Intelligent detection method for misloading and neglected loading of fasteners of airplane power distribution system under complex illumination

Abstract

The invention discloses a method for intelligently detecting mistaken and neglected installation of fasteners of an airplane power distribution system under complex illumination, and the method comprises the steps: installing a tested piece on a precise rotary table, and forming an automatic image collection platform through an industrial camera and an annular LED light source which are disposed on an automatic guide rail; and then a high-quality fusion reconstruction image is obtained by using a multi-exposure fusion technology, and misloading and neglected loading positions of the fasteners are automatically detected through a detection model with a deep learning algorithm, so that comprehensive, efficient and accurate detection of the positions of all the fasteners on the surface of the power distribution unit is realized, and the safety and reliability of an aircraft power distribution system are ensured.

Description

Technical Field

[0001] The present invention belongs to the technical field of surface assembly defect detection of aircraft power distribution equipment. More specifically, it relates to an intelligent detection method for misinstallation and missing installation of fasteners in an aircraft power distribution system under complex lighting conditions. Background Art

[0002] The safe operation of an aircraft is closely related to the assembly quality of its key components. As the core component of the aircraft electrical system, the power distribution system fixes and connects various electronic components together through numerous fasteners (such as screws, nuts, etc.) to ensure the stable and reliable operation of the system. If misinstallation (wrong installation) or missing installation (omission of installation) of fasteners occurs during the assembly process, it may lead to loose connections, increased vibration, and even potential safety hazards such as electrical failures. Therefore, after the assembly of the aircraft power distribution system is completed, it is very important to conduct a comprehensive inspection of all surface fasteners to ensure no misinstallation or missing installation.

[0003] Traditionally, the detection of such fasteners mainly relies on manual visual inspection, but there are many deficiencies in manual inspection: First, manual detection is subject to the experience and attention of inspectors, and it is easy to cause missed detection or misjudgment due to fatigue or subjective factors; Second, manually checking a large number of fasteners one by one is inefficient and difficult to meet the requirements of high efficiency and high precision in modern aviation manufacturing; Finally, under complex lighting conditions (such as reflection on the surface of fasteners, coexistence of high light and shadow), it is difficult for the human eye to clearly see all areas at the same time, resulting in the inability to discover some detailed problems.

[0004] In the prior art, there are also some detection methods based on machine vision. These methods mainly use traditional image processing techniques or deep learning models to analyze the assembly situation of fasteners. However, most current detection methods related to deep learning focus on the processing of single images, emphasizing the detection of fasteners in local areas or from specific perspectives, and the focus is on improving the precision rate of target detection in a single image, while ignoring the higher requirement for overall coverage in the detection of surface fasteners in the aircraft power distribution system. The detection of a single photo often fails to achieve complete coverage of all surface fasteners in the power distribution system, easily leading to missed detection of fasteners. Summary of the Invention

[0005] The purpose of the present invention is to overcome the deficiencies of the prior art and provide an intelligent detection method for misinstallation and missing installation of fasteners in an aircraft power distribution system under complex lighting conditions. Based on an intelligent detection method combining multi-angle, multi-exposure fusion and deep learning, it realizes comprehensive and accurate detection of all fasteners on the entire surface of the aircraft power distribution system, so as to solve the problems of incomplete coverage, insufficient detection accuracy and low efficiency existing in the existing detection methods, thereby achieving efficient, precise and comprehensive detection of surface fasteners of the aircraft power distribution unit and ensuring the flight safety of the aircraft.

[0006] To achieve the above-mentioned invention object, an intelligent detection method for misassembly and missing installation of fasteners in an aircraft power distribution system under complex lighting conditions according to the present invention is characterized by including:

[0007] (1) Commission the detection device. After the detection device is commissioned, initialize the imaging parameters of the industrial camera.

[0008] (2) Set the shooting distance range, and collect sample images and test images through the detection device.

[0009] (3) Add target boxes to the targets in each sample image, and mark the category and position coordinates of each target.

[0010] (4) Perform multi-exposure image fusion reconstruction to obtain a fused and reconstructed image.

[0011] (5) Construct an improved Faster R-CNN target detection model.

[0012] (6) Train the improved Faster R-CNN target detection model.

[0013] (7) Detect misassembly and missing installation defects of fasteners on the surface of aircraft power distribution equipment.

