Method and apparatus for identifying the integrity of the recovery of landing gear safety pins and pitot tube covers

Abstract

The present application relates to the technical field of aviation safety, and particularly relates to a landing gear safety pin and air speed tube sleeve recovery integrity identification method and device, wherein the identification method comprises: acquiring a real-time image inside a tool box; the tool box is used for placing a plurality of safety pins and a plurality of air speed tube sleeves recovered from a landing gear, and a plurality of first areas for placing safety pins and a plurality of second areas for placing air speed tube sleeves are arranged in the tool box; based on a target detection model, it is identified whether safety pins exist in each first area and whether air speed tube sleeves exist in each second area in the real-time image, and if yes, it is determined that all safety pins and air speed sleeves have been completely recovered; wherein the target detection model is a deep learning model with adaptive weight parameters obtained through a pre-training process. The present application can accurately identify the recovery integrity of landing gear safety pins and air speed tube sleeves in real time.

Description

Technical Field

[0001] The present invention relates to the field of aviation safety technology, and in particular to a method and device for identifying the recovery integrity of a landing gear safety pin and a pitot tube sleeve. Background Art

[0002] In the civil aviation maintenance industry, maintenance personnel must remove the landing gear safety pin and pitot tube sleeve during each aircraft's pre-flight preparations. Failure to remove the safety pin or pitot tube sleeve can cause serious errors after the aircraft takes off. The current method for detecting whether the landing gear safety pin and pitot tube sleeve are fully recovered is manual visual inspection. Maintenance personnel remove the landing gear safety pin and pitot tube sleeve, take photos, and send them to the work group. The on-duty officer or other designated personnel will review the photos and manually check the integrity of the safety pin and pitot tube sleeve. The disadvantages of this method are that the manual review is labor-intensive, time-consuming, and slow. Furthermore, manual review is prone to omissions or negligence, and failure to detect abnormalities in the photos or omissions in the review can result in the aircraft taking off with the safety pin or pitot tube sleeve, posing a safety hazard. Summary of the Invention

[0003] (1) Technical issues to be resolved

[0004] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a method and device for identifying the recovery integrity of a landing gear safety pin and a pitot tube sleeve, which solves the technical problems of the manual visual method, which is labor-intensive and prone to omissions.

[0005] (2) Technical solution

[0006] In order to achieve the above objectives, the main technical solutions adopted by the present invention include:

[0007] In a first aspect, an embodiment of the present invention provides a method for identifying the recovery integrity of a landing gear shear pin and pitot tube sleeve, comprising: acquiring a real-time image of the interior of a toolbox; the toolbox is used to store multiple shear pins and multiple pitot tube sleeves recovered from the landing gear, and the toolbox has multiple first areas designated for storing shear pins and second areas designated for storing pitot tube sleeves;

[0008] identifying, based on the target detection model, whether a safety pin exists in each first region of the real-time image and whether a pitot tube sleeve exists in each second region; and if so, determining that all safety pins and pitot tube sleeves have been completely recovered;

[0009] The target detection model is a deep learning model with adapted weight parameters obtained through a pre-training process.

[0010] Optionally, the target detection model is obtained by replacing the convolution module, CSPLayer_2Conv module and SPPF module in the backbone network of the YOLOv8n neural network model with an inverted residual module;

[0011] Among them, the inverted residual module is a module in the MobileNetV3 neural network model.

[0012] Optionally, the input end of the backbone network of the target detection model is a Conv_BN_hswish module;

[0013] The Conv_BN_hswish module includes a common convolution unit, a batch normalization unit and an h-swish activation function unit connected sequentially.

[0014] Optionally, the activation function in the h-swish activation function unit is:

[0015]

[0016] Wherein, x represents the pixel value of each pixel in the real-time image, ReLU6=(ReLU,6), ReLU=max(0,x).

