model defects, and strengthen the sensitivity of the model to key parts and defects, a "recognition-feedback-optimization" closed loop is formed to solve the long-tail distribution problem.

[0062] S105, distribute the calculation task of the second defect distribution map through a distributed control architecture, use a multi-node parallel computing method to process high-load data, extract the defect density and priority of key parts from the data, and obtain a first priority list.

[0063] The computing task of the second defect distribution map is decomposed by the distributed control architecture, a task allocation algorithm is used to distribute the task to a multi-node computing cluster to obtain a task allocation result. According to the task allocation result, a multi-node parallel computing method is used to process high-load data, a MapReduce algorithm is used to perform partition calculation on the defect distribution map to obtain partition defect data. If the partition defect data meets a preset integrity threshold, a data aggregation method is used to merge the partition defect data to generate a complete second defect distribution map. The defect area of the key part is extracted from the second defect distribution map by image segmentation technology, the defect density of each area is calculated, and the defect density distribution is obtained. If the defect density exceeds a preset density threshold, the key parts are sorted according to the density values to generate a preliminary priority sequence. A K-means clustering algorithm is used to classify the key parts in the preliminary priority sequence to determine high-priority defect parts and generate a first priority list.

[0064] For example, the distributed control architecture can effectively handle the high-load computing requirements of the second defect distribution map by decomposing complex tasks into multiple subtasks and distributing them to computing nodes. The distributed control architecture decomposes the defect detection task into subtasks such as image preprocessing, feature extraction, and defect classification. Each subtask is assigned to a different computing node, and a unified scheduling mechanism ensures task balancing.

[0065] For example, assuming that the radar detection equipment generates 1000 high-resolution images per second, the distributed architecture can distribute these images to 10 nodes, with each node processing 100 images, thereby greatly improving efficiency. The task allocation algorithm is the core of the distributed system and is used to optimize the utilization of computing resources.

[0066] In one possible implementation, a greedy algorithm based on load balancing is used to preferentially assign tasks to idle or low-load nodes.

[0067] Specifically, assuming that there are 5 computing nodes in the cluster, and the current loads are 20%, 30%, 50%, 10%, and 40%, respectively, the algorithm will preferentially assign the new task to the node with a load of 10%.

[0068] It should be noted that the task allocation also needs to consider the communication delay between nodes to ensure that data transmission does not become a bottleneck. For example, in bridge defect detection, the task allocation algorithm will dynamically adjust the allocation strategy according to the image data volume and node processing capacity to ensure real-time performance. The multi-node parallel computing method accelerates the generation of the defect distribution map by simultaneously processing multiple data partitions. Preferably, a heterogeneous computing cluster is used, combining CPU and GPU nodes, with the CPU handling logical scheduling and the GPU accelerating image processing.

[0069] In one embodiment, the bridge image data is divided into 100 partitions, each containing 500 point cloud images, 10 GPU nodes are processed in parallel, and each node processes 10 partitions. This way, the calculation time is significantly shortened.

[0070] It can be understood that parallel computing requires an efficient synchronization mechanism to avoid data conflicts between nodes. For example, through a distributed lock mechanism, it is ensured that the partition data processed by each node is independent and complete. The MapReduce algorithm is used for partition computing of defect distribution maps and is suitable for processing large-scale data sets. For example, in defect detection, the Map stage maps image data to feature vectors, and the Reduce stage summarizes feature vectors to generate partition defect data.

[0071] Specifically, MapReduce can divide the defect points on the surface of the bridge into 1000 grids according to the area, the Map task extracts the defect features of each grid, and the Reduce task summarizes to generate partition defect distribution.

[0072] Preferably, MapReduce ensures the reliability of the calculation through a fault-tolerant mechanism, so that even if a node fails, the task can be redistributed to other nodes.

[0073] In one possible implementation, a hierarchical aggregation strategy is adopted, which first locally merges the defect data of adjacent partitions and then globally summarizes.

