Intelligent analysis method for realizing body health management of unmanned aerial vehicle

Abstract

The invention relates to an intelligent analysis method for realizing health management of an unmanned aerial vehicle body, and the method comprises the steps: carrying out the physical modeling and stress analysis of the unmanned aerial vehicle body, and determining key detection parts and corresponding damage types of the unmanned aerial vehicle body; collecting image data of key detection parts of an entity unmanned aerial vehicle body and preprocessing the image data to construct an image data set; an unmanned aerial vehicle structure damage detection network model is constructed, the network model comprises a preprocessing module, a backbone network module, a neck network module and a head network module, and the network model is trained; constructing a to-be-detected unmanned aerial vehicle image set, and performing unmanned aerial vehicle body structure damage detection on the to-be-detected unmanned aerial vehicle image set by using the trained unmanned aerial vehicle structure damage detection network model to obtain a detection result; and analyzing the detection result based on an empirical empowerment-based airframe health evaluation algorithm, and performing comprehensive health state evaluation on the to-be-detected unmanned aerial vehicle airframe. The method realizes intelligent detection and evaluation of potential damage of the unmanned aerial vehicle body.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent monitoring and fault diagnosis, and particularly to an intelligent analysis method for realizing the health management of an unmanned aerial vehicle (UAV) airframe. Background Art

[0002] Logistics transportation is an important application segment in the low-altitude economy, requiring the flight platform to have the capabilities of long endurance, large payload, and high-speed cruise. Small and medium-sized fixed-wing UAVs play an important role currently. With the deepening of the reform of low-altitude airspace management and the surging demand for end-delivery, the production demand for such UAVs shows a continuous growth trend. The production process of small and medium-sized fixed-wing UAVs is strict, the system integration is difficult, and the factory inspection procedures are complex, resulting in their manufacturing costs increasing exponentially compared with those of quadrotor UAVs. This prompts users to position small and medium-sized fixed-wing UAVs as high-value assets that need to be recycled. However, existing small and medium-sized fixed-wing UAVs mostly adopt hard landing methods such as parachute buffering, which are extremely likely to cause damage to the surface composite materials and even structural damage to the airframe. Therefore, frequent maintenance and support are required to extend the service life. However, whether in the factory inspection or maintenance and support stage, existing damage detection technologies cannot accurately detect the flaws of the airframe of small and medium-sized fixed-wing UAVs. The main reasons include: the external damage scales such as skin and cracks vary greatly, while the internal damages such as beam fractures and rib fractures are complex and difficult to directly observe. Therefore, in view of the current development needs of the low-altitude economy, it is very necessary to design a detection method with low cost, high accuracy, high efficiency, and the ability to comprehensively evaluate the internal and external damages of the airframe. Summary of the Invention

[0003] In view of the above problems, the present invention proposes an intelligent analysis method for realizing the health management of an unmanned aerial vehicle airframe. By physically modeling and analyzing the forces of the UAV and conducting random landing experiments, the damage characteristics of the UAV airframe are obtained. Visual sensing technology and ray detection technology are used to collect damage images, and combined with deep learning technology, intelligent detection of the internal and external damages of the UAV airframe is realized. According to the detection results and the airframe health assessment algorithm based on empirical weighting, the health status of the UAV airframe is evaluated.

[0004] An intelligent analysis method for realizing the health management of an unmanned aerial vehicle airframe proposed by the present invention includes the following steps: Step S1, physically model the UAV airframe to obtain the model of the UAV airframe, and analyze the forces on the model of the UAV airframe through random landing experiments to determine the key detection parts of the UAV airframe and the corresponding damage types; Step S2, use a small field-of-view camera and an X-ray machine to collect image data of the key detection parts of multiple physical UAV airframes, collect image data with different viewing distances and different ray intensities, and preprocess the collected image data to construct an image dataset; Step S3: Construct a drone structural damage detection network model based on the Hyper-YOLO network. The drone structural damage detection network model includes a preprocessing module, a backbone network module, a neck network module, and a head network module. The preprocessing module is used to process the images in the image dataset to obtain normalized tensors. The backbone network module is used for feature extraction. The neck network module is used to achieve multi-scale feature fusion. The head network module is used to classify and regress the results of feature extraction and feature fusion, and output the target detection results including the damage location, damage type, damage quantity, and damage degree value of the drone airframe structure. Step S4: Divide the image dataset obtained in Step S2 into a training set and a validation set. Based on the image characteristics of the image dataset, adjust the parameters of the drone structural damage detection network model constructed in Step S3, and use the training set to train the network model to obtain a trained drone structural damage detection network model. Step S5: Collect images of the airframe of the drone to be tested to form an image set of the drone to be tested. Use the trained drone structural damage detection network model in Step S4 to detect the damage of the drone airframe structure in the image set of the drone to be tested to obtain the detection results. Step S6: Based on the body health assessment algorithm with empirical weighting, analyze the damage information recorded in the detection results and comprehensively evaluate the health status of the airframe of the drone to be tested.

[0005] Furthermore, Step S1 specifically includes the following steps: Step S11: Obtain the material parameters of the drone airframe and the composition of the drone airframe structure. Based on the obtained material parameters of the drone airframe and the composition of the drone airframe structure, perform physical modeling on the drone airframe to obtain the model of the drone airframe. Step S12: Divide the model of the drone airframe into two parts, the external structure and the internal structure, and each part is divided into multiple main component units. Step S13: Set the landing simulation parameters for the random landing experiment. The landing simulation parameters include the landing height and the landing tilt angle, and perform N random landing tests on the model of the drone airframe. Step S14: Statistically classify the damage conditions of each main component unit in the model of the drone airframe during the N random landing tests, record the damage quantity of each main component unit in the external structure and the internal structure of the model of the drone airframe during each random landing test, and determine the key detection parts of the drone airframe as the main component units with more damage quantity. Step S15: Statistically analyze the damage types that occur in the key detection parts of the drone airframe during the random landing test in Step S13 respectively.

[0006] Further, in step S3, the preprocessing module includes an image size adjustment module, a normalization processing module, a color space conversion module, and an output dimension adjustment module connected in sequence. The image size adjustment module is used to adjust the image to the image size adapted to the input of the network model; the normalization processing module is used to normalize the pixel values in the image to values between 0 and 1 to reduce the oscillation and instability of the network model during training and accelerate the convergence speed of the model; the color space conversion module is used to convert the image channel order to the BGR format; the output dimension adjustment module is used to convert the input image into a four-dimensional tensor to ensure the smooth execution of feature extraction and processing tasks by subsequent network layers.