[0017] Optionally, the number of the first areas is three, the number of the second areas is three, and the pre-training process includes:

[0018] S1. Obtain at least 64 initial training images;

[0019] In the initial training image, 64 placement methods are classified according to whether there is a safety pin in each first area and whether there is a pitot tube sleeve in each second area;

[0020] S2. performing labeling and enhancement processing on the safety pin and the pitot tube sleeve in the initial training image to obtain a final training image;

[0021] S3. Among all the final training images, select the final training images in which all the safety pins and pitot tube sleeves are present as a positive example set, and the remaining final training images as a negative example set. The positive example set and the negative example set together constitute a training data set;

[0022] S4. Use the training data set to train the initial target detection model to obtain a target detection model with adapted weight parameters.

[0023] Optionally, the S1 includes:

[0024] According to multiple shooting angles, shooting directions and shooting brightness, a camera device is used to obtain initial training images based on 64 placement methods.

[0025] Optionally, in the S2,

[0026] The marking processing includes: using labelme software to mark the safety pin and the pitot tube cover in the initial training image, and generating a data file in JSON format.

[0027] The enhancement processing includes: cropping, rotating, translating, changing brightness, introducing noise, HSV enhancement, Cutout, flipping, introducing motion blur and introducing speckle noise.

[0028] Optionally, the S4 includes:

[0029] The training data set is divided into a training set, a validation set and a test set according to a ratio of 8:1:1;

[0030] The initial target detection model is trained using the training set, and after each round of training, the validation set is used to verify and update the model parameters of the current target detection model;

[0031] The current training process ends after 300 rounds of training;

[0032] It is judged whether the performance index reaches a preset value, and if so, the current target detection model is taken as a target detection model with adaptive weight parameters;

[0033] The performance index includes:

[0034]

[0035] Wherein P represents the accuracy, TP represents the number of images in the positive example set that are correctly detected, FP represents the number of images in the positive example set that are not correctly detected, FN represents the number of images in the negative example set that are incorrectly detected as positive examples, R represents the recall rate, P(R) represents the relationship curve between the accuracy P and the recall rate R, AP represents the average accuracy, mAP represents the average value of the average accuracy AP of all classification categories, n represents the classification category, and AP i represents the average accuracy of the i-th classification.

[0036] In a second aspect, an embodiment of the present application provides a landing gear safety pin and pitot tube cover recovery integrity identification device, comprising a processing device;

[0037] The processing device includes a memory, a processor, and a computer program stored on the memory and executable on the processor, and the computer program is executed by the processor to implement the landing gear safety pin and pitot tube cover recovery integrity identification method of the first aspect.

[0038] Optionally, an interactive device is also included;

[0039] The interactive device is used to interact with the user, obtain real-time images uploaded by the user, send the real-time images to the processing device to identify the recovery integrity of the landing gear safety pin and pitot tube sleeve in the real-time image and obtain an identification result; and receive and display the identification result returned by the processing device to the user.

[0040] (3) Beneficial effects

[0041] The beneficial effects of the present invention are as follows: the method and device for identifying the recovery integrity of landing gear safety pins and pitot tube sleeves of the present invention, after acquiring a real-time image of the interior of a toolbox, uses a target detection model to identify, based on a first area designated for safety pin placement and a second area designated for pitot tube sleeve placement within the toolbox, whether a safety pin is present in each first area of ​​the real-time image and a pitot tube sleeve is present in each second area, thereby determining that all safety pins and pitot tube sleeves have been completely recovered. Compared to the prior art, after acquiring a real-time image, the target detection model can be used to complete the remaining review work, effectively reducing manual workload. The recovery integrity of landing gear safety pins and pitot tube sleeves can be accurately and in real time, ensuring that the safety pins and pitot tube sleeves are removed and recovered into the toolbox before takeoff, thereby ensuring the safe flight of the aircraft. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A schematic flow chart of a method for identifying the recovery integrity of a landing gear shear pin and pitot tube sleeve provided in an embodiment;

[0043] Figure 2 This is a schematic diagram of the structure of the existing YOLOv8n neural network model;

[0044] Figure 3 A schematic diagram of the structure of the backbone network of the target detection model provided in the embodiment;