[0074] For example, the defect data of 100 partitions is first merged into 50 groups, and then gradually merged into a complete distribution map.

[0075] It should be noted that the data integrity needs to be verified during the aggregation process to ensure that there is no data loss or duplication.

[0076] Illustratively, assuming that 50 defect points are detected on a bridge, the total number needs to be consistent after aggregation, and the spatial position of the distribution map is accurate. Image segmentation technology is used to extract the defect area of the key part to calculate the defect density.

[0077] Preferably, a deep learning-based segmentation algorithm is used to separate the defect area from the background. Specifically, the segmented area can calculate the number of defects per unit area, and this method can accurately locate high-risk areas. The K-means clustering algorithm is used to classify key parts to generate a first priority list. In one embodiment, the defect areas of the bridge are divided into high, medium and low priority according to the defect density and the importance of the location. For example, the central area is classified as high priority due to its key function, and the edge area is classified as low priority. Specifically, assuming there are 50 key parts, K-means clustering can divide them into 3 categories according to the density value (such as 10, 5, 2 per square meter) to generate a priority list to guide subsequent repair work.

[0078] It can be understood that the accuracy of clustering depends on feature selection, which needs to consider defect density and area function. This step is based on the high-precision identification described above as a basic step, and innovatively combines identification and related parallel computing, allowing parallel computing to calculate according to bridge structure partitioning through the spatial continuity of the high-precision defect distribution map (such as crack extension path), ensuring the physical reasonableness of task splitting. Further dividing the calculation area and merging the area results avoids the cross-area data exchange bottleneck.

[0079] S106, according to the defect density and spatial coordinates in the first priority list, a genetic algorithm is used to optimize the unmanned aerial vehicle inspection path, and the flight trajectory is dynamically adjusted through real-time environmental data fusion to obtain a first path planning.

[0080] The defect density data and spatial coordinate data of the inspection area are obtained, and the inspection area is divided by a preset grid division method to obtain a region division result. According to the region division result and the spatial coordinate data, a genetic algorithm is used to calculate an initial inspection path to obtain an initial path planning. Real-time environmental data is obtained, and if the wind speed or obstacle position in the environmental data changes, the initial path planning is updated through an environmental data fusion method to obtain a dynamically adjusted trajectory. For the dynamically adjusted trajectory, a flight trajectory is generated by using a path smoothing processing to obtain a smoothed flight trajectory. According to the smoothed flight trajectory, control parameters are calculated by using a UAV control instruction generation method to obtain flight control instructions. The actual flight data of the UAV is obtained, and if the deviation between the actual flight data and the smoothed flight trajectory exceeds a preset threshold, the flight control instructions are adjusted by using a deviation correction method to obtain corrected control instructions. According to the corrected control instructions, a path optimization result generation method is used to update the first path planning to obtain an optimized path planning.

[0081] Illustratively, when obtaining the defect density data and spatial coordinate data of the inspection area, data can be collected by laser radar and infrared sensors.

[0082] For example, within a bridge inspection area, lidar scans the bridge surface to generate point cloud data and calculates defect density; infrared sensors record spatial coordinates with centimeter-level accuracy. Defect density data reflects the degree of cracking or wear on the bridge surface, while spatial coordinate data provides the basis for subsequent path planning.

[0083] It should be noted that data collection must be conducted under stable weather conditions to ensure accuracy.

[0084] Specifically, the inspection area is divided using a preset grid division method, which can be a uniform grid division.

