Aviation equipment detection and maintenance method and equipment based on target detection and medium

Abstract

The invention discloses an aviation equipment detection and maintenance method and equipment based on target detection and a medium, and relates to the technical field of aviation equipment detection, and the method comprises the steps: inputting a correction image set into a pre-trained target detection network, carrying out the multi-scale feature analysis and edge texture recognition, and generating a part detection result; establishing a component topological relation according to a component detection result, and comparing the component topological relation with a standard configuration by using a graph structure analysis method to generate topological consistency information; performing defect positioning and quantitative analysis on the topological consistency information, extracting a defect position, a defect type and a defect quantitative index, tracking an extension track of the defect quantitative index along with time through continuous time frame difference, and generating a defect parameter; and performing risk assessment on the defect parameters, generating and executing a maintenance scheme, verifying the maintenance effect through real-time image detection, and forming a detection and maintenance closed loop. According to the invention, the accuracy and integrity of aviation equipment detection are finally improved.

Description

Technical Field

[0001] This invention relates to the field of aviation equipment testing technology, and in particular to a method, equipment and medium for testing and maintaining aviation equipment based on target detection. Background Technology

[0002] With the rapid development of the aviation industry, the structure of modern aviation equipment is becoming increasingly complex, and its operational safety and reliability are directly related to flight safety and operational efficiency. Existing aviation equipment inspection and maintenance technologies have made progress in several areas, such as using high-resolution imaging, ultrasonic testing, infrared thermal imaging, and sensor fusion to inspect the external structure and key components of aviation equipment. These inspection technologies can, to a certain extent, identify macroscopic defects or temperature anomalies on the surface of aviation equipment, and, combined with traditional maintenance procedures, allow for manual confirmation and handling of these anomalies.

[0003] In practical applications, existing detection methods still have certain limitations in scenarios with high structural complexity, irregular defect distribution, or complex environmental conditions. Traditional image recognition methods struggle to model the topological relationships between components, making it difficult to accurately determine whether component configurations have deviated, thus affecting the accuracy of defect localization and risk assessment. Furthermore, regarding the evolution trend of defects over time, most existing methods remain at a single-moment static detection level, lacking continuous parameter tracking and dynamic risk analysis, which makes it difficult to provide sufficient support for the generation of maintenance plans. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an aviation equipment inspection and maintenance method based on target detection, which solves the problems of incomplete defect identification, insufficient topology analysis and lack of maintenance verification in the prior art.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for the inspection and maintenance of aviation equipment based on target detection, which includes acquiring an initial multispectral image, performing geometric distortion correction and illumination normalization processing, and generating a corrected image set;

[0008] The corrected image set is input into a pre-trained target detection network for multi-scale feature analysis and edge texture recognition to generate part detection results.

[0009] Based on the component inspection results, the component topology relationship is established, and the graph structure analysis method is used to compare the component topology relationship with the standard configuration to generate topology consistency information.

[0010] Defect location and quantification analysis are performed on topology consistency information to extract defect location, defect type and defect quantification index. Then, through continuous time frame difference, the expansion trajectory of defect quantification index over time is tracked to generate defect parameters.

[0011] Risk assessment of defect parameters is conducted, repair plans are generated and implemented, and the repair effect is verified through real-time image detection, forming a closed loop of detection and repair.

[0012] As a preferred embodiment of the target detection-based aviation equipment inspection and maintenance method of the present invention, the specific steps of generating the calibration image set are as follows: performing temporal and spatial synchronization calibration on the initial multispectral image set to generate a synchronized image set;

[0013] Perform surface geometric correction on the synchronized image set to generate a geometrically corrected image set;

[0014] The geometrically corrected image set is subjected to illumination normalization, noise suppression, and edge texture enhancement to generate a clear feature image set;

[0015] Multi-angle and multispectral information fusion is performed on a set of clear feature images to generate a corrected image set.

[0016] As a preferred embodiment of the target detection-based aviation equipment inspection and maintenance method of the present invention, the pre-trained target detection network is obtained by iteratively training the network through supervised learning based on a diverse image dataset containing various components and defects of aviation equipment, and by optimizing the weights of the target detection network using multi-scale feature extraction and edge texture recognition.

[0017] As a preferred embodiment of the target detection-based aviation equipment inspection and maintenance method of the present invention, the specific steps for generating component inspection results are as follows: inputting the corrected image set into a pre-trained target detection network, performing multi-scale feature extraction, and generating a preliminary feature map;

[0018] Edge texture enhancement processing is performed on the preliminary feature map to generate an enhanced feature map. The enhanced feature map is then used for position and category recognition to generate preliminary detection results.

[0019] The non-maximum suppression algorithm is used to remove overlapping and redundant candidate boxes from the preliminary detection results and generate component detection results.

[0020] As a preferred embodiment of the target detection-based aviation equipment inspection and maintenance method of the present invention, the standard configuration is obtained by collecting design drawings, three-dimensional models and historical layout data of various components of aviation equipment, and establishing standard topological relationships between components based on structured analysis.

[0021] As a preferred embodiment of the target detection-based aviation equipment inspection and maintenance method of the present invention, the specific steps for generating topology consistency information are as follows: extracting the candidate box position and category probability of each component from the component inspection results, calculating the center coordinates and relative distance of each component as spatial relationships, constructing the components as component nodes, constructing the spatial relationships and functional associations between components as edges, and establishing a preliminary component topology graph.