[0045] Figure 4 A schematic structural diagram of an inverted residual module provided in an embodiment;

[0046] Figure 5 A schematic diagram of a flow chart of a pre-training process provided in an embodiment;

[0047] Figure 6 An initial training image partially collected for the embodiment;

[0048] Figure 7 This is a line chart of the performance of the target detection model on the training set and validation set during the training process;

[0049] Figure 8 A schematic diagram of an interface for displaying recognition results of an interactive device provided in an embodiment. DETAILED DESCRIPTION

[0050] In order to better explain the present application, in order to facilitate understanding, the present application is described in detail by specific embodiments in combination with the drawings.

[0051] Embodiment one

[0052] As Figure 1 shown, the embodiment provides a landing gear safety pin and pitot tube sleeve recovery integrity identification method, comprising steps A1 and A2, as follows:

[0053] A1, acquiring real-time images inside a tool box; the tool box is used to place a plurality of safety pins and a plurality of pitot tube sleeves recovered from a landing gear, and the tool box is provided with a plurality of first regions for placing safety pins and a plurality of second regions for placing pitot tube sleeves.

[0054] A2, based on a target detection model, identifying whether each first region in the real-time image exists a safety pin, and whether each second region exists a pitot tube sleeve, if yes, it is determined that all safety pins and pitot tube sleeves have been completely recovered; wherein the target detection model is a deep learning model with adaptive weight parameters obtained through a pre-training process.

[0055] The landing gear safety pin and pitot tube sleeve recovery integrity identification method provided by the embodiment, after acquiring real-time images inside a tool box, based on a target detection model, according to the first regions for placing safety pins and the second regions for placing pitot tube sleeves inside the tool box, identifies whether each first region in the real-time image exists a safety pin, and whether each second region exists a pitot tube sleeve, and determines that all safety pins and pitot tube sleeves have been completely recovered. Compared with the prior art, after acquiring real-time images, the remaining review work can be completed using the target detection model, effectively reducing the amount of manual work, and identifying the recovery integrity of the landing gear safety pin and pitot tube sleeve in real time and accurately, ensuring that the safety pin and pitot tube sleeve are removed and recovered into the tool box before the aircraft takes off, and ensuring the safe flight of the aircraft.

[0056] In a preferred embodiment of the present embodiment, in order to further improve the identification speed and reduce the requirement for the computing power of the equipment, the target detection model is obtained by replacing the convolution module, CSPLayer_2Conv module and SPPF module in the backbone network of the YOLOv8n neural network model with a reverse residual module; wherein the reverse residual module is a module in the MobileNetV3 neural network model.

[0057] The YOLOv8n neural network model itself has a good speed advantage in the field of image detection and has better detection performance in application scenarios with high real-time requirements. It usually includes a backbone network (backbone), a neck network (neck) and a head network (head). The structure of the existing YOLOv8n neural network model is as follows: Figure 2 shown.

[0058] This embodiment improves the backbone network of the existing YOLOv8n neural network model, such as Figure 3 As shown in Figure 1, the inverted residual module is used to replace the convolution module (ConvModule), CSPLayer_2Conv module and SPPF module in the backbone network of the YOLOv8n neural network model.

[0059] The structure of the inverted residual module is as follows Figure 4 As shown, it is a key component in the design of lightweight convolutional neural networks. It is innovatively adjusted based on the traditional residual block to adapt to environments with limited computing resources. In the standard residual structure, the input feature map undergoes a sequence of convolutions and is added to the original input to produce an output; the inverted residual adopts a different strategy: first, the number of channels of the input feature map is expanded using an expanded convolution with a small number of output channels to enrich the feature representation, and then a depthwise separable convolution is used. This operation is performed in two steps - depthwise convolution reduces the computational burden and is implemented through channel-by-channel convolution; the subsequent point convolution not only restores the number of channels, but also embeds nonlinear characteristics, enhancing the expressive power of the model. This design exquisitely balances computational efficiency and feature extraction capabilities, and is particularly effective in lightweight networks. It can effectively reduce the computing power requirements of the device and improve the recognition speed of the target detection model.