[0085] For example, a 1000m x 1000m inspection area is divided into 100 10m x 10m grids, with each grid recording the average defect density and center coordinates. The area division provides a basis for path planning, prioritizing the inspection of grids with high defect density. Preferably, the grid size can be dynamically adjusted according to equipment performance; smaller drones can use finer 5m x 5m grids. A genetic algorithm is used to calculate the initial inspection path, with a population size of 50 and 100 iterations, selecting the shortest path covering high defect density grids. For example, the algorithm prioritizes connecting grids with a defect density higher than 0.5, generating an initial path approximately 2000m long. The genetic algorithm optimizes path coverage efficiency by simulating natural selection. It should be noted that the initial path must consider the drone's endurance to avoid excessively long paths. For example, when acquiring real-time environmental data, wind speed sensors and cameras can monitor environmental changes. Suppose during inspection, the wind speed is detected to increase from 5m / s to 10m / s, and the camera identifies temporary obstacles such as birds. Environmental data fusion methods integrate sensor data through weighted averaging to update path planning. For example, it avoids high-wind-speed areas, bypasses obstacles, and generates dynamically adjusted trajectories, increasing the path length to 2200 meters. Dynamic adjustments ensure the safe operation of the drone.

[0086] For dynamic adjustment trajectory, path smoothing processing is adopted, and a smooth flight trajectory can be generated by cubic spline interpolation. For example, the adjusted trajectory contains sharp corners, and after smoothing processing, the corner radius is increased to 2 meters, and the trajectory is more consistent with the flight characteristics of the unmanned aerial vehicle. The smooth flight trajectory reduces the control difficulty and improves the flight stability. Specifically, the control parameters are calculated by the unmanned aerial vehicle control instruction generation method, and the speed and yaw angle instructions can be generated according to the smooth flight trajectory. For example, the speed of the straight line segment is set to 5 meters per second, and the speed of the turning segment is reduced to 3 meters per second, and the yaw angle is adjusted to 30 degrees. The control parameters ensure that the unmanned aerial vehicle accurately follows the trajectory. Preferably, the instruction generation considers the load of the unmanned aerial vehicle to avoid overload. It can be understood that when the actual flight data of the unmanned aerial vehicle is obtained, the actual position can be recorded by the GPS and inertial navigation system. If the actual position deviates from the smooth flight trajectory by more than 0.5 meters, the control instruction is adjusted by a deviation correction method. For example, the deviation caused by the crosswind is detected, the correction method increases the yaw angle by 5 degrees, and the corrected control instruction is generated. The correction instruction improves the path tracking accuracy. For example, according to the corrected control instruction, the path optimization result generation method is used to update the path planning by using the simulated annealing algorithm. For example, the optimized path length is shortened to 2100 meters, avoiding high wind speed areas while covering all high defect density grids. Optimizing the path planning improves the inspection efficiency and prolongs the service life of the unmanned aerial vehicle.

[0087] In this content, the above defect density clustering result is taken as a guide, the defect recognition is taken as a part of the optimization algorithm, the defect is directly mapped as the fitness function weight of the genetic algorithm, and the high-risk areas are driven to be covered by the path, and the areas where more defects may occur are further inspected in detail and comprehensively. It can ensure that the path planning target is accurately aligned with the structural safety demand.

[0088] S107, if the key part coverage rate of the first path planning is lower than the preset threshold, the parameters of the genetic algorithm are adjusted by the reinforcement learning algorithm, the path is regenerated and the coverage rate is verified, and the second path planning is obtained.

[0089] If the path coverage rate of the key part is lower than the preset threshold, the parameter adjustment strategy of the genetic algorithm is obtained by the reinforcement learning algorithm, and an optimized parameter set is obtained. The genetic algorithm is driven by the optimized parameter set to generate a new path set, and a candidate path planning is obtained. For the candidate path planning, the path coverage rate of the key part is calculated by using the coverage rate evaluation model, and coverage rate data is obtained. If the coverage rate data is lower than the preset threshold, the reinforcement learning algorithm is triggered again to adjust the parameters by the iteration control module, and an updated parameter set is obtained. According to the updated parameter set, the path planning is regenerated and the coverage rate is verified, and a verification path planning is obtained. By comparing the coverage rates of the verification path planning and the historical path planning, it is judged whether the optimization termination condition is reached, and the final path planning, i.e. the second path planning, is obtained.