[0014] The invention object of the present invention is achieved as follows:

[0015] An intelligent detection method for misassembly and missing installation of fasteners in an aircraft power distribution system under complex lighting conditions according to the present invention installs the test piece on a precision rotary turntable, forms an automated image acquisition platform through an industrial camera and a ring-shaped LED light source installed on an automatic guide rail, then uses multi-exposure fusion technology to obtain a high-quality fused and reconstructed image, and automatically detects the misassembly and missing installation positions of fasteners through a detection model with a deep learning algorithm, realizing comprehensive, efficient and accurate detection of the positions of all fasteners on the surface of the power distribution unit, and ensuring the safety and reliability of the aircraft power distribution system.

[0016] Meanwhile, an intelligent detection method for misassembly and missing installation of fasteners in an aircraft power distribution system under complex lighting conditions according to the present invention also has the following beneficial effects:

[0017] (1) By using multi-exposure image fusion technology, it effectively solves the problem of detail loss in traditional single-exposure images due to complex lighting conditions. Especially for fasteners made of metal materials, severe specular reflections are likely to occur on the surface under strong light, resulting in blurred edge contours of the fasteners; while under low light conditions, the hole areas of the uninstalled fasteners are too dark and lack details, making it difficult to identify missing fasteners. The multi-exposure image fusion technology reconstructs a high-quality image that retains both bright and dark details, thus effectively solving the technical problem of difficult to balance details under complex lighting conditions.

[0018] (2) A technical solution based on the Faster R-CNN object detection model is proposed. The ResNet50 convolutional neural network is used as the backbone network for feature extraction. Aiming at the problems in fastener detection, such as diverse shapes, large size variations, and complex background interference, an attention mechanism is introduced at appropriate positions in the object detection model, including the SE channel attention module, the Transformer global self-attention module, and the CBAM channel-spatial joint attention module, to construct an improved Faster R-CNN architecture with multiple attention fusions. This effectively enhances the model's ability to focus on key feature information, significantly improves the richness and representativeness of feature expressions, and enables the model to accurately detect and classify the positions and assembly states of various fasteners in complex background environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 FIG. is a flowchart of an intelligent detection method for misinstallation and missing installation of fasteners in an aircraft power distribution system under complex lighting;

[0020] Figure 2 FIG. is a schematic diagram of assembly states of different categories;

[0021] Figure 3 FIG. is an image after fusion and reconstruction of test images under different exposure intensities;

[0022] Figure 4 FIG. is an architecture diagram of the improved Faster R-CNN object detection model;

[0023] Figure 5 FIG. is an architecture diagram of the traditional Faster R-CNN object detection model;

[0024] Figure 6 FIG. is a schematic diagram of the inspection results of fasteners on the surface of aircraft power distribution equipment. DETAILED DESCRIPTION OF THE INVENTION

[0025] The following describes the specific embodiments of the present invention with reference to the drawings, so that those skilled in the art can better understand the present invention. It should be particularly noted that in the following description, when the detailed description of known functions and designs may obscure the main content of the present invention, these descriptions will be omitted here.

[0026] Embodiment

[0027] For convenient description, the relevant technical terms appearing in the specific embodiments are first explained:

[0028] Faster R-CNN (Faster Region-based Convolutional Neural Network): Fast Region-based Convolutional Neural Network;

[0029] ResNet50 (Residual Network 50): Residual Network 50

[0030] SE (Squeeze-and-Excitation): Squeeze-and-Excitation module

[0031] CBAM (Convolutional Block Attention Module): Convolutional Block Attention Module

[0032] Transformer: Self-attention neural network

[0033] LayerNorm (Layer Normalization): Layer Normalization

[0034] ReLu (Rectified Linear Unit): Rectified Linear Unit

[0035] RPN (Region Proposal Network): Region Proposal Network

[0036] ROI (Region of Interest): Region of Interest, referring to a specific area in an image that may contain the target object.