[0060] Specifically, the inverted residual module includes a pointwise convolution submodule, a depthwise separable convolution submodule, and a residual connection submodule. The working principle of the inverted residual module is as follows:

[0061] Point-by-point convolution submodule: First, the input feature map is transformed through a 1×1 convolution layer. Usually this step increases the number of channels. This is the origin of the word "inverted" because it is the opposite of the traditional convolution operation that reduces the number of channels.

[0062] Depthwise Convolution: Next, a filter is applied independently to each input channel through depthwise convolution, which greatly reduces the number of parameters. Subsequently, the channels are fused again through pointwise convolution, usually reducing the number of channels so that the number of output channels is equal to or close to the number of input channels.

[0063] Residual connection submodule: After the depthwise separable convolution submodule, if the input and output have the same shape, the input can be directly added to the output through the residual connection submodule, which helps the propagation of gradients and also enhances the expressiveness of the object detection model.

[0064] In addition, the activation function in the inverted residual module also adopts the h-swish activation function, which enables the target detection model to provide better performance while maintaining computational efficiency; the expansion rate of the inverted residual module is no longer fixed, but can be adjusted according to needs, which increases the flexibility and efficiency of the model architecture; in addition, the SE module is introduced between the above sub-modules, which is an attention mechanism that can adaptively recalibrate channel weights, thereby enhancing the expressiveness of the model.

[0065] In contrast, the classic ResNet residual architecture first reduces the number of channels through 1×1 convolutions, then performs 3×3 convolutions, and finally expands the channels through 1×1 convolutions again, creating a "bottleneck" structure. The inverted residual approach, however, reverses this approach, initially increasing the number of channels with 1×1 convolutions, then applying 3×3 depthwise convolutions (DW convolutions) for efficient feature extraction, and finally reducing the number of channels back to their initial size. This series of operations not only improves model efficiency but also maintains strong feature learning capabilities.

[0066] More preferably, the input end of the backbone network of the target detection model is the Conv_BN_hswish module.

[0067] The Conv_BN_hswish module includes a common convolution unit, a batch normalization unit and an h-swish activation function unit connected sequentially.

[0068] The activation function in the h-swish activation function unit is:

[0069]

[0070] Wherein, x represents the pixel value of each pixel in the real-time image, ReLU6=(ReLU,6), ReLU=max(0,x).

[0071] Through the above-mentioned improvements to the backbone network of the YOLOv8n neural network model, the resulting target detection model is smaller in size, requires lower computing power resources for the recognition equipment it uses, and has better adaptability to the application environment.

[0072] Example 2

[0073] Based on the target detection model provided in Example 1, this example specifically describes the pre-training project of the target detection model.

[0074] In a specific implementation of this embodiment, the number of the first areas of the toolbox is three, and the number of the second areas is three.

[0075] like Figure 5 As shown, the pre-training process includes steps S1 to S4:

[0076] S1. Obtain training images of 64 placement methods.

[0077] In the initial training image, a total of 64 placement methods are included according to whether there is a safety pin in each first area and whether there is a pitot tube sleeve in each second area.

[0078] Specifically, during pre-flight preparation for each aircraft, maintenance personnel must remove the landing gear safety pins and pitot tube sleeves and place them in a toolbox. Failure to remove the safety pins or pitot tube sleeves can result in serious errors after takeoff. Each aircraft has three landing gear safety pins and three pitot tube sleeves, and there are 64 possible permutations and combinations of these placements. Mobile phones were used to capture images of the landing gear safety pins and pitot tube sleeves in various placements, from multiple angles and directions, with varying degrees of brightness. A total of 8,612 images were collected.

[0079] Ideally, to better reflect real-world scenarios, additional images taken by maintenance personnel on-site can be collected as initial training images. These images typically show toolboxes, safety pins, and pitot tube sleeves in a more worn-out state, with potentially irregular placement. This enriches the dataset's diversity, ensuring better generalization of the resulting object detection model in real-world application scenarios.