[0022] The node feature encoding and edge weight calculation are performed on the preliminary component topology map to quantify the spatial relationships and functional associations between the components and generate an encoded topology map.

[0023] The encoded topology graph is compared with the standard configuration using graph matching. The degree of matching between nodes and edges is calculated using a similarity measurement algorithm to generate a topology consistency index.

[0024] The abnormal relationships between components are analyzed using topology consistency indicators, and components with displacement, missing parts, and incorrect installations are marked to generate topology consistency information.

[0025] As a preferred embodiment of the target detection-based aviation equipment inspection and maintenance method of the present invention, the specific steps for generating defect parameters are as follows: identifying abnormal component regions based on topology consistency information, extracting the bounding box and key points of each abnormal component, and generating defect locations;

[0026] The defect locations are classified and analyzed to extract morphological and texture features and identify the defect types.

[0027] Perform quantitative analysis on defect types, calculate defect geometry and strength indices, and generate quantitative defect indices;

[0028] The defect quantification index is compared with the continuous time series images in the calibration image set to calculate the difference, track the expansion trajectory of cracks and damage over time, and generate the defect expansion path.

[0029] Defect parameters are generated by combining defect location, defect type, defect quantification indicators, and defect propagation path.

[0030] As a preferred embodiment of the target detection-based aviation equipment inspection and maintenance method of the present invention, the specific steps of forming the inspection and maintenance closed loop are as follows: performing multi-dimensional risk assessment on defect parameters, quantifying the impact of defects on the safety and performance of aviation equipment, and generating a risk level and maintenance priority list;

[0031] Based on the risk level and maintenance priority list, a maintenance plan is generated by defining the maintenance sequence, required tools, operating procedures and expected repair results for each high-risk and priority maintenance component.

[0032] During the implementation of the maintenance plan, real-time image detection is carried out using the calibration image set to monitor the condition of the maintenance components and generate real-time detection data.

[0033] The real-time detection data is compared with the expected repair results in the repair plan to evaluate the repair completion and quality, and generate repair verification results.

[0034] Adjust subsequent maintenance plans based on maintenance verification results and implement corresponding repeat maintenance operations to form a closed loop of testing and maintenance.

[0035] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the target detection-based aviation equipment inspection and maintenance method described in the first aspect of the present invention.

[0036] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the target detection-based aviation equipment inspection and maintenance method described in the first aspect of the present invention.

[0037] The beneficial effects of this invention are as follows: By constructing a component topology map based on the component inspection results, and further performing node feature encoding and edge weight calculation to generate a quantifiable topology structure, and then comparing the topology structure with the standard configuration, the modeling of the relative positions, assembly relationships and functional associations between components is realized. This not only identifies defects in individual components, but also detects abnormalities such as assembly deviations, displacements, missing parts or incorrect installations. Furthermore, the inspection results are expanded from single-point identification to overall assembly consistency analysis, ultimately improving the accuracy and completeness of aviation equipment inspection. Attached Figure Description

[0038] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a flowchart of an aircraft equipment inspection and maintenance method based on target detection.

[0040] Figure 2 A flowchart for generating the corrected image set.

[0041] Figure 3 A flowchart generated from the component inspection results.

[0042] Figure 4A flowchart for generating topology consistency information. Detailed Implementation

[0043] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0044] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0045] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0046] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for inspecting and maintaining aviation equipment based on target detection, including the following steps:

[0047] S1. Acquire initial multispectral images, perform geometric distortion correction and illumination normalization processing, and generate a corrected image set.

[0048] S1.1 Perform temporal and spatial synchronization correction on the initial multispectral image set to generate a synchronized image set.

[0049] Specifically, multispectral imaging is used to photograph the surface of aerospace equipment, acquiring an initial multispectral image set covering multiple bands such as visible and infrared light. This initial multispectral image set includes spatial and spectral resolution information. During time synchronization correction of the initial multispectral image set, initial multispectral images of different bands within the same detection period are registered based on timestamp information, ensuring consistency across bands in the time dimension. For example, millisecond-level time deviations are corrected to a unified time point. During spatial synchronization correction of the initial multispectral image set, geometric registration methods are used to align the positions of pixels in images of different bands, mapping the same physical location in different bands to a unified coordinate system across all bands. For example, affine or projection transformation methods are used to control pixel errors to the sub-pixel level, thus ensuring that the initial multispectral image set remains consistent in both time and space, generating a synchronized image set.

[0050] S1.2 Perform surface geometric correction on the synchronized image set to generate a geometrically corrected image set.

[0051] Specifically, image regions containing surface feature points of aviation equipment are selected from the synchronized image set. Surface calibration point coordinates are extracted using an automatic detection method to obtain the correspondence between the two-dimensional pixel coordinates in the synchronized image set and the actual three-dimensional coordinates in space. A polynomial surface fitting method is used to establish the mapping relationship between pixel coordinates and spatial coordinates, for example, by generating a transformation matrix through second-order polynomial fitting. After establishing the mapping relationship, each image in the synchronized image set undergoes coordinate transformation according to the transformation matrix. Pixel resampling and interpolation correction are performed on distorted areas caused by surface curvature, for example, using bilinear interpolation for pixel filling. All images after surface geometric correction are unified onto the same geometric reference plane to obtain a geometrically corrected image set.