[0037] In this embodiment, as Figure 1 shown, an intelligent detection method for misassembly and missing parts of fasteners in an aircraft power distribution system under complex illumination according to the present invention includes:

[0038] S1. Debugging the detection device;

[0039] Taking the aircraft power distribution equipment to be tested as the test piece, fixing the test piece at the center of the rotating table, and making the axis of the rotating table coincide with the geometric center axis of the test piece; fixing the industrial camera on the automatic guide rail device, and the automatic guide rail device can achieve precise two-dimensional movement in the X-axis direction and the Y-axis direction;

[0040] Arranging a ring-shaped LED light source with high uniformity around the test piece, adjusting the position, height and incident angle of the LED ring light source so that the ring-shaped LED light source can stably and uniformly illuminate the test piece; adjusting the position of the industrial camera so that the optical axis of its lens is precisely aligned with the test piece, so that the test piece is located within the effective imaging area of the industrial camera;

[0041] After the device to be detected is debugged, initialize the imaging parameters of the industrial camera, specifically including the aperture, focusing position, and exposure time of the industrial camera; among them, by adjusting the lens aperture, the depth of field range and image brightness can be optimized to meet the best combination of clarity and brightness; by finely adjusting the focusing ring, the camera can clearly and accurately focus on the surface to be detected of the power distribution unit; according to the on-site lighting conditions, the surface material of the object, and the reflection characteristics, reasonably set the exposure time to ensure that the acquired image has appropriate brightness and contrast;

[0042] S2. Image data acquisition;

[0043] S2.1. Set the shooting distance range;

[0044] Let the coincidence position of the axis of the turntable and the geometric center axis of the test piece be O; with the coincidence point O as the center, the distance range for the automatic guide device to move on the X-axis and Y-axis is [D min , D max , in this embodiment, [D min , D max takes [0 cm, 40 cm].

[0045] S2.2. Collect sample images;

[0046] Set the rotation angle step Δθ of the turntable. When the turntable rotates one week, at any rotation angle θ i below, with the coincidence point O as the starting point, first slide left and right on the X-axis, with [D min , D max as the sliding limit, set the collection points D j at the collection interval step Δd, and at each collection point D j take low-exposure sample images medium-exposure sample images high-exposure sample images After the left and right sliding on the X-axis is completed, then slide up and down on the Y-axis to take pictures, and then increase the rotation angle θ i by Δθ, and repeat the above operations until the turntable rotates one week. Among them, record the sample image taken at the jth collection point D i at the ith rotation angle θ j as In this embodiment, Δθ takes 15°, and Δd takes 5 cm.

[0047] S2.3. Collect test images;

[0048] Increase the rotation angle step of the turntable to where λ is a positive integer, and then take test images according to step (2.2). Among them, record the jth collection point D i at the ith rotation angle θj The captured test images are In this embodiment, λ is taken as 3, is 45°

[0049] S3. Manual annotation of sample images;

[0050] Add a target box for the target in each sample image, and label the category and position coordinates of each target;

[0051] In this embodiment, each target includes 7 categories, and the 7 category numbers are "1, 2, 3, 4, 5, 6, 7" in sequence. The assembly states corresponding to the 7 categories are shown in Table 1 and Figure 2 shown as follows;

[0052] Category Number Assembly Status Installation Status 1 Correct Countersunk screw installed in counterbore 2 Correct Non-countersunk screw installed in through-hole 3 Correct Non-countersunk screw with flat washer installed in through-hole 4 Correct Non-countersunk screw with spring washer and flat washer installed in through-hole 5 Incorrect Countersunk screw installed in through-hole or non-countersunk screw installed in counterbore 6 Incorrect Countersunk screw incorrectly equipped with spring washer or flat washer 7 Missing Screw not installed in counterbore or through-hole

[0053] Table 1

[0054] S4. Multi-exposure image fusion and reconstruction;

[0055] S4.1. Traverse each group of test images Calculate the contrast weight local variance weight and brightness weight of each pixel point (x, y) in each test image under different exposure conditions

[0056]

[0057] Among them, is the Laplacian operator; represents the gray value of the pixel point (x, y) in, and Var(x, y) is the gray statistical variance within the m×m window centered on the pixel (x, y); represents the mean of the squares of the pixels within the m×m window centered on the pixel point (x, y), represents the square of the mean of the pixels within the m×m window centered on the pixel point (x, y); ε is a constant; is the average gray value of the test image under the k-th exposure condition; is the ideal brightness value under the k-th exposure condition, where k = 1, 2, 3, representing low exposure, medium exposure, and high exposure respectively; σ is the sensitivity to control the brightness weight; in this embodiment, takes the gray median value of 128; the value of σ is taken as 0.2; the size of the m×m window can be determined according to the actual image size, and in this embodiment, it is taken as 5×5; ε is a very small positive number, and in this embodiment, it is taken as 10 -12 .