[0007] Further, in step S3, the backbone network module includes a main path, which is composed of five blocks D1 to D5 connected in sequence. The first block D1 is composed of a convolutional layer, and the second block D2, the third block D3, and the fourth block D4 are each composed of a convolutional layer and a MANet module. The fifth block D5 is composed of a convolutional layer, a MANet, and a fast spatial pyramid pooling module; the five blocks D1 to D5 respectively output feature maps B1 to B5.

[0008] Further, the MANet module first passes through a 1×1 bypass convolutional module Conv. After the 1×1 bypass convolutional module Conv, a middle main path, a left branch, and a right branch are connected in parallel. Among them, the middle main path outputs to a connection block after passing through a depthwise separable convolution DSCconv. The depthwise separable convolution DSCconv includes a 1×1 bypass convolutional module Conv, a depthwise convolution DwConv with a kernel size of k×k, and a pointwise convolution PwConv with a kernel size of k×k connected in sequence. The left branch is composed of a 1×1 bypass convolutional module Conv. The right branch is divided into a first left branch, a first middle branch, and a first right branch after passing through a Split layer. The first left branch and the first middle branch directly output to the connection block. The first right branch contains at least three sequentially connected ConvNeck layers, and each ConvNeck layer outputs to the connection block. Finally, the outputs of the middle main path, the left branch, and the right branch are connected by the connection block and then input into a 1×1 bypass convolutional module Conv for processing to obtain the output of the MANet module.

[0009] Further, in step S3, the neck network module includes a semantic collection module, a hypergraph calculation module, and a semantic scattering module connected in sequence. The semantic collection module is used to convert the visual feature map into the semantic space and construct a hypergraph. The hypergraph calculation module enables the propagation of high-order information between different layers and positions based on the constructed hypergraph. The semantic scattering module is used to perform fusion processing on the information to generate semantic features of different scales. Among them, the semantic collection module respectively performs downsampling operations on the output feature maps B1, B2, and B3 of the first block D1, the second block D2, and the third block D3 in the backbone network module, and simultaneously performs an upsampling operation on the output feature map B5 of the fifth block D5 in the backbone network module. Then, the results of the above downsampling and upsampling operations are concatenated with the output feature map B4 of the fourth block D4 in the backbone network module through a connection block and input to the hypergraph calculation module for processing. The hypergraph calculation module includes a convolutional layer with a kernel size of 1×1, a HyperConv module, and a MANet module connected in sequence. The result output by the hypergraph calculation module is input to the semantic scattering module for processing.

[0010] Further, the semantic scattering module includes three parallel paths. The first path includes a downsampling operation, a convolutional layer with a kernel size of 1×1, and a C2f module. The second path includes two MANet modules. The third path includes an upsampling operation and a MANet module. The first path first performs a downsampling operation, and then concatenates the output feature map after the downsampling operation with the output feature map B5 of the fifth block D5 in the backbone network module. The concatenated result is dimensionally reduced through a convolutional layer with a kernel size of 1×1, and then concatenated with the output of the second path after the downsampling operation. The concatenated result is feature-enhanced through the C2f module to obtain the output of the first path. The second path concatenates the input with the output feature map B4 of the fourth block D4 in the backbone network module and then performs feature extraction through the MANet module. Then, it is concatenated with the output of the third path after the downsampling operation, and the concatenated result is further subjected to feature extraction through a MANet module to obtain the output of the second path. The third path concatenates the input after the upsampling operation with the output feature map B3 of the third block D3 in the backbone network module, and then performs feature extraction through the MANet module on the concatenated result to obtain the output of the third path.

[0011] Further, the C2f module includes a 1×1 bypass convolution module Conv and a Split layer connected in sequence. After the Split layer, it is divided into a second left branch, a second middle branch, and a second right branch. The second left branch and the second middle branch are directly output to the connection block. The second right branch includes multiple sequentially connected ConvNeck layers, and each ConvNeck layer outputs to the connection block. The outputs of the second left branch, the second middle branch, and the second right branch are connected through the connection block and then input into a 1×1 bypass convolution module Conv for processing to obtain the output of the C2f module.

[0012] Further, the HyperConv module first performs a reshaping operation on the input feature map. The reshaped feature map is divided into two paths. One path directly enters the HGNN + convolutional layer for fusion processing. The other path sequentially performs distance calculation and threshold processing to screen and determine whether to continue the iteration, and conveys the result of the screening decision to the HGNN + convolutional layer; if it is decided to continue the iteration, the fusion result of the HGNN + convolutional layer is re-input into the HGNN + convolutional layer for iterative fusion. If it is decided to terminate the iteration, the current fusion result feature map of the HGNN + convolutional layer is reshaped back to the output with the same dimension as the input of the HyperConv module.

[0013] Further, step S6 specifically includes the following steps:

[0014] Step S61, according to the influence degree of different damage types on the health condition of the drone, define a weighted mapping function for calculating the damage score , such that under the same damage type, the greater the damage degree, the higher the damage score. Under the same damage degree, the damage score of the damage type that poses a greater threat to the health of the drone is higher. The full score of the damage score is 100 points; where represents the weighted mapping function, x represents the damage type, y represents the damage quantity of the corresponding damage type, and z represents the damage degree value of the corresponding damage type; the weighted mapping function for calculating the damage score is expressed by the following formula: , where represents the damage degree value of each damage type, represents the damage score of each damage type, and e represents the natural constant;

[0015] Step S62, according to the weighted mapping function Map the detection results of each image in the UAV image set to be measured, and calculate the overall damage score G of each image Score , and take the overall damage score G Score of all images, and take the maximum value G max as the overall health score of the UAV to be measured; the calculation formula of the overall damage score G of each image Score is as follows: , wherein, represents the maximum value of the damage degree values of each damage type, represents the adjusted weight of each damage type;

[0016] Step S63, stipulate the health status levels of the UAV and the corresponding score ranges of each level, and obtain the health status level of the UAV to be measured according to the range where the overall health score of the UAV to be measured obtained in step S62 is located; The health status levels of the UAV and the corresponding score ranges of each level are specifically as follows: .