[0052] S1.3 Perform illumination normalization, noise suppression, and edge texture enhancement on the geometrically corrected image set to generate a clear feature image set.

[0053] Specifically, illumination normalization is performed on the geometrically corrected image set according to the spectral channels. Homomorphic filtering with logarithmic transformation and high-low frequency separation is used to separate the reflection and illumination components, and local contrast correction is performed on the brightness channel. Noise suppression and edge texture enhancement are performed on the illumination normalization results. First, non-local mean filtering is used to suppress salt-and-pepper noise and random noise. Then, Laplacian operator is used to sharpen or unsharpen the mask and guided filtering is used to improve edge sharpness and texture contrast. Color channel consistency correction and subpixel interpolation sharpening are performed on the enhanced image, and a clear feature image set is output according to the original geometric reference.

[0054] It should also be noted that the original geometric datum refers to the reference standard used to describe and constrain the geometry of a component before acquiring images of the aircraft equipment or measuring the component. It typically includes design dimensions, shape profiles, key point locations, and relative relationships between components, used for subsequent image correction, spatial positioning, and topology analysis. This ensures that measurement data or image features are compared and processed within a unified geometric framework, thereby improving the accuracy and reliability of inspection and analysis.

[0055] S1.4. Perform multi-angle and multispectral information fusion on the clear feature image set to generate a corrected image set.

[0056] Specifically, multi-angle images and multispectral channel images are extracted from the set of sharp feature images. Coarse registration is achieved through affine or projection transformation, followed by sub-pixel-level fine registration using phase correlation. Photometric normalization is performed on the registered multi-angle and multispectral channel images, and histogram matching or linear scaling correction is used to ensure consistent brightness distribution across different angles and bands. Histogram equalization is then used to adjust local contrast. Multi-scale wavelet transform decomposition is performed on the registered and photometric normalized multispectral channel images. Principal component analysis (PCA) is used to fuse the approximation coefficients, and detail coefficients are weighted and merged based on local energy and Laplacian response. Inverse transform is then performed to reconstruct the spectral fusion image. Simultaneously, local sharpness is assessed using the Laplacian operator. Pixel selection and weighted averaging are performed on multi-angle images at the same pixel location based on local sharpness indices to generate the fusion image. Guided filtering is applied to the fusion image to maintain edge consistency, and bilinear resampling is used to output the corrected image set.

[0057] S2. Input the corrected image set into the pre-trained target detection network, perform multi-scale feature analysis and edge texture recognition, and generate part detection results.

[0058] S2.1 The pre-trained target detection network is obtained by iteratively training it through supervised learning based on a diverse image dataset containing various components and defects of aviation equipment, and by optimizing the weights of the target detection network using multi-scale feature extraction and edge texture recognition.

[0059] Specifically, a diverse image dataset containing various components and defects of aviation equipment is acquired, and each diverse image is labeled, including the location of the component bounding box and the defect category information. The diverse image dataset is divided into a training set and a validation set, for example, the training set accounts for 80% and the validation set accounts for 20%. Under supervised learning, the images in the training set are input into the object detection network, and multi-scale feature extraction layers are used to perform convolution operations on features at different scales of the images to obtain multi-scale feature maps. Edge texture recognition processing is performed on the multi-scale feature maps, for example, using the Sobel operator to extract edge information, and the edge features are fused with the convolution features to generate an enhanced feature map. The cross-entropy loss function is used to compare the candidate box category probabilities predicted by the object detection network with the labeled information, and the weights of the object detection network are adjusted iteratively through stochastic gradient descent. After each iteration, the diverse images in the validation set are input into the object detection network to calculate the validation loss, and the convergence of the object detection network is monitored. The training is repeated iteratively until the training rounds are reached, for example, 500 rounds, to obtain a pre-trained object detection network capable of identifying various components and defects of aviation equipment.

[0060] S2.2 Input the corrected image set into the pre-trained target detection network to perform multi-scale feature extraction and generate a preliminary feature map.

[0061] Specifically, the corrected image set is input frame by frame into the input layer of the pre-trained object detection network, and the image size is adjusted to the example resolution, such as 1024×1024 pixels. In the multi-scale feature extraction layer, low-level, mid-level, and high-level features are extracted using different convolutional kernel sizes and strides to generate corresponding scale feature maps. The feature maps at each scale are normalized to ensure consistent feature amplitudes, and then non-linear mapping is performed using activation functions such as ReLU. The feature maps at different scales are spatially aligned by downsampling or upsampling operations, and local pooling is performed on the 3×3 pixel example window to enhance edge and texture information. The feature maps at each scale are concatenated according to the channel dimension to output a preliminary feature map of multi-scale fusion.

[0062] It should be noted that low-level feature maps mainly contain local details such as edges and textures, mid-level feature maps reflect the shape, outline and local structural patterns of the component, and high-level feature maps contain semantic information and the spatial relationship of the overall component, which helps to identify the global features of complex components and defects.

[0063] S2.3. Perform edge texture enhancement processing on the preliminary feature map to generate an enhanced feature map, and perform position and category recognition on the enhanced feature map to generate preliminary detection results.