[0058] S4.2. Calculate the fusion weights of each pixel under different exposure conditions;

[0059]

[0060] Among them, α, β, and γ are the fusion coefficients of the contrast weight, local variance weight, and brightness weight respectively, and the general value range is [0.5, 2.0]. In this embodiment, α, β, and γ are taken as 0.95, 0.99, and 0.97 respectively;

[0061] S4.3. To ensure that the sum of weights is 1, normalize the fusion weights of each pixel;

[0062]

[0063] S4.4. Arrange the normalized fusion weights of each pixel into a fusion weight map according to the positions of the pixels;

[0064] Then perform two-dimensional Gaussian filtering and downsampling on the fusion weight maps under different exposure conditions to obtain the weight Gaussian pyramid under different exposure conditions; in this embodiment, the number of layers of the weight Gaussian pyramid is 5.

[0065] S4.5. Perform two-dimensional Gaussian filtering and downsampling on the test images under different exposure conditions to obtain the image Gaussian pyramid under different exposure conditions;

[0066] Then subtract the adjacent two-layer Gaussian images in the image Gaussian pyramid to construct a Laplacian pyramid, where the top layer is directly stored; in this embodiment, the number of layers of the image Gaussian pyramid is 5.

[0067] S4.6. Use the weight Gaussian pyramid to perform layer-by-layer and pixel-by-pixel fusion on the Laplacian pyramid to obtain the Laplacian image after fusion for each layer;

[0068]

[0069] Among them, represents the pixel value after fusion of the pixel point (x, y) in the l-th layer under the k-th exposure condition; represents the pixel value of the pixel point (x, y) in the l-th layer under the k-th exposure condition, represents the weight of the pixel point (x, y) in the l-th layer under the k-th exposure condition;

[0070] Form the Laplacian image of the l-th layer by the pixel values after fusion of each pixel point in the l-th layer under the k-th exposure condition;

[0071] S4.7. Reconstruct the test images under different exposure conditions;

[0072] Upsample the Laplacian image of the first layer, add the upsampling result to the Laplacian image of the second layer, then upsample the added result, and then add the upsampling result to the Laplacian image of the third layer, and so on until the last Laplacian image. Finally, take the accumulated result as the final fused and reconstructed image, denoted as

[0073] In this embodiment, on the left are test images with three exposure intensities. After performing fused reconstruction through the above method, the fused reconstruction image shown on the right is obtained.

[0074] S5. Construct an improved Faster R-CNN object detection model;

[0075] In this embodiment, as Figure 4 shown, based on the Faster R-CNN as the basic model, insert an SE attention module between the ResNet50 network and the RPN network, and then sequentially connect a Transformer encoder module and a CBAM attention module at the output end of the RPN network. The output end of the CBAM attention module is connected to the classifier, thereby constructing an improved Faster R-CNN model;

[0076] Among them, the SE attention module includes a global average pooling layer and two fully connected layers;

[0077] The Transformer encoder module includes a multi-head self-attention mechanism, a residual connection, LayerNorm normalization, and two feed-forward fully connected networks;

[0078] The CBAM attention module includes a channel attention sub-module and a spatial attention sub-module. After the two act on the ROI feature map in series, they are then flattened and followed by full connection classification;

[0079] In this embodiment, the traditional Faster R-CNN model is as Figure 5 shown. By adding an SE attention module, the global channel weights are adjusted to feature responses, effectively improving the semantic sensitivity and discriminative ability of the features in the RPN proposal stage; by the Transformer module, the dependencies and context semantic information between the internal positions of the ROI are effectively captured, strengthening the sufficient interaction and fusion within the ROI features; by the CBAM attention module, valuable feature channels are simultaneously screened and key spatial regions are highlighted, thus significantly optimizing the quality of the input features of the second-stage classifier and ultimately improving the discriminative accuracy of the target categories.

[0080] S6. Train the improved Faster R-CNN object detection model;

[0081] S6.1. Linearly normalize the labeled sample images, thus linearly scaling the original grayscale range [0, 255] to the model input range [0, 1], and then divide them into several batches, and input each batch into the improved Faster R-CNN object detection model in turn;

[0082] S6.2. The linearly normalized sample images first extract feature maps through the ResNet50 network, and then input the feature maps into the SE attention module. After compression, excitation, and recalibration, the feature maps with enhanced channel attention are obtained;

[0083] S6.3. Input the feature maps with enhanced channel attention into the RPN network, and generate a series of bounding boxes and classification probabilities for each fastener through the RPN network;