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] The present invention conducts physical modeling and force analysis on the UAV, obtains the damage characteristics of the UAV body, can simulate various damages generated by the UAV in the real use scenario, obtains the key detection parts and damage types through statistics, and makes the detection more efficient; uses visual sensing technology and ray detection technology to collect damage images, constructs a UAV structure damage detection network model, accurately identifies various damages of the UAV to be measured, realizes intelligent detection of the damage of the internal and external structures of the UAV body, and obtains information such as the damage type, damage quantity, damage location and the scale of the damage degree value of the UAV to be measured, enabling the maintainer to intuitively understand the damage situation of the UAV; according to the detection results and the body health assessment algorithm based on empirical weighting, conducts a weighted assessment of the health status of the UAV body, can directly obtain the damage degree of the UAV to be measured, and intelligently obtains whether the damage contained in the current UAV to be measured seriously affects flight, so as to facilitate the maintainer to plan the maintenance plan.

[0019] The present invention can realize intelligent detection and evaluation of potential damages of the UAV body, can optimize the UAV maintenance process, reduce the UAV maintenance cost, and is an effective method to ensure the safe flight and efficient operation of the UAV. Description of the Drawings

[0020] Figure 1 is the implementation flowchart of an intelligent analysis method for realizing UAV body health management of the present invention;

[0021] Figure 2 It is a schematic structural diagram of the UAV structural damage detection network model in the present invention;

[0022] Figure 3 It is a schematic structural diagram of the semantic scattering module in the neck network module in the embodiment of the present invention;

[0023] Figure 4 It is a schematic structural diagram of the MANet module in the embodiment of the present invention;

[0024] Figure 5 It is a schematic structural diagram of the C2f module in the embodiment of the present invention;

[0025] Figure 6 It is a schematic structural diagram of the HyperConv module in the embodiment of the present invention;

[0026] Figure 7 It is a statistical chart of the number of each damage type in the expanded image dataset in the embodiment of the present invention;

[0027] Figure 8 It is a damage detection result diagram of the internal structure of the UAV body in the embodiment of the present invention;

[0028] Figure 9 It is a damage detection result diagram of the external structure of the UAV body in the embodiment of the present invention. Detailed implementation manners

[0029] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific implementation manners described herein are only used to explain the present invention and are not used to limit the present invention.

[0030] Embodiment:

[0031] As Figure 1 shown, this embodiment provides an intelligent analysis method for realizing the health management of the UAV body, including the following steps:

[0032] Step S1, perform physical modeling on the UAV body to obtain the model of the UAV body; through randomly adjusting the landing attitude and flight state of the model of the UAV body, perform a random landing experiment, and conduct a force analysis on the model of the UAV body, so as to determine the parts that need to be focused on detecting and the possible damage types of the UAV body. Specifically, it includes:

[0033] Step S11: Obtain the material parameters of the UAV airframe, including the composition materials of the external skin of the airframe, the composition materials of the internal beams and ribs of the airframe, the out-of-plane uniaxial compressive yield strength, the strength of the carbon fiber material used, and the density elastic modulus of the carbon fiber. In this embodiment, the composition material of the external skin of the airframe is Nomex aramid honeycomb material, the composition materials of the internal beams and ribs of the airframe are carbon fiber, the out-of-plane uniaxial compressive yield strength is 3.34 MPa, the strength of the carbon fiber material used is 1000 MPa, and the density elastic modulus of the carbon fiber is 200 GPa to 600 GPa. Obtain the structural composition of the UAV airframe, including the structural dimensions of the wing (including the middle wing, aileron, outer wing, etc.), the fuselage (including beams, ribs, tail struts, etc.), the tail wing (including the vertical tail, horizontal tail, etc.), and the landing gear. According to the obtained UAV airframe material parameters and the UAV airframe structural composition, use Abaqus software to perform physical modeling on the UAV airframe to obtain the model of the UAV airframe.

[0034] Step S12: Divide the model of the UAV airframe into two parts, the external structure and the internal structure. Each part can be divided into multiple main component units. The external structure includes component units such as the fuselage, middle wing, aileron, outer wing, tail strut, vertical tail, and horizontal tail. The internal structure includes component units such as beams, ribs, vertical single and double ears at the connection of the middle wing and the outer wing, and internal bushings at the front end of the tail strut.

[0035] Step S13: Set the landing height of the random landing experiment to be between 10 m and 50 m, and the landing inclination angle to be between -30° and 30°. Conduct N random landing tests on the model of the UAV airframe obtained after the physical modeling in Step S11.

[0036] Step S14: Statistically classify the damage conditions that occur in each main component unit in the model of the UAV airframe during the N random landing tests in Step S13. Record the number of damages that occur in each main component unit in the external structure and the internal structure of the model of the UAV airframe during each random landing test. Designate the main component unit with a larger number of damages as the key detection part of the UAV airframe. In this embodiment, through statistical classification, it is obtained that the key detection part of the external structure of the UAV airframe is the middle wing, and the key detection parts of the internal structure of the UAV airframe are the beams and ribs.

[0037] Step S15: Statistically count the damage types that occur in the key detection parts of the UAV airframe, namely the middle wing, beams, and ribs, during the random landing tests in Step S13. Among them, the damage types that occur in the middle wing are skin peeling, perforation, cracks, and fractures. The damage types that occur in the beams are beam fractures and beam adhesions. The damage types that occur in the ribs are rib fractures and rib adhesions.

[0038] Step S2: Use a small field-of-view camera and an X-ray machine to collect image data of key detection parts of multiple physical UAV airframes. Collect image data with different viewing distances and ray intensities, and preprocess the collected image data to construct an image dataset. Specifically, it includes:

[0039] Step S21: Select a small field-of-view camera, set the size (pixels) of the collected image data to 3000×3000, and select shooting distances of 10 cm, 15 cm, and 20 cm.

[0040] Step S22: At shooting distances of 10 cm, 15 cm, and 20 cm, use the small field-of-view camera to take pictures of different parts of the outer middle wing of multiple physical UAV airframes, and collect a total of 1000 damaged images of the outer structure of the UAV airframe.

[0041] Step S23: Select a 2512 X-ray machine. According to the UAV airframe structure composition and UAV airframe material parameters, set the maximum tube current of the 2512 X-ray machine to 15 mA and the maximum tube voltage to 250 KV. Use the 2512 X-ray machine to irradiate the physical UAV airframe to obtain damaged images of the internal structure of the UAV airframe including beam fractures, beam adhesions, rib fractures, and rib adhesions. A total of 400 damaged images of the internal structure of the UAV airframe with a size of 3000×3000 are obtained.

[0042] Step S24: Combine the damaged images of the UAV airframe external structure and the damaged images of the UAV airframe internal structure to form an airframe structure damaged image set.