[0064] Specifically, edge detection is performed on the initial feature map using convolutional kernels such as 3×3 or 5×5, while texture features are extracted using the Laplacian operator to generate an enhanced feature map. This enhanced feature map is then input into a candidate box generation network, which performs a sliding window operation on the two-dimensional plane of the enhanced feature map with a preset stride. The area covered by each sliding window is extracted as a candidate region, and the candidate box generation network performs bounding box regression on these regions to obtain component candidate boxes. Simultaneously, the classification branch in the candidate box generation network (e.g., a classification head formed by combining convolutional and fully connected layers) is used to discriminate each component candidate box, calculating the corresponding class probability distribution. This generates multiple component candidate boxes with class probability distributions on the enhanced feature map. Non-maximum suppression is used to remove overlapping and redundant candidate boxes, retaining the candidate boxes with the highest confidence. The center coordinates, width, height, and class labels of the retained candidate boxes are output to form the preliminary detection results.

[0065] It should also be noted that the candidate box generation network is a deep learning structure used to quickly locate potential target regions in an image. It generates a set of candidate boxes that may contain the target by sliding preset anchor boxes or prior boxes on the initial feature map and combining convolution operations to predict the existence probability and offset of each box, thus providing an initial region reference for subsequent accurate classification and localization.

[0066] The preset step size refers to the pixel interval in which the sliding window moves horizontally and vertically each time. The specific steps are as follows: starting from the upper left corner of the preliminary feature map, the window is moved horizontally to cover the entire row, and then moved downwards to cover all rows with a vertical step size. The process is repeated until the entire preliminary feature map area is traversed. The area corresponding to each window position is extracted as a candidate region, and the coordinates of the upper left corner of the window are recorded for subsequent candidate box positioning and offset calculation.

[0067] S2.4. Using the non-maximum suppression algorithm, overlapping and redundant candidate boxes are removed from the preliminary detection results to generate component detection results.

[0068] Specifically, the candidate boxes in the preliminary detection results are grouped by category and sorted from high to low according to the category probability of each candidate box, for example, the candidate box with the highest probability is placed at the top; the candidate box with the highest probability is selected from the sorted candidate box list as the reference box and added to the component detection result list; the intersection-union ratio (IoU) is obtained by calculating the ratio of the intersection area and the union area of ​​the reference box and the candidate box, which is used as the overlap index; when the IoU of the candidate box and the reference box is greater than the preset candidate box overlap threshold, the candidate box is removed from the candidate box list, and the candidate boxes with the IoU ratio lower than the candidate box overlap threshold are retained; the candidate box with the highest probability is selected from the remaining candidate box list as the new reference box, and overlapping redundant candidate boxes are removed in the same way until the candidate box list is empty; all the retained candidate boxes and their corresponding category probabilities are integrated to generate the component detection result.

[0069] It should also be noted that the steps for setting the preset candidate box overlap threshold include: selecting an appropriate candidate box overlap threshold between 0.5 and 0.7 based on the size characteristics and detection accuracy requirements of the aerospace equipment components. For example, setting the candidate box overlap threshold to 0.6 as an example value is used to determine whether the degree of overlap between candidate boxes needs to be removed, and applying the preset candidate box overlap threshold to remove redundant candidate boxes during the candidate box screening process of the non-maximum suppression algorithm.

[0070] S3. Based on the component inspection results, establish the component topology relationship, and use graph structure analysis method to compare the component topology relationship with the standard configuration to generate topology consistency information.

[0071] S3.1 Standard configuration is obtained by collecting design drawings, 3D models and historical layout data of various components of aviation equipment, and establishing standard topological relationships between components based on structured analysis.

[0072] Specifically, design drawings, 3D models, and historical layout data of various components of aviation equipment are collected and organized using structured analysis methods. Spatial location, size, connection relationship, and functional association information of each component are extracted. A standard topology table is constructed according to the spatial and functional relationships between components. Each node is labeled to represent each component, and edges represent the connections and functional associations between components, forming a complete standard topology relationship between components. The standard topology table is verified for consistency using a topology consistency verification method to ensure the accuracy and completeness of the weights of each component node and edge, generating a standard configuration.

[0073] S3.2 Extract the candidate box position and category probability of each component from the component detection results, calculate the center coordinates and relative distance of each component as spatial relationships, construct the components as component nodes, construct the spatial relationships and functional associations between components as edges, and establish a preliminary component topology graph.

[0074] Specifically, the candidate bounding box position and class probability of each part are extracted from the part detection results, using the coordinates of the top-left corner of the candidate bounding box. and the coordinates of the bottom right corner The coordinates of the candidate box center are calculated using the following expression:

[0075] ;

[0076] in, Indicates the center coordinates of the candidate box. This indicates that the top left corner of the candidate box is horizontal. Pixel coordinates of direction This indicates that the top left corner of the candidate box is vertical. Pixel coordinates of direction This indicates that the bottom right corner of the candidate box is horizontal. Pixel coordinates of direction This indicates that the bottom right corner of the candidate box is vertical. Pixel coordinates of direction Indicates that the center of the candidate box is horizontal. Pixel coordinates of direction This indicates that the center of the candidate box is vertical. Pixel coordinates of the direction;

[0077] The relative positional relationship between components is calculated using the Euclidean distance method, expressed as follows:

[0078] ;

[0079] in, Indicates the first The component node and the first The relative positional relationships between the components Indicates the first The center of each component node is horizontal Directional coordinates Indicates the first The center of each component node is vertical Directional coordinates Indicates the first The center of each component node is horizontal Directional coordinates Indicates the first The center of each component node is vertical Directional coordinates Indicates the index of the component node;

[0080] Each component is constructed as an independent component node based on the candidate box position, category probability, and size information. Edges between components are constructed based on the Euclidean distance calculation results and functional association information, and the weight of the edges is represented by the inverse distance ratio. By traversing all component nodes, the feature vectors of component nodes are calculated in turn, and the edge weights are generated by combining the Euclidean distance between component nodes and the functional association information. Feature encoding and normalization processing are performed on all component nodes and edges to form a complete preliminary component topology graph.