[0084] Sort the bounding boxes from largest to smallest according to the classification probabilities, and perform non-maximum suppression processing. Select the top N bounding boxes with the highest probabilities as the candidate regions ROI;

[0085] S6.4. Project and crop the feature maps with enhanced channel attention according to the candidate region positions to obtain the RoI feature maps; then expand the RoI feature maps into two-dimensional sequences and send them into the Transformer encoder. Calculate the self-attention features through the Transformer encoder, and then reshape them back into the RoI feature maps according to the self-attention features;

[0086] S6.5. Input the reshaped RoI feature maps into the CBAM attention module to perform channel attention processing and spatial attention processing in turn, and obtain the RoI feature maps with dual-dimensional joint enhancement;

[0087] S6.6. Input the RoI feature maps with dual-dimensional joint enhancement into the classifier, predict the class probabilities and bounding boxes of each fastener, match the predicted bounding boxes with the truly labeled target boxes, determine the truly labeled target boxes corresponding to each predicted box, and then use the cross-entropy loss function to calculate the classification loss value, use the smooth L1 loss function to calculate the bounding box regression loss value, perform weighted summation on the classification loss and the position regression loss to obtain the total loss value of the current batch of training samples. Finally, take the total loss value as the optimization target, use the gradient descent method for backpropagation to update the network model parameters, and then repeat the above process by inputting the next batch of training images until the improved Faster R-CNN object detection model converges;

[0088] S7. Detect the missing and misassembled defects of the fasteners on the surface of the aircraft electrical distribution equipment;

[0089] S7.1. Linearly normalize the test image and then input it into the trained Faster R-CNN object detection model to obtain each fastener category and bounding box in the test image;

[0090] S7.2. Determine the misassembly and missing defects based on the predicted category of each fastener;

[0091] When the predicted category is "1", it means that the installation state of the fastener is correct, and the countersunk head screw is installed in the counterbore;

[0092] When the predicted category is "2", it means that the installation state of the fastener is correct, and the non-countersunk head screw is installed in the through hole;

[0093] When the predicted category is "3", it means that the installation state of the fastener is correct, and the non-countersunk head screw equipped with a flat washer is installed in the through hole;

[0094] When the predicted category is "4", it means that the installation state of the fastener is correct, and the non-countersunk head screw equipped with a spring washer and a flat washer is installed in the through hole;

[0095] When the predicted category is "5", it means that the installation state of the fastener is incorrect, and the countersunk head screw is installed in the through hole or the non-countersunk head screw is installed in the counterbore;

[0096] When the predicted category is "6", it means that the installation state of the fastener is incorrect, and the countersunk head screw is incorrectly equipped with a spring washer or a flat washer;

[0097] When the predicted category is "7", it means that the fastener is missing, and no screw is installed in the counterbore or through hole.

[0098] Experimental verification

[0099] In this example, the experimental model is built and trained based on the TensorFlow framework. The hardware configuration is NVIDIA GeForce GTX 1650, the operating system is Windows 10, Python 3.6 and TensorFlow-GPU 1.5.0 are used, CUDA 9.0 is supported, and TensorFlow 2.3.0 and OpenCV 4.5.0 are adopted.

[0100] Experimental verification is carried out on the surface of the actual aircraft power distribution system. The test data set used includes a total of 2370 images with a resolution of 960×960, covering all seven types of fastener installation states.

[0101] For the experimental test results, the Intersection over Union (IoU) and Dice similarity coefficient are used to comprehensively evaluate the experimental test results. The IoU refers to the ratio of the intersection area to the union area of the predicted region and the ground truth region; the Dice similarity coefficient is defined as the ratio of twice the overlapping area of the predicted region and the ground truth region to the sum of their areas. The value ranges of both of these two metrics are from 0 to 1, and the larger the value, the closer the predicted result is to the actual situation.

[0102] The average IoU of the detection status of various fasteners reaches 98.41%, and the average Dice similarity coefficient reaches 98.70%, which proves that the method of the present invention can provide comprehensive and accurate fastener status detection results in the actual environment;

[0103] Figure 6 Some detection results are shown, where the solid bounding boxes represent correctly assembled fasteners, the long dashed bounding boxes represent incorrectly assembled fasteners, the dotted bounding boxes represent missing fasteners, and the corresponding category information is automatically marked in the upper left corner of the bounding boxes.

[0104] Although the above describes the illustrative specific embodiments of the present invention for the convenience of those skilled in the art to understand the present invention, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions made using the concept of the present invention are within the scope of protection.