[0090] Exemplarily, in the case that the path coverage of the critical part is lower than the preset threshold, the reinforcement learning algorithm is used to optimize the genetic algorithm parameters. The core of reinforcement learning is to learn the optimal strategy through trial and error. Assuming that the unmanned aerial vehicle inspects a bridge, the critical part such as the bridge joint needs high coverage. The reinforcement learning algorithm takes the coverage rate as the reward function, and the initial state is the current path coverage data. The action space is the parameter adjustment range of the genetic algorithm, such as the crossover rate and the mutation rate. In one possible implementation, the reinforcement learning adopts the Q-learning method. After initializing the Q table, the parameters are updated through multiple iterations to obtain an optimized parameter set. For example, the initial crossover rate is 0.7 and the mutation rate is 0.05. After reinforcement learning, the crossover rate is adjusted to 0.8 and the mutation rate is adjusted to 0.03, which improves the path generation efficiency. Specifically, the genetic algorithm is driven by the optimized parameter set to generate a new path set. The genetic algorithm is based on population evolution, and the path with high coverage rate is preferred. Preferably, the population size is set to 100, the iteration number is 50, and the fitness function is mainly based on the critical part coverage rate. Assuming that the bridge is divided into a 10x10 grid, and the defect area coverage rate needs to reach 90%. After the new path set is generated, the coverage rate is improved from 85% to 88%. The candidate path planning is thus formed, which contains multiple paths covering the critical part. It should be noted that the path planning needs to consider the endurance limit of the unmanned aerial vehicle, such as the single flight time not exceeding 30 minutes. In one embodiment, the coverage rate evaluation model uses the grid statistical method to calculate the critical part coverage rate. The model divides the inspection area into grids and counts the number of times the unmanned aerial vehicle passes through the critical part. For example, the defect area grid coverage needs to reach 5 times or more. If the evaluation coverage rate is only 88%, which is lower than the 90% threshold, the iteration control module triggers the reinforcement learning to adjust the parameters again. After updating the parameter set, the crossover rate is adjusted to 0.85 and the mutation rate remains 0.03, and the path planning is regenerated. The verification path planning shows that the coverage rate reaches 91%, meeting the requirements. It can be understood that the verification path planning is compared with the historical path planning to determine the optimization termination condition. The termination condition can be set as stable coverage rate or reaching the maximum iteration number. For example, the historical path coverage rate is 85% and the verification path is 91%, which has a significant improvement, and the coverage rate changes less than 1% for three consecutive iterations, so the optimization is terminated and the second path planning is obtained. This comparison ensures the stability of the path planning and avoids over-optimization. For example, the final path planning needs to cover different defect areas, and the path length is controlled within 10 kilometers, and the endurance time is 25 minutes. The reinforcement learning optimizes the parameters, the genetic algorithm generates the path, the coverage rate is evaluated, and the iteration control forms a closed loop to ensure that the critical part coverage rate meets the requirements. Preferably, the path planning can also integrate wind speed data to ensure that the unmanned aerial vehicle avoids strong wind areas, further improving the inspection efficiency.

[0091] In the content, the path set generated by the genetic algorithm provides an action space for reinforcement learning (such as discretization of crossover / mutation operations), and Q-learning guides the evolution direction of parameters through a reward function (coverage), which further improves the effectiveness of the genetic algorithm through the above-mentioned reinforcement learning method, thereby improving the path planning effect and the effectiveness of the inspection.

[0092] In S108, the second path planning is used to guide the UAV to perform the inspection task, and the second defect distribution map is updated in real time by using the radar scanning, and the third data set is obtained by extracting the newly added defect data from the second defect distribution map.

[0093] The second path planning is used to generate a set of inspection instructions, and the UAV is driven to perform the inspection task along the specified path to obtain the inspection trajectory data. The radar scanning module is used to scan the inspection area in real time, and the second defect distribution map is generated by obtaining the environmental reflection signal from the inspection trajectory data. The feature extraction algorithm is applied to the second defect distribution map to identify the newly added defect area, and the third data set is obtained by classifying and arranging the newly added defect feature set.