[0081] S3.3. Encode node features and calculate edge weights for the preliminary component topology map, quantify the spatial relationships and functional associations between components, and generate an encoded topology map.

[0082] Specifically, for each component node in the preliminary component topology graph, the spatial coordinates, candidate box category probabilities, and component size of the component node are extracted as component node feature vectors, and the component node feature vectors are normalized. The Euclidean distance between nodes is calculated based on the center coordinates of the component nodes. The Euclidean distance is combined with the category probability and functional association weights, and the edge weights are calculated according to the weighted method. The Euclidean distance weight is normalized by the reciprocal of the distance between component nodes, the category similarity weight is calculated by the intersection of probability distributions, and the functional association weight is assigned according to the adjacency relationship of the standard configuration. The component node feature vectors and edge weights are combined to generate the coded topology graph.

[0083] S3.4. Perform graph matching comparison between the encoded topology graph and the standard configuration, calculate the matching degree of nodes and edges through a similarity measurement algorithm, and generate a topology consistency index.

[0084] Specifically, the encoded topology graph is matched one by one with the component nodes and edges in the standard configuration. The spatial similarity of each component node is calculated and represented by the Euclidean distance of the component node center coordinates. The category matching degree is obtained based on the intersection of the candidate box category probability distributions. The edges between component nodes are matched, and the edge matching degree is calculated based on the spatial relationship and functional association weight of the edge connecting the component nodes. The component node matching degree and the edge matching degree are combined according to a weighted method to obtain the matching score of each pair of encoded topology graphs and standard configurations. The matching scores of all component nodes and edges are summarized to generate a topology consistency index.

[0085] S3.5 Analyze the abnormal relationships between components using topology consistency indicators, mark components with displacement, missing parts, and incorrect installations, and generate topology consistency information.

[0086] Specifically, based on the topology consistency index, the matching score of each component node is analyzed. The spatial deviation between the center coordinates of the component node and the center coordinates of the standard configuration node is calculated using the Euclidean distance method to obtain the positional deviation of the component node. The Euclidean distance between the center coordinates of the component node and the coordinates of the standard configuration node is calculated. When the Euclidean distance is greater than the positional deviation, the component is marked as a displacement anomaly. The component nodes in the coded topology graph are compared with the standard configuration nodes. When a component node in the coded topology graph fails to match a standard configuration node, the component is marked as a missing anomaly. When calculating the category similarity of successfully matched component nodes, if the category similarity is less than the category matching positional deviation, the component is marked as an incorrect installation anomaly. The marking results of displacement anomalies, missing anomalies, and incorrect installation anomalies are summarized to generate topology consistency information.

[0087] It should be noted that, based on the component inspection results, topological relationship modeling is combined, and the established component topological relationship is compared with the standard configuration using graph structure analysis methods to generate topological consistency information. This expands traditional target inspection from "single component identification" to overall modeling of "spatial position and assembly relationship between components." By establishing a quantifiable component topological structure and comparing it with the standard configuration, not only can single-point defects be identified, but also assembly deviations and correlation anomalies between components can be discovered. This achieves a breakthrough from local inspection to overall assembly consistency analysis, effectively improving the accuracy and completeness of inspection.

[0088] S4. Perform defect location and quantification analysis on topology consistency information, extract defect location, defect type and defect quantification index, and track the expansion trajectory of defect quantification index over time through continuous time frame difference to generate defect parameters.

[0089] S4.1 Identify abnormal component regions based on topology consistency information, extract the bounding box and key points of each abnormal component, and generate defect locations.

[0090] Specifically, based on the components marked as abnormal in the topology consistency information, the component node identifier, center coordinates, category information, and time index of the abnormal components are read. After locating the corresponding image in the calibration image set according to the time index, the region determined by the candidate box position and category probability in the component detection results is used as the initial bounding box. Canny edge detection is performed within the initial bounding box, and morphological closing and opening operations are performed to connect broken edges and remove isolated noise (e.g., the structuring element size is 3×3 pixels). Then, connected component analysis is performed to extract the connected component with the largest area, and the bounding box is updated according to the minimum bounding rectangle or rotated bounding rectangle. Within the updated bounding box, Harris corner detection is used to obtain candidate key points, and the coordinates of the candidate key points are refined using a sub-pixel corner precision method. The coordinates of the candidate key points are filtered according to the response intensity threshold and minimum spacing to obtain a set of stable key points. The four corner coordinates of the updated bounding box and the set of stable key points are recorded one-to-one according to the abnormal components, and the output is the defect location.