Claims

1. An intelligent detection method for misinstallation and missing installation of fasteners in an aircraft power distribution system under complex lighting conditions, characterized in that, Including: (1) Detection device debugging; Use the aircraft power distribution equipment to be tested as the test piece, fix the test piece at the center of the rotating table, and make the axis of the rotating table coincide with the geometric center axis of the test piece; fix the industrial camera on the automatic guide rail device, and the automatic guide rail device can achieve precise two-dimensional movement in the X-axis direction and the Y-axis direction; Arrange a high-uniformity annular LED light source around the test piece, adjust the position, height and incident angle of the LED annular light source so that the annular LED light source can stably and uniformly irradiate the test piece; adjust the position of the industrial camera so that the optical axis of its lens is precisely aligned with the test piece, so that the test piece is located within the effective imaging area of the industrial camera; After the detection device debugging is completed, initialize the imaging parameters of the industrial camera; (2) Image data acquisition; (2.1) Set the shooting distance range; Let the coincidence position of the axis of the rotating table and the geometric center axis of the test piece be O; with the coincidence point O as the center, the distance range for the automatic guide device to move along the X-axis and Y-axis is [D min , D max ; (2.2) Collect sample images; Set the rotation angle step Δθ of the rotary table. When the rotary table rotates one week, at any rotation angle θ i , starting from the coincidence point O, first slide left and right on the X-axis, with [D min , D max as the sliding limit, set the acquisition point D according to the acquisition interval step Δd j , and at each acquisition point D j , capture a low-exposure sample image a medium-exposure sample image a high-exposure sample image . After the left and right sliding on the X-axis is completed, then slide up and down on the Y-axis to capture images. Then increase the rotation angle θ i by Δθ, and repeat the above operations until the rotary table rotates one week. Among them, denote the sample image captured at the j-th acquisition point D i at the i-th rotation angle θ j as (2.3) Collect test images; Increase the rotation angle step of the rotary table to where λ is a positive integer, and then capture test images according to step (2.2). Among them, denote the j-th acquisition point D at the i-th rotation angle θ i as j The captured test image is (3) Manual annotation of sample images; Add a target box to the target in each sample image, and mark the category and position coordinates of each target; (4) Multi-exposure image fusion and reconstruction to obtain a fused and reconstructed image (5) Construct an improved Faster R-CNN object detection model; Based on Faster R-CNN as the basic model, insert an SE attention module between the ResNet50 network and the RPN network, and then sequentially connect a Transformer encoder module and a CBAM attention module at the output end of the RPN network. The output end of the CBAM attention module is connected to the classifier, thus constructing an improved Faster R-CNN model; (6) Train the improved Faster R-CNN object detection model; (6.1) Perform linear normalization processing on the annotated sample images, and then divide them into several batches, and input each batch into the improved Faster R-CNN object detection model in turn; (6.2) The linearly normalized sample images first extract feature maps through the ResNet50 network, and then input the feature maps into the SE attention module. After compression, excitation and recalibration, the feature maps with enhanced channel attention are obtained; (6.3) Input the feature maps with enhanced channel attention into the RPN network, and generate a series of bounding boxes and classification probabilities for each fastener through the RPN network; Sort the bounding boxes according to the classification probability from large to small, perform non-maximum suppression processing, and select the top N bounding boxes with the highest probability as the candidate region ROI; (6.4) Project and crop the feature maps with enhanced channel attention according to the candidate region positions to obtain RoI feature maps; then expand the RoI feature maps into two-dimensional sequences and send them into the Transformer encoder. Calculate the self-attention features through the Transformer encoder, and then reshape them back into RoI feature maps according to the self-attention features; (6.5) Pass the reshaped RoI feature maps through the CBAM attention module to perform channel attention processing and spatial attention processing in turn, and obtain RoI feature maps with double-dimensional joint reinforcement; (6.6) Input the RoI feature map with two-dimensional joint enhancement into the classifier to predict the class probability and bounding box of each fastener. Match the predicted bounding box with the true labeled target box to determine the true labeled target box corresponding to each predicted box. Then, use the cross-entropy loss function to calculate the classification loss value, use the smooth L1 loss function to calculate the bounding box regression loss value, and perform weighted summation of the classification loss and the position regression loss to obtain the total loss value of the current batch of training samples. Finally, taking the total loss value as the optimization target, use the gradient descent method to backpropagate and update the network model parameters, and then repeat the process by inputting the next batch of training images until the improved Faster R-CNN object detection model converges; (7) Detect the missing or misassembled defects of the fasteners on the surface of the aircraft electrical distribution equipment; (7.1) After linearly normalizing the test image, input it into the trained Faster R-CNN object detection model to obtain the category and bounding box of each fastener in the test image; (7.2) Determine the missing or misassembled defects according to the predicted category of each fastener; When the predicted category is "1", it means that the installation state of the fastener is correct, and the countersunk head screw is installed in the counterbore; When the predicted category is "2", it means that the installation state of the fastener is correct, and the non-countersunk head screw is installed in the through hole; When the predicted category is "3", it means that the installation state of the fastener is correct, and the non-countersunk head screw equipped with a flat washer is installed in the through hole; When the predicted category is "4", it means that the installation state of the fastener is correct, and the non-countersunk head screw equipped with a spring washer and a flat washer is installed in the through hole; When the predicted category is "5", it means that the installation state of the fastener is incorrect, and the countersunk head screw is installed in the through hole or the non-countersunk head screw is installed in the counterbore; When the predicted category is "6", it means that the installation state of the fastener is incorrect, and the countersunk head screw is incorrectly equipped with a spring washer or a flat washer; When the predicted category is "7", it means that there is a missing installation of the fastener, and no screw is installed in the counterbore or the through hole.