[0094] For example, the process of generating the inspection instruction set for the second path planning. For example, the second path planning can be decomposed into a series of consecutive waypoint coordinates and flight parameters through path resolution. For example, the path planning covers a certain bridge area, generating an instruction set containing 100 waypoints, each waypoint including latitude, longitude, height and flight speed, such as waypoint 1 (120.1234, 30.5678, 50 meters, 5 meters / second). Specifically, the instruction set also embeds the trigger conditions of the inspection task, such as automatically starting the high-definition camera shooting at the key position. Preferably, the instruction set will be transmitted to the UAV control system through the wireless communication module to ensure that the UAV flies accurately along the path. This way can effectively drive the UAV to perform the inspection task and generate continuous inspection trajectory data. In the generation of inspection trajectory data. For example, the UAV flies along the surface of the bridge, records the GPS coordinates, flight attitude and timestamp in real time, forming a trajectory data set. In one embodiment, the trajectory data is sampled once per second, including position (latitude and longitude), height and yaw angle, such as (120.1235, 30.5679, 50.5 meters, 45 degrees). It should be noted that the accuracy of the trajectory data directly affects the accuracy of the subsequent defect distribution map, so it is usually combined with an inertial navigation system to correct the deviation. This high-precision trajectory data provides a reliable basis for the analysis of environmental reflection signals. When using a radar scanning module for real-time scanning. Specifically, the millimeter wave radar can be used to scan the physical structure in the inspection area, such as the surface of the bridge or the bridge connection point. For example, the radar emits 10 beams per second, receives the environmental reflection signal, and generates a signal data set containing distance, angle and intensity. For example, the reflection signal shows that there is an abnormal echo on the surface of the bridge, and the intensity is 20% higher than the normal value. In one possible implementation, the signal data can be filtered by a filtering algorithm to remove noise and further improve the clarity of the defect distribution map. This method can efficiently capture abnormal changes in the environment. In the process of generating the second defect distribution map.

[0095] It can be understood that the reflected signal is converted into a two-dimensional or three-dimensional defect distribution map through a spatial mapping algorithm. For example, the distribution map shows that there is an abnormal area of 2m x 1m on the top of a certain bridge, and the intensity anomaly value is concentrated in the center of the area. Preferably, the distribution map will mark the defect probability of the key parts, such as the defect probability of a certain area is 85%. This visual distribution map provides an intuitive basis for subsequent feature extraction. The feature extraction algorithm identifies the newly added defect area. For example, a convolutional neural network can be used to analyze the defect distribution map and extract features such as texture, shape, and intensity. In one embodiment, the algorithm identifies that a crack feature is newly added in the middle area of a certain bridge, and the feature set includes a crack length of 0.5m, a width of 0.02m, and an abnormal value of 0.1m. In particular, feature extraction will preferentially focus on key parts, such as different connection points of a bridge, because defects in these areas have a greater impact on safety. This precise feature extraction helps to quickly locate newly added defects. When the data processing module classifies and organizes the feature set of the newly added defects. For example, the feature set can be divided into three categories of cracks, corrosion, and loosening according to the defect type, and divided into light, medium, and severe according to the severity. For example, a certain crack is classified as medium to severe, with a length of 0.5m and located at a key connection point of the bridge. Preferably, the classification result will generate a third data set containing defect ID, location and priority, such as (ID001, a certain area on the upper surface of the bridge, high priority). In one possible implementation, the third data set is also compared with the historical defect database to filter out newly added defects. This classification and organization method is convenient for subsequent maintenance decisions.

[0096] In this step, the above effective path planning result is used as the basis, and the smooth flight trajectory is combined to reduce radar scanning jitter and ensure the spatial consistency of point cloud data. Through real-time data quality, the requirements of defect incremental analysis are met, and the defects of the bridge are further identified and the missed detection of the bridge defects is prevented as much as possible, and the effectiveness of the inspection is improved.

