Marine parking platform dynamic position prediction method based on Bayesian deep learning

Abstract

The invention provides an offshore parking platform dynamic position prediction method based on Bayesian deep learning, relates to the technical field of dynamic prediction of the dynamic position of an offshore parking platform by an unmanned aerial vehicle, and solves the technical problem of dynamic prediction of the position of the offshore parking platform by the unmanned aerial vehicle in a complex environment. According to the technical scheme, the method comprises the following steps: S1, establishing a marine unmanned aerial vehicle parking platform image data set; s2, establishing an offshore parking platform dynamic position prediction model based on Bayesian deep learning; s3, acquiring a dynamic position prediction model of the offshore parking platform; and S4, obtaining the position and uncertainty of the offshore parking platform relative to the unmanned aerial vehicle. The method has the beneficial effects that the influence of factors such as height, position, angle, sea wind, sea wave, jitter and the like in the landing process of the unmanned aerial vehicle on the sea surface is solved, the dynamic position of the offshore parking platform is accurately predicted, the prediction uncertainty is given, and the robustness of the model is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle (UAV) dynamic prediction of the dynamic position of an offshore parking platform, and in particular to a method for predicting the dynamic position of an offshore parking platform based on Bayesian deep learning. Background Art

[0002] With the rapid development of drone technology and the gradual move of drone applications to the deep sea, drone landing has become the final step for drones to complete various tasks at sea. The prediction of the dynamic position of the offshore landing platform is crucial for drone landing. The dynamic position prediction of the offshore landing platform involves complex marine environments and changeable target forms, including ships and platforms. It requires high-precision sensors, such as high-definition cameras, height sensors, etc. It also requires a stable and reliable offshore drone landing platform dynamic position prediction algorithm to analyze the various information obtained by each sensor.

[0003] CN116894936A provides a method and system for identifying and locating offshore targets based on drone vision, which can realize the positioning of offshore targets, but the sensors involved are complex and the impact of the complex offshore environment is not considered. The present invention can well realize the dynamic prediction of the position of the offshore parking platform in a complex sea surface environment. Summary of the invention

[0004] The purpose of the present invention is to provide a method for predicting the dynamic position of an offshore parking platform based on Bayesian deep learning. The method can more accurately obtain the position of the offshore parking platform relative to the UAV, and give the uncertainty of its position, thereby improving the robustness of the neural network.

[0005] To achieve the above object, the present invention adopts the following technical solution: a method for predicting the dynamic position of an offshore shutdown platform based on Bayesian deep learning, comprising the following steps:

[0006] S1. Establish an image dataset of offshore drone parking platforms;

[0007] S2. Establish a dynamic position prediction model for offshore parking platforms based on Bayesian deep learning;

[0008] S3. Use the constructed data set to perform model training to obtain a dynamic position prediction model for the offshore parking platform;

[0009] S4. Analyze and process the image of the UAV during landing according to the trained dynamic position prediction model of the offshore parking platform to obtain the position and uncertainty of the offshore parking platform relative to the UAV;

[0010] Preferably, the step S1 specifically includes the following steps:

[0011] S11: Collect environmental images of offshore drone parking platforms, including T-shaped parking platforms, H-shaped parking platforms, concentric circular parking platforms, rectangular parking platforms, and offshore unmanned boat platforms.

[0012] S12: In order to simulate the various unfavorable factors encountered by drones during the recognition process, the collected images are rotated to different degrees and uncertainty noise is randomly added to improve the image gradient and highlight the main features.

[0013] S13: Mark and save the offshore parking platform area in the form of (v i ,u i ,d i ), where v i is the candidate region ROI,u i is the pixel coordinate of the ROI midpoint relative to the image center point, in the form of (x i ,y i ), d i The ratio of the ROI area to the entire image is . The original images and processed images are randomly combined as the offshore parking platform image dataset.

[0014] Preferably, the dynamic position prediction model of the offshore parking platform based on Bayesian deep learning described in step S2 consists of 1 input layer, 3 convolutional pooling layers, 2 convolutional layers, and 1 output layer. Each convolutional layer is followed by Rule as an activation function, and the result of the output layer is passed to the SSoftmax layer to obtain the dynamic position prediction model of the offshore parking platform. Its prior probability is randomly obtained using a Gaussian distribution, and the posterior probability is approximated using the MCO_Dropout regularization method. The specific method is: add a Dropout before each layer, and then sample the posterior probability of the weight. In this way, the posterior distribution of the probability of the SSoftmax layer is obtained.