[0043] Step S25: Set label names corresponding to the damage type names for the airframe structure damaged image set. Among them, the label names corresponding to skin peeling, perforation, crack, fracture, rib fracture, rib adhesion, beam fracture, and beam adhesion in the damage type are set as SD, TR, CRK, FRA, RF, RA, BF, and BA respectively. Perform annotation processing on the damaged images in the airframe structure damaged image set according to the label names, and jointly form an image dataset with the generated annotation result file and the airframe structure damaged image set.

[0044] Step S3: Build a UAV structure damage detection network model based on the Hyper-YOLO network. The UAV structure damage detection network model includes a preprocessing module, a backbone network module, a neck network module, and a head network module. The preprocessing module is used to process the original damaged images in the image dataset to obtain a standardized tensor. The backbone network module is used for feature extraction. The neck network module is used to achieve multi-scale feature fusion. The head network module is used to classify and regress the results of feature extraction and feature fusion, and output the target detection results including the damage location, damage type, damage quantity, and damage degree value of the UAV airframe structure.

[0045] Step S31, construct a preprocessing module. As Figure 2 shown, the preprocessing module includes an image size adjustment module, a normalization processing module, a color space conversion module, and an output dimension adjustment module connected in sequence. The image size adjustment module is used to adjust the image to the image size adapted to the input of the network model; the normalization processing module is used to normalize the pixel values in the image to values between 0 and 1 to reduce the oscillation and instability of the network model during training and accelerate the convergence speed of the model; the color space conversion module is used to convert the image channel order to the BGR format; the output dimension adjustment module is used to convert the input image into a four-dimensional tensor to ensure the smooth execution of the feature extraction and processing tasks by the subsequent network layers. The standardized tensor output by the preprocessing module is input into the backbone network module for processing.

[0046] Step S32, construct a backbone network module. As Figure 2 shown, the backbone network module includes a main path, which is composed of five blocks D1 - D5 connected in sequence. The first block D1 is composed of a convolutional layer. The first block D1 is used to compress the size of the input image, increase the number of channels at the same time, and perform basic feature extraction; the second block D2, the third block D3, and the fourth block D4 are each composed of a convolutional layer and a MANet module. The second block D2, the third block D3, and the fourth block D4 are respectively used for abstract feature enhancement, extraction of deep semantic features, and extraction of cross-region correlation features; the fifth block D5 is composed of a convolutional layer, a MANet module, and a fast spatial pyramid pooling module. The fifth block D5 is used to integrate multi-scale features; the output feature maps B1 of the first block D1, B2 of the second block D2, B3 of the third block D3, B4 of the fourth block D4, and B5 of the fifth block D5 obtained are respectively input into the neck network module.

[0047] The MANet module is used to enhance the feature extraction ability of the backbone network. As Figure 4As shown, the MANet module first passes through a 1×1 bypass convolution module Conv. After the 1×1 bypass convolution module Conv, there are a middle main path, a left branch, and a right branch connected in parallel. Among them, the output of the middle main path after passing through the depthwise separable convolution DSCconv is sent to the connection block. The depthwise separable convolution DSCconv includes a 1×1 bypass convolution module Conv, a depthwise convolution DwConv with a convolution kernel size of k×k, and a pointwise convolution PwConv with a convolution kernel size of k×k connected in sequence. The left branch consists of a 1×1 bypass convolution module Conv. The right branch is divided into a first left branch, a first middle branch, and a first right branch after passing through a Split layer. The first left branch and the first middle branch are directly output to the connection block. The first right branch contains at least three sequentially connected ConvNeck layers, and the output of each ConvNeck layer is sent to the connection block. Finally, the outputs of the middle main path, the left branch, and the right branch are connected by the connection block and then input into a 1×1 bypass convolution module Conv for processing to obtain the output of the MANet module.

[0048] Among them, the k×k bypass convolution module Conv is used for channel feature recalibration; the k×k bypass convolution module Conv consists of a convolution layer with a convolution kernel size of k×k, a batch normalization layer BN, and an activation function SiLU connected in sequence. In this embodiment, a 1×1 bypass convolution module Conv is used, and the convolution kernel size of the convolution layer included therein is 1×1.

[0049] The ConvNeck layer is used for fusing and compressing the semantic information of different features; the ConvNeck layer includes two convolution layers with a convolution kernel size of 3×3 connected in sequence, and a residual connection operation is added to directly add the input feature to the output of the second convolution layer.

[0050] The fast spatial pyramid pooling module is used to efficiently capture multi-scale features. The fast spatial pyramid pooling module first sequentially connects a convolution block with a convolution kernel size of 1×1 and three MaxPool2d layers. At the same time, the output results of the convolution block and the three MaxPool2d layers are respectively input into the splicing layer for splicing, and finally the splicing result is input into a convolution block with a convolution kernel of 1×1; among them, both convolution blocks include a Conv2d convolution layer, a BatchNorm2d batch normalization layer, and an activation function; the first convolution block is used to reduce the amount of calculation and extract preliminary features; the MaxPool2d layer is used to extract features of each scale; the splicing layer is used to fuse features of multiple scales; the second convolution block is used to convert the number of channels into the output feature channel number and further fuse features of different scales.

[0051] Step S33: Construct a neck network module. The neck network module includes a semantic collection module, a hypergraph calculation module, and a semantic scattering module connected in sequence. The semantic collection module is used to transform the visual feature map into the semantic space and construct a hypergraph. The hypergraph calculation module enables the propagation of high-order information between different layers and positions based on the constructed hypergraph. The semantic scattering module is used to perform fusion processing on the information to generate semantic features of different scales. The semantic collection module respectively performs downsampling operations on the output feature maps B1, B2, and B3 of the first block D1, the second block D2, and the third block D3 in the backbone network module, and simultaneously performs an upsampling operation on the output feature map B5 of the fifth block D5 in the backbone network module. Then, the results of the above downsampling and upsampling operations are concatenated with the output feature map B4 of the fourth block D4 in the backbone network module through a connection block and input to the hypergraph calculation module for processing. The hypergraph calculation module includes a convolutional layer with a kernel size of 1×1, a HyperConv module, and a MANet module connected in sequence. The result output by the hypergraph calculation module is input to the semantic scattering module for processing.