[0097] S109, according to the feature value of the third data set, using a convolutional neural network algorithm to classify and locate the newly added defects, update the second defect distribution map and the first priority list, and obtain the final inspection result.

[0098] The original data is obtained from the third data set, and feature values are extracted through data processing to obtain a feature matrix. A convolutional neural network is used to infer the feature matrix, output the classification result and positioning coordinates of the new defect, and obtain a defect feature set. If the defect feature set contains a classification result, the second defect distribution map is updated according to the classification result to obtain an updated distribution map. If the defect feature set contains positioning coordinates, the second defect distribution map is updated according to the positioning coordinates to obtain an enhanced distribution map. Through the classification result and positioning coordinates in the defect feature set, the first priority list is updated according to the preset priority sorting rule to obtain a sorted priority list. According to the enhanced distribution map and the sorted priority list, a final inspection result is generated to obtain inspection output data. For the inspection output data, a clustering algorithm is used to group the defect distribution to obtain a defect grouping result.

[0099] For example, the original data obtained from the third data set usually involves various sensor data collected from the inspection task. For example, point cloud data or image data generated by radar scanning, which contains preliminary information about defects. In one possible implementation, the data processing module will first preprocess the original data, such as denoising and format standardization, to ensure the accuracy of subsequent feature extraction.

[0100] Specifically, noise points in the radar point cloud can be removed by a filtering algorithm, and the data can be converted into a unified format to obtain a clean original data set. This step lays the foundation for subsequent feature extraction. The process of extracting feature values through data processing to generate a feature matrix aims to extract meaningful patterns from the original data.

[0101] For example, edge detection algorithms can be used to extract shape features of defects from image data or depth and density features from point cloud data. These feature values are organized into a feature matrix, with each row representing a defect sample and each column corresponding to a feature type. For example, a feature matrix may contain 100 defect samples, each with 5 features such as size, depth, etc. This provides a structured input for subsequent inference. A convolutional neural network is used to infer the feature matrix, outputting classification results and positioning coordinates. In one embodiment, the convolutional neural network can be designed as a multi-layer structure containing convolutional layers and fully connected layers to identify defect types (such as cracks or corrosion) and calculate two-dimensional coordinates of defects in the inspection area.

[0102] For example, the network can output a defect as “crack” with coordinates (50, 30). It should be noted that the network needs to be trained in advance on the labeled data set to ensure the reliability of the inference. This step realizes the accurate classification and positioning of defects. If the defect feature set contains classification results, update the second defect distribution map. Preferably, the classification results can be mapped to the distribution map, for example, mark “crack” as a red area to generate an updated distribution map. This facilitates intuitive display of defect type distribution. If the positioning coordinates are included, an enhanced distribution map is generated. Specifically, the specific position of the defect can be marked on the distribution map, such as adding a marker point at coordinates (50, 30). This way enhances the spatial information of the distribution map, facilitating subsequent analysis. Through the classification results and positioning coordinates in the defect feature set, combined with the priority sorting rule, the first priority list is updated.

[0103] In one possible implementation, the priority rule can be sorted according to the severity of the defect and the importance of the location. For example, a large crack near a critical device has a higher priority than a small corrosion point in the edge area. The sorted priority list may place a certain crack at the top because of its large size and location in the core area. This provides a basis for optimizing the inspection task. According to the enhanced distribution map and the sorted priority list, the final inspection result is generated. For example, the inspection output data can include the total number of defects, type distribution and priority sorting, which facilitates maintenance personnel to develop repair plans. It can be understood that this step integrates all analysis results to form an operational output. For the inspection output data, a clustering algorithm is used to group the defect distribution.

[0104] In one embodiment, the K-means algorithm can be used to group defects by spatial location. For example, 100 defects can be divided into 3 groups, each representing a concentrated area such as the top, side, etc. of the device. This grouping facilitates targeted maintenance and optimizes resource allocation.