[0015] The goal is to minimize the loss function of the Bayesian deep convolutional network:

[0016]

[0017] D represents the number of samples, and i is the training sample number. The above loss consists of two parts: one is the residual of the model, which is used to capture the cognitive uncertainty of the model parameters; the other is the accidental uncertainty of the prediction, which acts as a regularization term. After training, the following output results are obtained:

[0018] in To predict the pixel coordinate of the parking platform relative to the midpoint of the image, To predict the pixel ordinate of the parking platform relative to the midpoint of the image, Indicates the percentage of the downtime platform in the image size. represents the uncertainty of the predicted position. Reflects the height of the parking platform from the drone, from which the distance between the parking platform and the drone can be estimated as: K is the distance between the parking platform and the drone when the parking platform occupies the entire image. Represents the weight parameter.

[0019] Combine the above values ​​into As output, the predicted position of the offshore landing platform relative to the UAV and uncertainty

[0020] Preferably, the model training in step S3 specifically includes the following steps:

[0021] S31: Define a dynamic position prediction model for offshore parking platforms based on Bayesian deep learning, including convolutional layers, pooling layers, fully connected layers, and add a Dropout layer to introduce randomness.

[0022] S32: In each training iteration, the Dropout layer is turned on to randomly discard some neurons, and the Monte Carlo algorithm is used for multiple forward propagations to calculate the output under different sampling conditions.

[0023] S33: Calculate the loss function according to formula (1) and update the model parameters.

[0024] S34: Terminate the training process according to the number of training rounds or convergence conditions.

[0025] Preferably, the model detection in step S4 specifically includes the following steps:

[0026] S41: Collecting pictures of the sea surface taken by drone.

[0027] S42: The image data passes through multiple convolutional layers and maximum pooling layers, and finally obtains the prediction result through the SSoftmax layer.

[0028] S43: Perform multiple forward propagations through the Monte Carlo sampling method to obtain multiple prediction results.

[0029] S44: Take the average of multiple predictions to form the final prediction result And calculate the variance of multiple forecasts to estimate forecast uncertainty.

[0030] The specific process is: input the drone shooting images captured by the camera into the model trained by S3 to obtain the probability distribution of the position of the offshore parking platform relative to the drone, providing data support for the next movement of the drone.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] 1. The present invention collects a large number of images of offshore UAV parking platforms, randomly rotates the original images, and adds uncertainty noise to form a mixed image data set, thereby improving image gradients and highlighting trunk features. It can simulate the images of UAV parking platforms taken by UAVs under complex sea surfaces, thereby improving the authenticity and reliability of training data and ensuring the accuracy of detection.

[0033] 2. In the real sea environment, due to the complexity and changeability of the sea environment, the collected data generally has uncertainty. The recognition method proposed in the present invention can well solve the influence of factors such as height, position, angle, sea breeze, waves, and jitter during the landing of the UAV on the sea surface, and realize a more accurate position prediction of the offshore parking platform, and give the uncertainty of the prediction, which is not available in traditional neural networks. It helps to alleviate the problem of overconfidence of deep neural networks and improve the robustness of the model.

[0034] 3. The present invention introduces the distance prediction of the offshore parking platform of the UAV relative to the UAV. The UAV infers the distance of the offshore parking platform relative to the UAV through the relative size of the offshore parking platform, provides a three-dimensional position prediction of the parking platform, and provides more accurate data support for the next movement of the UAV. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 Schematic diagram of the working of the present invention.

[0036] Figure 2 Flowchart of the present invention.

[0037] Figure 3 Schematic diagram of the drone randomly rotating to take images and randomly adding uncertain noise in the present invention.

[0038] Figure 4 Model training flow chart in the present invention.

[0039] Figure 5 Schematic diagram of the parking platform position prediction in the present invention.

[0040] Figure 6 Example 2 Prediction effect diagram

[0041] Figure 7 Example 3 Prediction effect diagram DETAILED DESCRIPTION

[0042] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0043] Example 1

[0044] Reference Figure 2 This embodiment provides a method for predicting the dynamic position of an offshore parking platform based on Bayesian deep learning, comprising the following steps:

[0045] Step 1: Establish an image dataset of offshore UAV parking platforms;

[0046] Step 2: Establish a dynamic position prediction model for offshore parking platforms based on Bayesian deep learning;

[0047] Step 3: Use the constructed data set to perform model training to obtain a dynamic position prediction model for the offshore parking platform;

[0048] Step 4: Analyze and process the images of the UAV during landing according to the trained dynamic position prediction model of the offshore parking platform to obtain the position and uncertainty of the offshore parking platform relative to the UAV;

[0049] Step 1: Establish an offshore drone parking platform image dataset, which includes three steps:

[0050] 1-1) Collect environmental images of the arrow-shaped drone parking platform at sea.

[0051] 1-2) In order to simulate the various unfavorable factors encountered by drones during the recognition process, the collected images are rotated to different degrees and uncertainty noise is randomly added to improve the image gradient and highlight the main features.