[0091] It should also be noted that the response intensity threshold is obtained as follows: the response intensity of candidate key points of abnormal components is sorted, and a quantile interval is selected as the threshold range. For example, the response intensity threshold is set to the top 20% to top 40% of the response intensity values ​​of candidate key points to retain more significant candidate key points. The minimum spacing is obtained as follows: a pixel-level range is set according to the structural size characteristics of the aerospace equipment components and the image resolution. For example, the minimum spacing is set to the example interval of 3 to 8 pixels to avoid candidate key points being too concentrated. The specific screening method is as follows: candidate key points are arranged from high to low response intensity, and only candidate key points with response intensity greater than the response intensity threshold are retained. Then, the coordinate distances are compared one by one in the retained candidate key point set. If the Euclidean distance between any two candidate key points is less than the minimum spacing, only the candidate key point with higher response intensity is retained.

[0092] S4.2 Classify and analyze the defect locations, extract morphological and texture features, and identify the defect types.

[0093] Specifically, image regions of each anomalous component are cropped from the correction image set based on time index and bounding box coordinates. These regions are then standardized to a fixed resolution, and analysis channels are selected. Morphological features of the cropped image regions are extracted using morphological operators, including both morphological and texture features. Morphological features are obtained through connected component analysis to acquire the area and perimeter of connected components, through bounding box measurement to acquire the aspect ratio, through minimum bounding rectangle fitting to acquire the area, through the convex hull algorithm to calculate the convex hull area, and from the area ratio to obtain the fullness and occupancy rate. Principal component analysis is then used to extract the principal components. The axis length and eccentricity are extracted using the Hu moment invariant moment method to extract seven invariant moment features. Texture features are calculated using the gray-level co-occurrence matrix at different distances and angles to determine contrast, correlation, energy, and uniformity. The LBP histogram is extracted using the local binary pattern method, and the pattern distribution is statistically analyzed. The image is decomposed using discrete wavelet transform, and the energy of each subband is calculated. The morphological and texture features are concatenated into a unified feature vector in field order, and the max-min normalization method is used to scale the features of each dimension to a unified range. The normalized unified feature vector is input into a classifier trained on a support vector machine to output the defect type.

[0094] S4.3 Quantitatively analyze the defect types, calculate the defect geometry and strength indices, and generate quantitative defect indices.

[0095] Specifically, the defect region is cropped within the defect location bounding box to generate a binary defect mask. Morphological features of the defect are extracted, including calculating the area through connected component analysis, extracting the contour to calculate the perimeter, fitting the aspect ratio using the minimum bounding rectangle, calculating the fullness based on the convex hull algorithm, obtaining the principal and secondary axis lengths using the least squares ellipse fitting, and applying a thinning algorithm combined with distance transformation to calculate the skeleton length and defect width. Simultaneously, intensity and texture features are extracted, including extracting the average gray-level difference between the defect region and the outer ring as contrast, calculating the gradient magnitude using the Sobel operator as edge intensity, and constructing a gray-level co-occurrence matrix to extract energy and correlation as texture indicators. The morphological features, defect width, and intensity texture indicators are combined to generate a defect quantification index.

[0096] S4.4 Perform differential calculation between the defect quantification index and the continuous time series images in the calibration image set to track the expansion trajectory of cracks and damage over time and generate the defect expansion path.

[0097] Specifically, continuous images are read from the calibration image set in chronological order, and the defect quantification index corresponding to each frame is extracted; the defect quantification index of the current frame is differiated from that of the previous frame to obtain the change values ​​of defect geometry and intensity indexes, and the defect contour increment map is updated; the defect center coordinate displacement and expansion direction are calculated based on the contour increment map to obtain the defect expansion vector; the defect expansion vector is used to perform connected component tracking of the defect region in continuous frames to form a preliminary defect path; the defect change value of each frame is combined with the preliminary path information to iteratively update the defect expansion trajectory and generate a complete defect expansion path.

[0098] S4.5. Based on the comprehensive defect location, defect type, defect quantification indicators, and defect propagation path, generate defect parameters.

[0099] Specifically, the defect location is read, and the corresponding category information is extracted from the defect type based on the defect location. At the same time, defect quantification indicators are obtained to describe the defect size, area, depth and severity. The defect location, defect type, defect quantification indicators and defect expansion path are spatially and temporally correlated, the changes of each defect during the expansion process are calculated, and integrated into a unified data structure. The center coordinates, bounding box, category probability, quantification indicator value and expansion trajectory of each defect are recorded to generate defect parameters.

[0100] It should be noted that, based on the generation of topological consistency information, defect location and quantitative analysis, combined with the continuous time frame difference method, are used to dynamically track the changes in defect quantitative indicators and generate defect parameters. This not only enables spatial identification and type classification of defects under static images, but also allows monitoring of the evolution process of defects using time-series images, thereby obtaining trend information such as crack propagation and damage accumulation over time. By introducing dynamic quantitative tracking in the time dimension, the defect detection results are upgraded from "qualitative identification" to "dynamic evolution analysis." This not only accurately locates and classifies defects, but also predicts development trends, thus providing a scientific basis for preventive maintenance and reliability assessment of aviation equipment, and enhancing the foresight and practicality of the detection.

[0101] S5. Conduct risk assessment on defect parameters, generate and execute repair plans, verify repair effectiveness through real-time image detection, and form a closed loop of detection and repair.

[0102] S5.1 Conduct a multi-dimensional risk assessment of the defect parameters, quantify the impact of defects on the safety and performance of aviation equipment, and generate a list of risk levels and maintenance priorities.