[0105] The deep learning model of the incremental defect, the classification result reversely updates the K-means clustering center, so that the priority list dynamically adapts to the structural state change, as the content of feedback, realizes the self-evolution cycle of “inspection-diagnosis-decision”.

[0106] In the above, the above steps are closely related, and the technical advantages are improved through the combination of single or multiple technologies, including the following: cross-layer cooperative mechanism: the multi-frequency signal fusion quality of the perception layer directly determines the upper limit of the recognition of the cognitive layer CNN; the path planning of the decision layer takes the defect distribution of the cognitive layer as input and outputs guidance for the action of the execution layer; the execution layer real-time data backflow reconstructs the decision basis, forming a dynamic optimization loop. Technical principle interlocking: anti-interference processing (time-frequency analysis) and CNN (spatial convolution) form an orthogonal feature extraction system; distributed computing (data parallelism) and reinforcement learning (policy optimization) jointly solve complex decision-making problems; genetic algorithm (global search) and real-time path adjustment (local correction) complement each other to cope with dynamic environments.

[0107] Experimental verification examples are given: For the annual inspection verification scheme of a cross-sea bridge (total length 8.2km), a comparative group is set up as a traditional manual inspection team (10 people) + single-frequency radar unmanned aerial vehicle; the experimental group is set up as the system (3 unmanned aerial vehicles + distributed computing cluster), and the results show that the inspection time, defect detection amount, minimum crack detection, void positioning error, high-risk defect omission rate and total cost of the traditional method comparative group are 15 days, 89, 2cm, ±5cm, 12%, 52w respectively, and the inspection time, defect detection amount, minimum crack detection, void positioning error, high-risk defect omission rate and total cost of the method of the system are 3 days, 157, 3mm, ±2cm, 0% and 18w respectively; at the same time, for the detection of internal steel bar corrosion voids (diameter 15cm, depth 20cm) of the main bridge pier, manual inspection cannot find. Compared with the traditional method, the system has great improvement in different indicators, and the system solves the problems of complex environment perception, real-time decision-making and efficient coverage, provides a full-stack solution for infrastructure intelligent operation and maintenance, and significantly improves the effectiveness of bridge inspection.

[0108] As shown in Figure 2 The application provides an unmanned aerial vehicle bridge inspection system based on radar scanning, which mainly comprises:

[0109] A three-dimensional point cloud and defect signal acquisition module is used to obtain three-dimensional point cloud data and surface defect signals of the bridge structure through radar scanning technology, and a multi-frequency radar fusion method is used to extract characteristic values from different frequency signals to obtain a first data set;

[0110] An anti-interference processing module is used to process the first data set by using an anti-interference algorithm if the first data set is affected by environmental noise or electromagnetic interference, to suppress noise and separate interference signals by using an adaptive filter, and to obtain a second data set;

[0111] A defect classification and positioning module is used to classify and position the bridge surface defects according to the characteristic value distribution of the second data set by using a convolutional neural network algorithm, to extract the spatial coordinates of defects such as cracks and corrosion from the second data set, and to obtain a first defect distribution map.

[0112] a data enhancement optimization module, configured to, if the detection accuracy of the first defect distribution map is lower than a preset threshold, expand the second data set by a data enhancement method, re-input the convolutional neural network algorithm for iterative optimization, and obtain a second defect distribution map;

[0113] a distributed computing module, configured to distribute the calculation task of the second defect distribution map by a distributed control architecture, process high-load data by a multi-node parallel computing method, extract the defect density and priority of key parts therefrom, and obtain a first priority list;

[0114] a path planning module, configured to optimize the unmanned aerial vehicle inspection path by a genetic algorithm according to the defect density and spatial coordinates in the first priority list, dynamically adjust the flight trajectory by real-time environmental data fusion, and obtain a first path planning;

[0115] a path optimization module, configured to, if the key part coverage rate of the first path planning is lower than a preset threshold, adjust the parameters of the genetic algorithm by a reinforcement learning algorithm, re-generate the path and verify the coverage rate, and obtain a second path planning;

[0116] a real-time inspection updating module, configured to guide the unmanned aerial vehicle to perform the inspection task by the second path planning, update the second defect distribution map by real-time radar scanning, extract new defect data therefrom, and obtain a third data set;

[0117] a defect data updating module, configured to classify and locate the new defects by the convolutional neural network algorithm according to the characteristic values of the third data set, update the second defect distribution map and the first priority list, and obtain a final inspection result.