[0080] S2. Perform labeling and enhancement processing on the safety pin and the pitot tube sleeve in the initial training image to obtain a final training image.

[0081] Specifically, the enhancement processing includes: cropping, rotation, translation, brightness change, noise introduction, HSV enhancement, cutout, flipping, motion blur introduction and speckle noise introduction. Among them, cutout refers to randomly cutting out part of the area in the sample and filling it with 0 pixel value, and the classification result remains unchanged. The above-mentioned cropping, rotation, translation, brightness change, noise introduction, HSV enhancement, Cutout and flipping can be implemented using the OpenCV function package, and the introduction of motion blur and speckle noise can be implemented using the Imgaug function package. After data enhancement, there are a total of 13,680 images. Some of the data-enhanced training images are as follows Figure 6 shown

[0082] The labeling process includes: using labelme software to label the safety pin and pitot tube sleeve in the initial training image and generate a data file in JSON format. Specifically, the safety pin in the training image can be labeled as "aqx" and the pitot tube sleeve in the image can be labeled as "gt".

[0083] More specifically, a Python script can be used to convert the above-mentioned JSON format data file into an XML format label file, and then the XML format label file can be converted into a TXT format label file using code, where the TXT format label file mainly includes data of five items: "class", "x_center", "y_center", "width" and "heighth".

[0084] S3. Among the 13,680 final training images, the final training images in which all safety pins and pitot tube sleeves are present are selected as the positive example set, and the remaining final training images are selected as the negative example set. The positive example set and the negative example set together constitute the training data set.

[0085] S4. Use the training data set to train the initial target detection model to obtain a target detection model with adapted weight parameters.

[0086] Specifically, S4 includes sub-steps S401 to S403:

[0087] S401. Randomly divide the training dataset into a training set, a validation set, and a test set in a ratio of 8:1:1, and use Pycharm to configure different categories of dataset yaml files.

[0088] S402: Use the training set to train the initial target detection model. After each round of training, use the validation set to validate and update the model parameters of the current target detection model.

[0089] S403: End the current training process after 300 rounds of training.

[0090] S404, determining whether the performance indicator reaches a preset value;

[0091] If not, jump back to S401 to continue training;

[0092] If yes, then stop training and use the current target detection model as the target detection model with adapted weight parameters. In fact, the adapted weight parameters can be exported as a best.pt weight file to facilitate migration to the recognition device together with the target detection model.

[0093] In S404, the performance indicators include:

[0094]

[0095] Where P represents precision, TP represents the number of images correctly detected in the positive set, FP represents the number of images not correctly detected in the positive set, FN represents the number of images incorrectly detected as positive in the negative set, R represents the recall rate, P(R) represents the relationship curve between precision P and recall rate R, AP represents average precision, mAP represents the average value of average precision AP of all placement classifications, n represents the classification category, AP i represents the average precision of the i-th classification.

[0096] Figure 7 It is a performance line chart of the target detection model on the training set and the validation set during the training process from S401 to S403 above.

[0097] Based on the above training process, this embodiment also trains and compares the performance of the existing YOLOv8n neural network model and the target detection model provided in Example 1. The initial learning rate is set to 0.01, the final learning rate is set to 0.01, the batch (batch size, that is, the number of training images input into the target detection model in each round of training) is 16, and the epoch (number of cycles, that is, the number of training rounds) is 300. The configuration of this experimental environment is shown in Table 1, and the results obtained by training the YOLOv8n neural network model are shown in Table 2. The results obtained by training the target detection model provided in Example 1 are shown in Table 3. The comparison of the results of the target detection model proposed in this embodiment and the YOLOv8n neural network model in terms of model parameter quantity, model computational complexity and model size is shown in Table 4.