[0052] 1-3) Mark and save the offshore parking platform area in the form of (v i ,u i ,d i ), where v i is the candidate region ROI,u i is the pixel coordinate of the ROI midpoint relative to the image center point, in the form of (x i ,y i ), d i The ratio of the ROI area to the entire image is . The original images and processed images are randomly combined as the offshore parking platform image dataset.

[0053] Step 2: Establish a dynamic position prediction model for offshore parking platforms based on Bayesian deep learning. The dynamic position prediction model for offshore parking platforms based on Bayesian deep learning consists of 1 input layer, 3 convolutional pooling layers, 2 convolutional layers, and 1 output layer. Each convolutional layer is followed by Rule as an activation function, and the results of the output layer are passed to the SSoftmax layer to obtain the dynamic position prediction model for offshore parking platforms.

[0054] Among them, the prior probability is randomly obtained using Gaussian distribution, and the posterior probability is approximated using MCO_Dropout regularization method. The specific method is: add a Dropout before each layer, and then sample the posterior probability of the weight. In this way, the posterior distribution of the probability of the SSoftmax layer is obtained.

[0055] The goal is to minimize the loss function of the Bayesian deep convolutional network:

[0056]

[0057] D represents the number of samples, and i is the training sample number. The above loss consists of two parts: one is the residual of the model, which is used to capture the cognitive uncertainty of the model parameters; the other is the accidental uncertainty of the prediction, which acts as a regularization term. After training, the following output results are obtained:

[0058] in To predict the pixel coordinate of the parking platform relative to the midpoint of the image, To predict the pixel ordinate of the parking platform relative to the midpoint of the image, Indicates the percentage of the downtime platform in the image size. represents the uncertainty of the predicted position. Reflects the height of the parking platform from the drone, from which the distance between the parking platform and the drone can be estimated as: K is the distance between the parking platform and the drone when the parking platform occupies the entire image. Represents the weight parameter.

[0059] Combine the above values ​​into As output, the predicted position of the offshore landing platform relative to the UAV and uncertainty

[0060] Step 3: Train the Bayesian deep learning model, such as Figure 3 As shown, the process training process is:

[0061] 3-1) Define a Bayesian deep convolutional network, including convolutional layers, pooling layers, and fully connected layers, and add a Dropout layer to introduce randomness.

[0062] 3-2) In each training iteration, the Dropout layer is turned on to randomly discard some neurons, and the Monte Carlo algorithm is used to perform multiple forward propagations to calculate the output under different sampling conditions.

[0063] 3-3) Calculate the loss function according to formula (1) and update the model parameters.

[0064] 3-4) Terminate the training process according to the number of training rounds or convergence conditions.

[0065] Step 4: Analyze and process the images of the drone during landing according to the trained dynamic position prediction model of the offshore parking platform to obtain the dynamic position and uncertainty of the offshore parking platform. The prediction effect is as follows: Figure 5 , which comprises the following steps:

[0066] 4-1) Use drones to collect images of the sea surface.

[0067] 4-2) The image data passes through multiple convolutional layers and maximum pooling layers, and finally passes through the SSoftmax layer to obtain the prediction result.

[0068] 4-3) Perform multiple forward propagations through the Monte Carlo sampling method to obtain multiple prediction results.

[0069] 4-4) Take the average of multiple predictions to form the final prediction result And calculate the variance of multiple forecasts to estimate forecast uncertainty.

[0070] The specific process is: input the collected drone shooting images into the model trained by S3 to obtain the position probability distribution of the offshore parking platform relative to the drone, providing data support for the next movement of the drone.

[0071] Example 2

[0072] Reference Figure 2 This embodiment provides a method for predicting the dynamic position of an offshore parking platform based on Bayesian deep learning, comprising the following steps:

[0073] Step 1: Establish an image dataset of offshore UAV parking platforms;

[0074] Step 2: Establish a dynamic position prediction model for offshore parking platforms based on Bayesian deep learning;

[0075] Step 3: Use the constructed data set to perform model training to obtain a dynamic position prediction model for the offshore parking platform;

[0076] Step 4: Analyze and process the images of the UAV during landing according to the trained dynamic position prediction model of the offshore parking platform to obtain the position and uncertainty of the offshore parking platform relative to the UAV;

[0077] Step 1: Establish an offshore drone parking platform image dataset, which includes three steps:

[0078] 1-1) Collect environmental images of the offshore X-shaped drone parking platform.

[0079] 1-2) In order to simulate the various unfavorable factors encountered by drones during the recognition process, the collected images are rotated to different degrees and uncertainty noise is randomly added to improve the image gradient and highlight the main features.

[0080] 1-3) Mark and save the offshore parking platform area in the form of (v i,u i ,d i ), where v i is the candidate region ROI,u i is the pixel coordinate of the ROI midpoint relative to the image center point, in the form of (x i ,y i ), d i The ratio of the ROI area to the entire image is . The original images and processed images are randomly combined as the offshore parking platform image dataset.