[0052] As Figure 3 shown, the semantic scattering module includes three parallel paths. The first path includes a downsampling operation, a convolutional layer with a kernel size of 1×1, and a C2f module. The second path includes two MANet modules. The third path includes an upsampling operation and a MANet module. The first path first performs a downsampling operation, and then concatenates the output feature map after the downsampling operation with the output feature map B5 of the fifth block D5 in the backbone network module. The concatenated result is subjected to dimensionality reduction processing through a convolutional layer with a kernel size of 1×1, and then concatenated with the output of the second path after the downsampling operation. The concatenated result is subjected to feature enhancement through the C2f module to obtain the output of the first path. The second path concatenates the input with the output feature map B4 of the fourth block D4 in the backbone network module and performs feature extraction through the MANet module, and then concatenates the result with the output of the third path after the downsampling operation. The concatenated result is further subjected to feature extraction through a MANet module to obtain the output of the second path. The third path concatenates the input after the upsampling operation with the output feature map B3 of the third block D3 in the backbone network module, and then performs feature extraction on the concatenated result through the MANet module to obtain the output of the third path. The outputs of the three paths of the semantic scattering module in the neck network module are respectively input to the head network module.

[0053] Among them, the C2f module is used to enhance the hierarchical integration of features. As Figure 5As shown in the figure, the C2f module includes a 1×1 bypass convolution module Conv and a Split layer connected in sequence. After the Split layer, it is divided into a second left branch, a second middle branch, and a second right branch. The second left branch and the second middle branch are directly output to the connection block. The second right branch includes multiple sequentially connected ConvNeck layers, and each ConvNeck layer outputs to the connection block. The outputs of the second left branch, the second middle branch, and the second right branch are connected through the connection block and then input into a 1×1 bypass convolution module Conv for processing to obtain the output of the C2f module.

[0054] The HyperConv module is used to fuse features of five different scales and enable the propagation of high-order information at different levels and positions. As Figure 6 shown, the HyperConv module first performs a reshaping operation on the input feature map. The reshaped feature map is divided into two paths. One path directly enters the HGNN + convolution layer for fusion processing. The other path sequentially performs distance calculation and threshold processing to screen and determine whether to continue the iteration, and conveys the result of the screening decision to the HGNN + convolution layer. If it is decided to continue the iteration, the fusion result of the HGNN + convolution layer is re-input into the HGNN + convolution layer through the feedback connection for iterative fusion. If it is decided to terminate the iteration, the current fusion result feature map of the HGNN + convolution layer is reshaped back to the output with the same dimension as the input of the HyperConv module through a reshaping operation. In this embodiment, the number of iterative fusions of the HGNN + convolution layer is between 1 and 5 times.

[0055] Step S34, construct the head network module. The head network module is used to classify and regress the results of feature extraction and feature fusion, and output the target detection results including the damage location, damage type, damage quantity, and damage degree value of the UAV airframe structure. The head network module includes three detection heads, and the decoupling head is included in the detection head. The head network module uses the decoupling head to separately hand the regression task and the classification task in the task to different branches of the detection head. In the regression task branch, the input data first passes through a Conv Module convolution module to initially extract the position offset of the target bounding box, and then enters a Conv2d convolution layer to expand the feature dimension, and finally generates the output related to the bounding box regression, and calculates the bounding box loss Bbox Loss. In the classification task branch, the input data also first passes through the Conv Module convolution module to initially extract the category semantic information of the target, and then enters another Conv2d convolution layer to expand the feature dimension, and finally generates the output related to the target classification, and calculates the classification loss Cls Loss.

[0056] Step S4: Divide the image dataset obtained in step S25 into a training set and a validation set. Based on the image characteristics of the image dataset, adjust the parameters of the UAV structural damage detection network model constructed based on the Hyper-YOLO network in step S3, and use the training set to train the network model to obtain a trained and stable UAV structural damage detection network model. Specifically, it includes:

[0057] Step S41: Divide the image dataset obtained in step S25 into a training set and a validation set at a ratio of 9:1.

[0058] Step S42: Use data augmentation methods such as randomly adjusting brightness, contrast, and saturation, adding Gaussian noise, salt noise, pepper noise, and horizontal rotation, vertical rotation, and central cropping to augment the training set and validation set obtained in step S41 to obtain the augmented training set and validation set. The augmented training set and validation set together constitute the augmented image dataset. In this embodiment, the quantity statistics of each damage type in the damage images included in the augmented image dataset are as Figure 7 shown.

[0059] Step S43: Based on the image characteristics of the training set obtained in step S42, such as image size, contrast, and brightness, adjust the parameters of the UAV structural damage detection network model constructed in step S3.

[0060] Step S44: Input the augmented training set obtained in step S42 into the UAV structural damage detection network model with adjusted parameters for training, and verify it with the validation set to obtain a trained and stable UAV structural damage detection network model.

[0061] Step S5: Collect images of the body of the UAV to be measured to form an image set of the UAV to be measured; use the trained and stable UAV structural damage detection network model in step S4 to detect the structural damage of the UAV body in the image set of the UAV to be measured to obtain the detection result. Specifically, it includes:

[0062] Step S51: Collect images of the key detection parts of the external and internal structures of the UAV body to be measured to form an image set of the UAV to be measured.

[0063] Step S52: Use the trained and stable UAV structural damage detection network model to detect the image set of the UAV to be measured, and obtain the detection result of each image. The detection result includes the damage type, the number of damages corresponding to the damage type, and the damage degree value corresponding to the damage type. In this embodiment, the damage detection effects of the internal and external structures of the UAV body are as Figure 8 、 Figure 9 shown.

[0064] Step S6: Based on the body health assessment algorithm with empirical weighting, analyze the damage information recorded in the detection results to comprehensively evaluate the health status of the UAV to be tested. The specific steps are as follows:

[0065] Step S61: There may be multiple damages with different damage types and different damage degree values on one image in the UAV image set to be tested. It is necessary to calculate the damage score of a single image according to all the damages in the detection results of each image. According to the influence degree of different damage types on the health of the UAV, define a weighted mapping function for calculating the damage score , such that under the same damage type, the greater the damage degree, the higher the damage score; under the same damage degree, the damage score of the damage type that poses a greater threat to the health of the UAV is higher, and the full score of the damage score is 100 points. Among them, represents the weighted mapping function, x represents the damage type, y represents the number of damages of the corresponding damage type, and z represents the damage degree value of the corresponding damage type. Define the weighted mapping function for calculating the damage score The specific process is as follows:

[0066] Step S611: Define the maximum shedding area of the skin shedding damage as the threshold for judging the initial weight of each damage type According to the range where the maximum shedding area of the skin shedding damage is located, stipulate the initial weights of each damage type , including: the initial weight W1 of the fracture damage, the initial weight W2 of the crack damage, the initial weight W3 of the perforation damage, and the initial weight W4 of the skin shedding damage. Specifically:

[0067] When the maximum shedding area of the skin shedding damage ≤ 16 cm 2 , define the initial weight W1 of the fracture damage as 0.4, the initial weight W2 of the crack damage as 0.3, the initial weight W3 of the perforation damage as 0.3, and the initial weight W4 of the skin shedding damage as 0; when the maximum shedding area of the skin shedding damage > 16 cm 2 , define the initial weight W1 of the fracture damage as 0.4, the initial weight W2 of the crack damage as 0.3, the initial weight W3 of the perforation damage as 0.25, and the initial weight W4 of the skin shedding damage as 0.05.