[0118] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for unmanned aerial vehicle (UAV) bridge inspection based on radar scanning, characterized in that, The method comprises the following steps: Obtaining original data of the bridge through radar scanning, fusing the original data to obtain a first data set; Preprocessing the first data set to obtain a second data set; Identifying the second data set through a deep learning model to obtain a defect distribution map; According to the defect distribution map, a priority processing list is obtained, and the unmanned aerial vehicle inspection path is optimized according to the priority processing list to obtain a path planning result, and the unmanned aerial vehicle inspection task is adjusted according to the path planning result, and a third data set is obtained in real time through radar scanning; Classifying and positioning the third data set through a deep learning model to obtain a final inspection result.

2. The method of claim 1, wherein the first data set is obtained by: Obtaining three-dimensional point cloud data of the bridge through a multi-frequency radar carried by the unmanned aerial vehicle, fusing the multi-frequency three-dimensional point cloud data, and extracting the spatial domain features of the fused data to obtain the first data set.

3. The method of claim 1, wherein the second data set is obtained by: Judging the first data set, and when the first data set has noise or interference, preprocessing the first data set through an anti-interference method and an adaptive filter to obtain the second data set.

4. The method of claim 1, wherein the defect distribution map is obtained by: Identifying the second data set through a deep learning model to obtain the category and location information of the bridge surface defects, extracting the spatial coordinates of the bridge label defects according to the category and location information of the bridge surface defects, visualizing the defect distribution map according to the spatial coordinates, and when the accuracy of the defect distribution map is lower than a threshold, enhancing and expanding the second data set, and iteratively optimizing the deep learning model to obtain and update the defect distribution map.

5. The method of claim 1, wherein the priority processing list is obtained by: Distributing the calculation tasks of the defect distribution map through a distributed control architecture, and calculating the distributed tasks through a multi-node calculation method to obtain the defect density and repair priority of the key parts of the bridge, and obtain the priority processing list.

6. The method of claim 1, wherein the path planning result is obtained by: Taking the spatial coordinates of the defect density of the key parts of the bridge in the priority processing list as reference points, optimizing the unmanned aerial vehicle inspection path through an optimization algorithm, and adjusting the optimized unmanned aerial vehicle inspection path through real-time environmental data to obtain the path planning result, and when the key part coverage rate of the path planning result is lower than a threshold, adjusting the parameters of the optimization algorithm through a reinforcement learning method to regenerate the path planning result.

7. The method of claim 1, wherein the final inspection result is obtained by: ​ ​ ​ ​ ​ ​ The defect distribution map is updated through the radar scanning result, and new defect data is extracted from the updated defect distribution map to obtain a third data set. The new defect data in the third data set is classified and positioned through a convolutional neural network. According to the defect and positioning classification result, the defect distribution map and the priority list are updated to obtain a final inspection result.

8. The method of claim 6, wherein, The regeneration process of the path planning result includes: The key position coverage rate of the path planning result is judged, and when it is lower than a threshold value, the parameters of the optimization algorithm are adjusted through reinforcement learning to obtain an optimization parameter set. The optimization parameter set is used to drive the optimization algorithm to generate a new path set to obtain a candidate path planning. The path coverage rate of the key position of the candidate path planning is calculated to obtain coverage rate data. The candidate path planning is iteratively regenerated according to the coverage rate data until the coverage rate of the candidate path planning and the historical path planning exceeds the threshold value to obtain the path planning result.

Images