[0098] Table 1 Experimental environment configuration table

[0099]

[0100]

[0101] Table 2 YOLOv8 network model training result table

[0102]

[0103] Table 3 Training result table of target detection model provided by example one

[0104]

[0105] Table 4 Result comparison

[0106]

[0107] From Table 2 and Table 3, it is found that the target detection model provided by example one and the YOLOv8n neural network model have little difference in the numerical changes of the indicators of precision, recall rate, mAP50 and mAP50-95. From Table 4, it can be found that after using the inverted residual module in MobileNetV3 to replace the backbone network of the YOLOv8n model, the model parameter quantity, GFLOPs (GigaFloating-point Operations Per Second, 10 billion floating-point operations per second) and model size are significantly reduced, the parameter quantity is reduced by 60.47%, the calculation amount of the model is reduced by 65.43%, and the model size is reduced by 57.45%, realizing the lightweight of the landing gear safety pin and air speed tube sleeve detection model.

[0108] Example three

[0109] Based on the recognition methods provided in examples one and two, this embodiment provides a landing gear safety pin and air speed tube sleeve recovery integrity recognition device, which comprises a processing device and an interactive device.

[0110] The processing device comprises a memory, a processor and a computer program stored on the memory and executable on the processor, and the computer program implements the landing gear safety pin and air speed tube sleeve recovery integrity recognition method of the first aspect when executed by the processor.

[0111] Specifically, in addition to including "whether all safety pins and air speed sleeves have been recovered completely", the recognition result of the processing device can also include the actual number of safety pins, the actual number of air speed sleeves and other information, so as to provide more reference information for the user and facilitate the user to further statistically process or process the data.

[0112] The interactive device is used to interact with the user, obtain the real-time image uploaded by the user, send the real-time image to the processing device to recognize the recovery integrity of the landing gear safety pin and the air speed tube sleeve in the real-time image and obtain the recognition result, and receive and show the recognition result returned by the processing device to the user.

[0113] In a specific implementation, an interactive terminal is installed on the interactive device, and the interactive terminal displays the interface of the recognition result as follows: Figure 8 As shown, in addition to marking each safety pin and pitot tube sleeve in the real-time image, the actual number of safety pins and the actual number of pitot tube sleeves are also displayed for users to review or perform other post-processing steps.

[0114] Preferably, to enhance the portability of the interactive device, the interactive device can be a mobile phone. Maintenance personnel can use the phone to take a photo, then upload it to the processing device for recognition via the interactive terminal. The interactive terminal then receives the recognition results returned by the processing device and displays them to the user. Furthermore, by using the mobile phone as the interactive device, the real-time images captured by the target detection model, which align with the data style of the training dataset used by the target detection model, can be used to further ensure the recognition accuracy of the target detection model.

[0115] In the description of the present invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.

[0116] In the present invention, unless otherwise expressly specified or limited, the terms "mounted," "connected," "connect," "fixed," etc. should be understood broadly. For example, they may refer to fixed connection, detachable connection, or integration; mechanical connection or electrical connection; direct connection or indirect connection through an intermediate medium; and internal communication between two components or interaction between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0117] In the present invention, unless otherwise expressly specified or limited, when a first feature is "above" or "below" a second feature, it may mean that the first and second features are in direct contact, or that the first and second features are in indirect contact through an intermediate medium. Furthermore, when a first feature is "above," "above," or "above" a second feature, it may mean that the first feature is directly above or obliquely above the second feature, or simply means that the first feature is at a higher level than the second feature. When a first feature is "below," "below," or "below" a second feature, it may mean that the first feature is directly below or obliquely below the second feature, or simply means that the first feature is at a lower level than the second feature.

[0118] In the description of the specification, the description of the terms "one embodiment", "some embodiments", "an embodiment", "example", "specific example" or "some examples" etc. means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are contained in at least one embodiment or example of the present application. In the specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. In addition, the person skilled in the art can combine and combine the different embodiments or examples described in the specification and the features of the different embodiments or examples without contradiction.

[0119] Although the embodiments of the present application have been shown and described above, it is understood that the above-described embodiments are exemplary and are not to be construed as limiting the present application, and the person skilled in the art can modify, modify, replace and modify the above-described embodiments within the scope of the present application.