[0068] Step S612: Allocate the weights of each damage type in the corresponding image according to the damage situation included in each image.

[0069] Count the number of each damage type existing in the image , including: the number N1 of the fracture damage, the number N2 of the crack damage, the number N3 of the perforation damage, and the number N4 of the skin shedding damage; according to the number of each damage type Initial weights for each type of damage Calculate the weighted weight values for each type of damage ;

[0070] Weighted weight values for each type of damage The calculation formula is as follows: , where represents the quantity of each type of damage in the image, represents the initial weight of each type of damage, represents the weighted weight value of each type of damage.

[0071] Based on the weighted weight values of each type of damage in each image calculate the total damage weight of the corresponding image , and the calculation formula is as follows: ,

[0072] If the total damage weight of the image is 0, directly define the adjusted weights of each type of damage in the image as all 0; if the total damage weight of the image is not 0, then based on the weighted weight values of each type of damage in each image and the total damage weight of the image perform weight normalization to calculate the adjusted weights of each type of damage in the corresponding image , and the calculation formula is: .

[0073] Step S613: Classify the damage degree levels of skin peeling damage, perforation damage, crack damage, and fracture damage, and assign damage score ranges to each damage degree level of each type of damage. Specifically as follows:

[0074] For fracture damage, use the longest cross-sectional distance (cm) as the basis for determining the damage degree, and use the value of the longest cross-sectional distance as the damage degree value of the fracture damage. When the damage degree value of the fracture damage is less than 1 cm, it is defined as mild damage, and the corresponding assigned damage score is from 0 to 30 points; when the damage degree value of the fracture damage is greater than or equal to 1 cm and less than or equal to 3 cm, it is defined as moderate damage, and the corresponding assigned damage score is from 30 to 80 points; when the damage degree value of the fracture damage exceeds 3 cm, it is defined as severe damage, and the corresponding assigned damage score is from 80 to 100 points.

[0075] The crack damage uses the crack length (cm) as the basis for judging the damage degree, and uses the numerical value of the crack length as the damage degree value of the crack damage. When the crack length of the crack damage is less than 1 cm, it is defined as mild damage, and the corresponding damage score is assigned from 0 to 30 points; when the crack length of the crack damage is greater than or equal to 1 cm and less than or equal to 3 cm, it is defined as moderate damage, and the corresponding damage score is assigned from 30 to 80 points; when the crack length of the crack damage exceeds 3 cm, it is defined as severe damage, and the corresponding damage score is assigned from 80 to 100 points.

[0076] The perforation damage uses the area (cm 2 ) as the basis for judging the damage degree, and uses the numerical value of the area as the damage degree value of the perforation damage. When the area of the perforation damage is less than 1 cm 2 , it is defined as mild damage, and the corresponding damage score is assigned from 0 to 30 points; when the area of the perforation damage is greater than or equal to 1 cm 2 and less than or equal to 3 cm 2 , it is defined as moderate damage, and the corresponding damage score is assigned from 30 to 80 points; when the area of the perforation damage exceeds 3 cm 2 , it is defined as severe damage, and the corresponding damage score is assigned from 80 to 100 points.

[0077] The skin shedding damage uses the shedding area (cm 2 ) as the basis for judging the damage degree, that is, uses the numerical value of the shedding area as the damage degree value of the skin shedding damage. When the shedding area of the skin shedding damage is less than or equal to 16 cm 2 , it belongs to mild damage, and the evaluated damage score is 0 points; when it is greater than 16 cm 2 , it belongs to severe damage, and the evaluated damage score is 100 points.

[0078] Step S614, define the logistic function as the weighted mapping function for calculating the damage score .

[0079] The total score of the damage score , in step S613, the two sets of data of the damage degree values and damage scores at the dividing boundaries of the damage degree levels (mild damage, moderate damage, severe damage) of the perforation damage, crack damage, and fracture damage are and , where S1 represents the damage degree value at the boundary dividing mild damage and moderate damage, represents the damage score at the boundary dividing mild damage and moderate damage, S2 represents the damage degree value at the boundary dividing moderate damage and severe damage, represents the damage score at the boundary dividing moderate damage and severe damage. According to the total score of the damage score and the damage degree values at the boundaries of the two damage degree levels , and the damage score , to obtain the coefficients of the logistic function and , thereby determining the logistic function. The logistic function is as follows: , wherein, represents the damage degree value of each damage type, represents the damage score of each damage type, and e represents the natural constant. Substitute the total score of the damage score into the logistic function, and substitute and respectively as and into the logistic function to form a system of equations, and solve the system of equations to obtain , that is, the weighted mapping function for calculating the damage score The corresponding logistic function is .

[0080] Step S62, according to the weighted mapping function for calculating the damage score, map the detection results of each image in the set of images of the UAV to be tested, and calculate the overall damage score G Score of each image. Take the maximum value of all the overall damage scores G Score of all the images as the overall health score result of the UAV to be tested. The specific method is as follows: According to the maximum value of the damage degree values of each damage type in each image and the adjusted weight of each damage type obtained in step S612 to calculate the weighted score of each damage type, that is, substitute the maximum value of the damage degree values of each damage type into in the logistic function to calculate the damage score of each damage type, multiply it by the adjusted weight Score to obtain the weighted score of each damage type; sum up the weighted scores of all damage types to obtain the overall damage score G max of each image; calculate the overall damage scores of all the images in the set of images of the UAV to be tested, and then take the maximum value G

[0081] of all the overall damage scores of all the images as the overall health score of the UAV to be tested. Score Among them, the calculation formula of the overall damage score G of each image is as follows: wherein, Represents the maximum value of the damage degree values for each damage type. Represents the adjusted weight for each damage type.