[0103] Specifically, the defect location, defect type, defect quantification index, and defect propagation path are read from the defect parameters. The weighted average method is used to calculate the index values ​​of the defect's impact on the structural safety, functional reliability, and performance of the aviation equipment, and these values ​​are normalized to obtain a comprehensive risk value. Based on a preset risk threshold, the comprehensive risk value is divided into different risk levels, such as high, medium, and low risk levels. According to the risk level and the importance of the defect location, a maintenance priority list is generated, recording the risk level, impact range, and recommended maintenance sequence for each defect, thus obtaining a risk level and maintenance priority list.

[0104] It should also be noted that the steps for setting the risk threshold include: selecting a suitable numerical range based on the safety and performance requirements of the aviation equipment and historical defect risk data, for example, setting the high-risk threshold to an example range of 0.7 to 1.0, the medium-risk threshold to an example range of 0.4 to 0.7, and the low-risk threshold to an example range of 0 to 0.4, to determine which risk level the comprehensive risk value of the defect parameter belongs to, and applying the preset risk threshold to classify the defect in the multidimensional risk assessment process.

[0105] S5.2 Based on the risk level and maintenance priority list, develop a maintenance sequence, required tools, operating steps, and expected repair results for each high-risk and priority maintenance component, and generate a maintenance plan.

[0106] Specifically, based on the risk level and maintenance priority list, information on high-risk and priority maintenance components is read sequentially to determine the maintenance order for each component; a list of required tools is compiled according to component type and maintenance requirements; specific operating procedures are written based on component structure and defect type, including disassembly, cleaning, repair, and reassembly processes; the expected repair effect indicators are determined by referring to maintenance specifications and historical repair data, such as the functional recovery rate or geometric accuracy of the repaired component; and the maintenance order, required tools, operating procedures, and expected repair effect are integrated to generate a maintenance plan.

[0107] S5.3 During the execution of the maintenance plan, real-time image detection is carried out using the calibration image set to monitor the condition of the maintenance components and generate real-time detection data.

[0108] Specifically, during the execution of the maintenance plan, consecutive image frames of the calibration image set are acquired sequentially. These image frames are then input into a pre-trained target detection network for feature extraction, generating enhanced feature maps. Based on these enhanced feature maps, a candidate box generation network generates candidate boxes for the maintenance components by sliding a window with a preset step size. The class score of each candidate box is normalized using the Softmax method to obtain the class probability distribution. Overlapping and redundant candidate boxes are removed using a non-maximum suppression algorithm based on a preset candidate box overlap threshold, resulting in the maintenance component detection results. The component location, category, and boundary information are extracted from the detection results, and the maintenance component status corresponding to each image frame is recorded to generate real-time detection data.

[0109] S5.4 Compare the real-time detection data with the expected repair effect in the repair plan, evaluate the repair completion and quality, and generate repair verification results.

[0110] Specifically, the bounding box position, category probability, and status information of each repair component in the real-time detection data are compared with the expected repair effect of the corresponding component in the repair plan. A similarity measurement method is used to calculate the repair completion rate of each component, such as quantification through position deviation, morphological matching degree, and category consistency score. At the same time, the repair completion rates of each repair component are weighted and summarized, and a weighted average method is used to calculate the overall repair quality index. The weights are determined according to the importance or risk level of the component. For example, the weight of critical load-bearing components is set to 0.7, and the weight of non-critical components is set to 0.3. The overall repair quality index reflecting the comprehensive completion status of all repair components is obtained by weighted averaging. The repair completion rates of each component and the overall quality index are compiled into the repair verification results.

[0111] S5.5 Adjust the subsequent maintenance plan based on the maintenance verification results and implement the corresponding repeated maintenance operations to form a closed loop of testing and maintenance.

[0112] Specifically, based on the maintenance verification results, the maintenance completion rate and overall maintenance quality index for each maintenance component are screened according to preset maintenance completion rate thresholds and maintenance quality thresholds. Components with a completion rate lower than the maintenance completion rate threshold or an overall maintenance quality index lower than the maintenance quality threshold are marked as substandard components. The maintenance order of substandard components is rearranged according to the maintenance priority list, the tools and operation steps required for each substandard component are determined, and targeted re-maintenance operations are carried out. At the same time, during the re-maintenance process, real-time detection is carried out using a calibration image set to obtain updated status data. The updated maintenance completion rate and quality index are compared again with the maintenance completion rate threshold and maintenance quality threshold, substandard components are screened, and feedback is given to the next round of maintenance arrangements, forming a closed loop of detection and maintenance.

[0113] It should also be noted that the steps for setting the maintenance completion threshold and maintenance quality threshold include: selecting a reasonable numerical range based on aviation equipment maintenance standards and safety performance requirements, for example, setting the maintenance completion threshold to an example range of 85% to 95% and the maintenance quality threshold to an example range of 0.8 to 0.95, to determine whether the component maintenance has met the expected standards, and applying the preset maintenance completion threshold and maintenance quality threshold to the screening process of maintenance verification results to identify non-compliant components.

[0114] This embodiment also provides a computer device applicable to the target detection-based aircraft equipment inspection and maintenance method, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the target detection-based aircraft equipment inspection and maintenance method proposed in the above embodiment.