2. The method for detecting the defect of missing or misassembled surface fasteners of an aircraft power distribution device according to claim 1, characterized in that, The initialization settings of the imaging parameters of the industrial camera include: the aperture, focusing position, and exposure time of the industrial camera.

3. The intelligent detection method for misassembly and missing installation of fasteners in an aircraft power distribution system under complex illumination according to claim 1, wherein, The initialization settings of the imaging parameters of the industrial camera include: the aperture, focusing position, and exposure time of the industrial camera.

4. An intelligent detection method for misinstallation and missing installation of fasteners in an aircraft power distribution system under complex lighting conditions according to claim 1, characterized in that, The method for multi-exposure image fusion and reconstruction is as follows: s1. Traverse each group of test images Calculate the contrast weight of each pixel point (x, y) in each test image under different exposure conditions Local variance weight Among them, is the Laplacian operator; represents the gray value of the pixel point (x, y) in the mean of the squares of the pixels within the m×m window centered on the pixel (x, y), represents the square of the mean of the pixels within the m×m window centered on the pixel point (x, y); ε is a constant; is the average gray value of the test image under the k-th exposure condition; is the ideal brightness value under the k-th exposure condition, where k = 1, 2, 3, representing low exposure, medium exposure, and high exposure respectively; σ is the sensitivity for controlling the brightness weight; s2. Calculate the fusion weights of each pixel point under different exposure conditions; Among them, α, β, and γ are the fusion coefficients of the contrast weight, local variance weight, and brightness weight respectively; s3. Normalize the fusion weights of each pixel point; s4. The fused weights after normalizing each pixel point Arrange them into a fused weight map according to the positions of the pixel points; Then, perform two-dimensional Gaussian filtering and downsampling on the fusion weight maps under different exposure conditions to obtain the weight Gaussian pyramids under different exposure conditions; s5. Test images under different exposure conditions Perform two-dimensional Gaussian filtering and downsampling to obtain the image Gaussian pyramid under different exposure conditions; Subtract the adjacent two-layer Gaussian images in the image Gaussian pyramid to construct the Laplacian pyramid, and directly store the top layer; s6. Use the weight Gaussian pyramid to perform layer-by-layer and pixel-by-pixel fusion on the Laplacian pyramid to obtain the fused Laplacian image of each layer; Among them, represents the pixel value after fusion of the pixel point (x, y) on the l-th layer under the k-th exposure condition; represents the pixel value of the pixel point (x, y) on the l-th layer under the k-th exposure condition, represents the weight of the pixel point (x, y) on the l-th layer under the k-th exposure condition; The pixel values of each pixel point after fusion in the l-th layer under the k-th exposure condition form the l-th layer Laplacian image; s7. Reconstruct the test images under different exposure conditions; Upsample the Laplacian image of the first layer, add the upsampling result to the Laplacian image of the second layer, then upsample the added result, and then add the upsampling result to the Laplacian image of the third layer, and so on until the Laplacian image of the last layer. Finally, take the accumulated result as the final fused reconstruction image, denoted as

Images