[0082] Among them, since the damage to the beams and ribs in the internal structure of the aircraft poses a greater threat to the safety of the aircraft, if the images in the image set of the UAV to be tested contain damage types such as rib fracture, rib adhesion, beam fracture, or beam adhesion, then the overall damage score G of this image Score Is rated 100 points.

[0083] Step S63: Specify the health status levels of the UAV and the corresponding score ranges for each level, and based on the range in which the overall health score of the UAV to be tested obtained in step S62 is located, obtain the health status level of this UAV to be tested.

[0084] In this embodiment, the health status levels of the UAV and the corresponding score ranges for each level are specifically as follows: when the overall health score of the UAV is greater than or equal to 0 points and less than 30 points, the health status level is mild airframe damage; when the overall health score of the UAV is greater than or equal to 30 points and less than or equal to 80 points, the health status level is moderate airframe damage; when the overall health score of the UAV is greater than 80 points and less than or equal to 100 points, the health status level is severe airframe damage.

[0085] Calculation example of the overall health score of the UAV to be tested:

[0086] Suppose that in an image in the image set of the UAV to be tested in step S52, there is a fracture damage with the longest cross-sectional distance of 2 cm, a crack damage with a crack length of 4 cm and two crack damages with a crack length of 1 cm, two perforation damages with an area of 1 cm 2 Each, and five skin peeling damages with the largest area of 10 cm 2 Each.

[0087] Then, the number of each damage included in this image includes the number of fracture damages N1 = 1, the number of crack damages N2 = 3, the number of perforation damages N3 = 2, and the number of skin peeling damages N4 = 5.

[0088] The maximum value of the damage degree values for each damage type contained in this image is for fracture damage: , for crack damage: , for perforation damage: , for skin peeling: .

[0089] According to the initial weight regulations for each damage type in step S611, the initial weight of fracture damage W1 = 0.4, the initial weight of crack damage W2 = 0.3, the initial weight of perforation damage W3 = 0.3, and the initial weight of skin peeling damage W4 = 0.

[0090] Adjust the weight of the fracture damage included in the image according to step S612 For , the adjusted weight of crack damage is , the adjusted weight of perforation damage is , the adjusted weight of skin peeling damage is .

[0091] According to the maximum value of the damage degree values of the above-mentioned various damage types and the adjusted weights of the various damage types, calculate the overall damage score G of this image by the calculation formula in step S62 Score It is about 56 points

[0092] If the overall damage score of this image is the highest among all the images in the image set of the UAV to be tested, then use the detection result of this image as the overall health score result of the UAV to be tested, that is, the overall health score of the UAV to be tested is 56 points. Then the health status level of the UAV to be tested is rated as moderate airframe damage

[0093] The above is only a preferred embodiment of the present invention, and is not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention

Claims

1. An intelligent analysis method for realizing the health management of an unmanned aerial vehicle body, characterized in that, It includes the following steps: Step S1: Physically model the UAV airframe to obtain a model of the UAV airframe. Conduct a force analysis on the model of the UAV airframe through a random landing experiment to determine the key detection parts of the UAV airframe and the corresponding damage types; Step S2: Use a small field-of-view camera and an X-ray machine to collect image data of the key detection parts of multiple physical UAV airframes. Collect image data with different viewing distances and ray intensities, and preprocess the collected image data to construct an image dataset; Step S3: Build a UAV structural damage detection network model based on the Hyper-YOLO network. The UAV structural damage detection network model includes a preprocessing module, a backbone network module, a neck network module, and a head network module. The preprocessing module is used to process the images in the image dataset to obtain a standardized tensor. The backbone network module is used for feature extraction. The neck network module is used to achieve multi-scale feature fusion. The head network module is used to classify and regress the results of feature extraction and feature fusion, and output a target detection result including the damage location, damage type, damage quantity, and damage degree value of the UAV airframe structure; Step S4: Divide the image dataset obtained in Step S2 into a training set and a validation set. Based on the image characteristics of the image dataset, adjust the parameters of the UAV structural damage detection network model constructed in Step S3, and use the training set to train the network model to obtain a trained UAV structural damage detection network model; Step S5: Collect images of the UAV airframe to be tested to form a UAV image set to be tested; Use the trained UAV structural damage detection network model in Step S4 to detect the UAV airframe structure damage of the UAV image set to be tested to obtain a detection result; Step S6: Based on the body health assessment algorithm with empirical weighting, analyze the damage information recorded in the detection result to comprehensively evaluate the health status of the UAV airframe to be tested.

2. The intelligent analysis method for realizing the health management of the UAV airframe according to claim 1, characterized in that, The specific steps of Step S1 include the following steps: Step S11: Obtain the UAV airframe material parameters and the UAV airframe structure composition. Physically model the UAV airframe according to the obtained UAV airframe material parameters and UAV airframe structure composition to obtain a model of the UAV airframe; Step S12: Divide the model of the UAV airframe into two parts, the external structure and the internal structure, and each part is divided into multiple main component units; Step S13: Set the landing simulation parameters of the random landing experiment. The landing simulation parameters include the landing height and the landing tilt angle, and conduct N random landing tests on the model of the UAV airframe; Step S14: Statistically classify the damage conditions of each main component unit in the model of the UAV airframe during the N random landing tests, record the damage quantity of each main component unit in the external structure and the internal structure of the model of the UAV airframe during each random landing test, and determine the main component units with more damage quantity as the key detection parts of the UAV airframe; Step S15: Count the damage types that occur in the key detection parts of the UAV body during the random landing test in Step S13 respectively.

3. The intelligent analysis method for realizing the health management of the UAV airframe according to claim 1, characterized in that, In Step S3, the preprocessing module includes an image size adjustment module, a normalization processing module, a color space conversion module, and an output dimension adjustment module connected in sequence. The image size adjustment module is used to adjust the image to the image size adapted to the input of the network model; the normalization processing module is used to normalize the pixel values in the image to values between 0 and 1 to reduce the oscillation and instability of the network model during training and accelerate the convergence speed of the model; the color space conversion module is used to convert the image channel order to the BGR format; the output dimension adjustment module is used to convert the input image into a four-dimensional tensor to ensure the smooth execution of feature extraction and processing tasks by subsequent network layers.