[0115] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0116] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the target detection-based aircraft equipment inspection and maintenance method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0117] In summary, this invention constructs a component topology map based on component inspection results, further encodes node features and calculates edge weights to generate a quantifiable topology structure, and then compares the topology structure with a standard configuration. This enables modeling of the relative positions, assembly relationships, and functional associations between components, thereby not only identifying defects in individual components but also detecting anomalies such as assembly deviations, displacements, missing parts, or incorrect installations. Furthermore, it expands the inspection results from single-point identification to overall assembly consistency analysis, ultimately improving the accuracy and completeness of aerospace equipment inspection.

[0118] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for inspecting and maintaining aviation equipment based on target detection, characterized in that: include, Acquire initial multispectral images, and perform geometric distortion correction and illumination normalization to generate a corrected image set; The corrected image set is input into a pre-trained target detection network for multi-scale feature analysis and edge texture recognition to generate part detection results. Based on the component inspection results, the component topology relationships are established. Then, using graph structure analysis methods, the component topology relationships are compared with the standard configuration to generate topology consistency information. The specific steps are as follows. Extract the candidate box position and category probability of each component from the component detection results, calculate the center coordinates and relative distance of each component as spatial relationships, construct the components as component nodes, construct the spatial relationships and functional associations between components as edges, and establish a preliminary component topology graph; The node feature encoding and edge weight calculation are performed on the preliminary component topology map to quantify the spatial relationships and functional associations between the components and generate an encoded topology map. The encoded topology graph is compared with the standard configuration using graph matching. The degree of matching between nodes and edges is calculated using a similarity measurement algorithm to generate a topology consistency index. The abnormal relationships between components are analyzed using topology consistency indicators, and components with displacement, missing parts, and incorrect installations are marked to generate topology consistency information. Defect location and quantification analysis are performed on topology consistency information to extract defect location, defect type and defect quantification index. Then, through continuous time frame difference, the expansion trajectory of defect quantification index over time is tracked to generate defect parameters. Risk assessment of defect parameters is conducted, repair plans are generated and implemented, and the repair effect is verified through real-time image detection, forming a closed loop of detection and repair.

2. The aircraft equipment inspection and maintenance method based on target detection as described in claim 1, characterized in that: The specific steps for generating the corrected image set are as follows: Temporal and spatial synchronization corrections are performed on the initial multispectral image set to generate a synchronized image set; Perform surface geometric correction on the synchronized image set to generate a geometrically corrected image set; The geometrically corrected image set is subjected to illumination normalization, noise suppression, and edge texture enhancement to generate a clear feature image set; Multi-angle and multispectral information fusion is performed on a set of clear feature images to generate a corrected image set.

3. The aircraft equipment inspection and maintenance method based on target detection as described in claim 1, characterized in that: The pre-trained target detection network is obtained by iteratively training it through supervised learning based on a diverse image dataset containing various components and defects of aviation equipment, and by optimizing the weights of the target detection network using multi-scale feature extraction and edge texture recognition.

4. The aircraft equipment inspection and maintenance method based on target detection as described in claim 1, characterized in that: The specific steps for generating the component detection results are as follows: The corrected image set is input into a pre-trained target detection network to perform multi-scale feature extraction and generate a preliminary feature map. Edge texture enhancement processing is performed on the preliminary feature map to generate an enhanced feature map. The enhanced feature map is then used for position and category recognition to generate preliminary detection results. The non-maximum suppression algorithm is used to remove overlapping and redundant candidate boxes from the preliminary detection results and generate component detection results.

5. The aircraft equipment inspection and maintenance method based on target detection as described in claim 1, characterized in that: The standard configuration is obtained by collecting design drawings, 3D models and historical layout data of various components of aviation equipment, and establishing standard topological relationships between components based on structured analysis.

6. The aircraft equipment inspection and maintenance method based on target detection as described in claim 1, characterized in that: The specific steps for generating defect parameters are as follows: Based on topological consistency information, abnormal component regions are identified, the bounding box and key points of each abnormal component are extracted, and the defect location is generated. The defect locations are classified and analyzed to extract morphological and texture features and identify the defect types. Perform quantitative analysis on defect types, calculate defect geometry and strength indices, and generate quantitative defect indices; The defect quantification index is compared with the continuous time series images in the calibration image set to calculate the difference, track the expansion trajectory of cracks and damage over time, and generate the defect expansion path. Defect parameters are generated by combining the defect location, defect type, defect quantification indicators, and defect propagation path.

7. The aircraft equipment inspection and maintenance method based on target detection as described in claim 1, characterized in that: The specific steps for forming a closed loop for testing and maintenance are as follows. A multidimensional risk assessment of defect parameters is conducted to quantify the impact of defects on the safety and performance of aviation equipment, and a list of risk levels and maintenance priorities is generated. Based on the risk level and maintenance priority list, a maintenance plan is generated by defining the maintenance sequence, required tools, operating steps, and expected repair results for each high-risk and priority maintenance component. During the implementation of the maintenance plan, real-time image detection is carried out using the calibration image set to monitor the condition of the maintenance components and generate real-time detection data. The real-time detection data is compared with the expected repair results in the repair plan to evaluate the repair completion and quality, and generate repair verification results. Adjust subsequent maintenance plans based on maintenance verification results and implement corresponding repeat maintenance operations to form a closed loop of testing and maintenance.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the target detection-based aircraft equipment inspection and maintenance method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the target detection-based aviation equipment inspection and maintenance method according to any one of claims 1 to 7.

Images