Claims

1. A method for recognizing the recovery integrity of a landing gear safety pin and pitot tube sleeve, characterized in that: include: Get a real-time image of the toolbox interior; The tool box is used to store a plurality of safety pins and a plurality of pitot tube sleeves recovered from the landing gear, and the tool box is provided with a plurality of first areas designated for storing safety pins and a second area designated for storing pitot tube sleeves; Based on the target detection model, identifying whether a safety pin is present in each first region of the real-time image and whether a pitot tube sleeve is present in each second region, and if so, determining that all safety pins and pitot tube sleeves are intact; The target detection model is a deep learning model with adaptive weight parameters obtained through a pre-training process; The target detection model is obtained by replacing the convolution module, CSPLayer_2Conv module and SPPF module in the backbone network of the YOLOv8n neural network model with the inverted residual module; The inverted residual module is a module in the MobileNetV3 neural network model; The number of the first regions is three, the number of the second regions is three, and the pre-training process includes: S1. Obtain at least 64 initial training images; In the initial training image, 64 placement methods are classified according to whether there is a safety pin in each first area and whether there is a pitot tube sleeve in each second area; S2. performing labeling and enhancement processing on the safety pin and the pitot tube sleeve in the initial training image to obtain a final training image; S3. Among all the final training images, select the final training images in which all the safety pins and pitot tube sleeves are present as a positive example set, and the remaining final training images as a negative example set. The positive example set and the negative example set together constitute a training data set; S4. Using the training data set to train the initial target detection model to obtain a target detection model with adapted weight parameters; The initial learning rate of the pre-training process is 0.01, the batch size is 16, and the epoch size is 300.

2. The identification method according to claim 1, characterized in that The input end of the backbone network of the target detection model is the Conv_BN_hswish module; The Conv_BN_hswish module includes a common convolution unit, a batch normalization unit and an h-swish activation function unit connected sequentially.

3. The identification method according to claim 2, characterized in that The activation function in the h-swish activation function unit is: Wherein, x represents the pixel value of each pixel in the real-time image, ReLU6=(ReLU,6), ReLU=max(0,x).

4. The identification method according to claim 1, wherein: Said S1 comprises: According to multiple shooting angles, shooting directions and shooting brightness, a camera device is used to obtain initial training images based on 64 placement methods.

5. The identification method according to claim 1, characterized in that In said S2, The labeling process includes: labeling the safety pin and pitot tube sleeve in the initial training image using labelme software to generate a data file in JSON format; The enhancement processing includes: cropping, rotation, translation, brightness change, noise introduction, HSV enhancement, cutout, flipping, motion blur introduction and speckle noise introduction.

6. The identification method according to claim 1, characterized in that The S4 includes: The training data set is divided into a training set, a validation set, and a test set in a ratio of 8:1:1; The initial target detection model is trained using the training set, and after each round of training, the model parameters of the current target detection model are verified and updated using the validation set; End the current training process after 300 rounds of training; Determine whether the performance index reaches the preset value. If so, use the current target detection model as the target detection model with the adapted weight parameters. The performance indicators include: Where P represents precision, TP represents the number of images correctly detected in the positive set, FP represents the number of images not correctly detected in the positive set, FN represents the number of images incorrectly detected as positive in the negative set, R represents the recall rate, P(R) represents the relationship curve between precision P and recall rate R, AP represents average precision, mAP represents the average value of average precision AP of all placement classifications, n represents the classification category, AP i represents the average precision of the i-th classification.

7. A landing gear safety pin and pitot tube recovery integrity identification device, characterized in that: including processing equipment; The processing device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, the method for identifying the recovery integrity of a landing gear safety pin and a pitot tube sleeve according to any one of claims 1 to 6 is implemented.

8. The identification device according to claim 7, characterized in that It also includes interactive devices; The interactive device is used to interact with the user, obtain real-time images uploaded by the user, send the real-time images to the processing device to identify the recovery integrity of the landing gear safety pin and pitot tube sleeve in the real-time image and obtain an identification result; and receive and display the identification result returned by the processing device to the user.

Images