4. The intelligent analysis method for realizing the health management of the UAV airframe according to claim 3, characterized in that, In Step S3, the backbone network module includes a main path, which is composed of five blocks D1 - D5 connected in sequence. The first block D1 is composed of a convolutional layer, the second block D2, the third block D3, and the fourth block D4 are each composed of a convolutional layer and a MANet module, and the fifth block D5 is composed of a convolutional layer, a MANet module, and a fast spatial pyramid pooling module; the five blocks D1 - D5 respectively output feature maps B1 - B5.

5. The intelligent analysis method for realizing the health management of the UAV airframe according to claim 4, characterized in that, The MANet module first passes through a 1×1 bypass convolutional module Conv. After the 1×1 bypass convolutional module Conv, there are a middle main path, a left branch, and a right branch connected in parallel. Among them, the middle main path outputs to the connection block after passing through the depthwise separable convolution DSCconv. The depthwise separable convolution DSCconv includes a 1×1 bypass convolutional module Conv, a depthwise convolution DwConv with a convolutional kernel size of k×k, and a pointwise convolution PwConv with a convolutional kernel size of k×k connected in sequence. The left branch is composed of a 1×1 bypass convolutional module Conv. The right branch is divided into a first left branch, a first middle branch, and a first right branch after passing through a Split layer. The first left branch and the first middle branch directly output to the connection block. The first right branch contains at least three sequentially connected ConvNeck layers, and each ConvNeck layer outputs to the connection block. Finally, the outputs of the middle main path, the left branch, and the right branch are connected by the connection block and then input into a 1×1 bypass convolutional module Conv for processing to obtain the output of the MANet module.

6. The intelligent analysis method for realizing the health management of the UAV airframe according to claim 5, characterized in that, In step S3, the neck network module includes a semantic collection module, a hypergraph calculation module, and a semantic scattering module connected in sequence. The semantic collection module is used to convert the visual feature map into the semantic space and construct a hypergraph; the hypergraph calculation module enables the propagation of high-order information between different layers and positions based on the constructed hypergraph; the semantic scattering module is used to perform fusion processing on the information to generate semantic features of different scales. Among them, the semantic collection module performs downsampling operations on the output feature maps B1, B2, and B3 of the first block D1, the second block D2, and the third block D3 in the backbone network module respectively, and at the same time performs an upsampling operation on the output feature map B5 of the fifth block D5 in the backbone network module. Then, the results of the above downsampling operation and upsampling operation are concatenated with the output feature map B4 of the fourth block D4 in the backbone network module through a connection block and input to the hypergraph calculation module for processing. The hypergraph calculation module includes a convolutional layer with a kernel size of 1×1, a HyperConv module, and a MANet module connected in sequence. The result output by the hypergraph calculation module is input to the semantic scattering module for processing.

7. The intelligent analysis method for realizing the health management of the UAV airframe according to claim 6, characterized in that, The semantic scattering module includes three parallel paths. The first path includes a downsampling operation, a convolutional layer with a kernel size of 1×1, and a C2f module. The second path includes two MANet modules. The third path includes an upsampling operation and a MANet module. The first path first performs a downsampling operation, and then concatenates the output feature map after the downsampling operation with the output feature map B5 of the fifth block D5 in the backbone network module. The concatenated result is processed by a convolutional layer with a kernel size of 1×1 for dimensionality reduction, and then concatenated with the output of the second path after the downsampling operation. The concatenated result is enhanced by the C2f module to obtain the output of the first path. The second path concatenates the input with the output feature map B4 of the fourth block D4 in the backbone network module and then extracts features through the MANet module, and then concatenates with the output of the third path after the downsampling operation. The concatenated result is then processed by another MANet module for feature extraction to obtain the output of the second path. The third path concatenates the input after the upsampling operation with the output feature map B3 of the third block D3 in the backbone network module, and then extracts features through the MANet module to obtain the output of the third path.

8. The intelligent analysis method for realizing the health management of the UAV airframe according to claim 7, characterized in that The C2f module includes a 1×1 bypass convolutional module and a Split layer connected in sequence. After the Split layer, it is divided into a second left branch, a second middle branch, and a second right branch. The second left branch and the second middle branch are directly output to the connection block. The second right branch includes multiple ConvNeck layers connected in sequence, and each ConvNeck layer is output to the connection block. The outputs of the second left branch, the second middle branch, and the second right branch are connected through the connection block and input to a 1×1 bypass convolutional module Conv for processing to obtain the output of the C2f module.

9. The intelligent analysis method for realizing the health management of the UAV airframe according to claim 8, characterized in that, The HyperConv module first performs a reshaping operation on the input feature map. The reshaped feature map is divided into two paths. One path directly enters the HGNN + convolutional layer for fusion processing. The other path performs distance calculation and threshold processing in sequence to screen and determine whether to continue the iteration, and conveys the result of the screening decision to the HGNN + convolutional layer. If it is decided to continue the iteration, the fusion result of the HGNN + convolutional layer is re-input into the HGNN + convolutional layer for iterative fusion. If it is decided to terminate the iteration, the current fusion result feature map of the HGNN + convolutional layer is reshaped back to the output with the same dimension as the input of the HyperConv module.

10. The intelligent analysis method for realizing the health management of the UAV airframe according to claim 9, characterized in that, Further, the step S6 specifically includes the following steps: Step S61: Define a weighted mapping function for calculating the damage score according to the influence degree of different damage types on the health of the drone , such that under the same damage type, the greater the damage degree, the higher the damage score; under the same damage degree, the damage score of the damage type that poses a greater threat to the health of the drone is higher, and the full score of the damage score is 100 points; among them, represents the weighted mapping function, x represents the damage type, y represents the number of damages of the corresponding damage type, and z represents the damage degree value of the corresponding damage type; the weighted mapping function is expressed by the following formula: ， Among them, represents the injury degree value of each injury type, represents the injury score of each injury type, and e represents the natural constant; Step S62, according to the weighted mapping function for calculating the damage score map the detection results of each image in the UAV image set to be measured, and calculate the overall damage score G of each image Score , and take the maximum value G Score in all the overall damage scores G of the images max as the overall health score of the UAV to be measured; the calculation formula of the overall damage score G of each image Score is as follows: ， Among them, represents the maximum value of the damage degree values of each damage type, represents the adjusted weight of each damage type; Step S63: Specify the health status levels of the UAVs and the corresponding score ranges for each level, and obtain the health status level of the UAV to be tested according to the range in which the overall health score of the UAV to be tested obtained in step S62 lies; The health status levels of the UAVs and the corresponding score ranges for each level are specifically as follows: 。

Images