[0081] Step 2: Establish a dynamic position prediction model for offshore parking platforms based on Bayesian deep learning. The dynamic position prediction model for offshore parking platforms based on Bayesian deep learning consists of 1 input layer, 3 convolutional pooling layers, 2 convolutional layers, and 1 output layer. Each convolutional layer is followed by Rule as an activation function, and the results of the output layer are passed to the SSoftmax layer to obtain the dynamic position prediction model for offshore parking platforms.

[0082] Among them, the prior probability is randomly obtained using Gaussian distribution, and the posterior probability is approximated using MCO_Dropout regularization method. The specific method is: add a Dropout before each layer, and then sample the posterior probability of the weight. In this way, the posterior distribution of the probability of the SSoftmax layer is obtained.

[0083] The goal is to minimize the loss function of the Bayesian deep convolutional network:

[0084]

[0085] D represents the number of samples, and i is the training sample number. The above loss consists of two parts: one is the residual of the model, which is used to capture the cognitive uncertainty of the model parameters; the other is the accidental uncertainty of the prediction, which acts as a regularization term. After training, the following output results are obtained:

[0086] in To predict the pixel coordinate of the parking platform relative to the midpoint of the image, To predict the pixel ordinate of the parking platform relative to the midpoint of the image, Indicates the percentage of the downtime platform in the image size. represents the uncertainty of the predicted position. Reflects the height of the parking platform from the drone, from which the distance between the parking platform and the drone can be estimated as: K is the distance between the parking platform and the drone when the parking platform occupies the entire image. Represents the weight parameter.

[0087] Combine the above values ​​into As output, the predicted position of the offshore landing platform relative to the UAV and uncertainty

[0088] Step 3: Train the Bayesian deep learning model, such as Figure 3 As shown, the process training process is:

[0089] 3-1) Define a Bayesian deep convolutional network, including convolutional layers, pooling layers, and fully connected layers, and add a Dropout layer to introduce randomness.

[0090] 3-2) In each training iteration, the Dropout layer is turned on to randomly discard some neurons, and the Monte Carlo algorithm is used to perform multiple forward propagations to calculate the output under different sampling conditions.

[0091] 3-3) Calculate the loss function according to formula (1) and update the model parameters.

[0092] 3-4) Terminate the training process according to the number of training rounds or convergence conditions.

[0093] Step 4: Analyze and process the images of the drone during landing according to the trained dynamic position prediction model of the offshore parking platform to obtain the dynamic position and uncertainty of the offshore parking platform. The prediction effect is as follows: Figure 5 , which comprises the following steps:

[0094] 4-1) Use drones to collect images of the sea surface.

[0095] 4-2) The image data passes through multiple convolutional layers and maximum pooling layers, and finally passes through the SSoftmax layer to obtain the prediction result.

[0096] 4-3) Perform multiple forward propagations through the Monte Carlo sampling method to obtain multiple prediction results.

[0097] 4-4) Take the average of multiple predictions to form the final prediction result And calculate the variance of multiple forecasts to estimate forecast uncertainty.

[0098] The specific process is: input the collected drone shooting images into the model trained by S3 to obtain the position probability distribution of the offshore parking platform relative to the drone, providing data support for the next movement of the drone.

[0099] Example 3

[0100] Reference Figure 2 This embodiment provides a method for predicting the dynamic position of an offshore parking platform based on Bayesian deep learning, comprising the following steps:

[0101] Step 1: Establish an image dataset of offshore UAV parking platforms;

[0102] Step 2: Establish a dynamic position prediction model for offshore parking platforms based on Bayesian deep learning;

[0103] Step 3: Use the constructed data set to perform model training to obtain a dynamic position prediction model for the offshore parking platform;

[0104] Step 4: Analyze and process the images of the UAV during landing according to the trained dynamic position prediction model of the offshore parking platform to obtain the position and uncertainty of the offshore parking platform relative to the UAV;

[0105] Step 1: Establish an offshore drone parking platform image dataset, which includes three steps:

[0106] 1-1) Collect environmental images of the unmanned boat-shaped drone parking platform at sea.

[0107] 1-2) In order to simulate the various unfavorable factors encountered by drones during the recognition process, the collected images are rotated to different degrees and uncertainty noise is randomly added to improve the image gradient and highlight the main features.

[0108] 1-3) Mark and save the offshore parking platform area in the form of (v i ,u i ,d i ), where v i is the candidate region ROI,u i is the pixel coordinate of the ROI midpoint relative to the image center point, in the form of (x i ,y i ), d i The ratio of the ROI area to the entire image is . The original images and processed images are randomly combined as the offshore parking platform image dataset.

[0109] Step 2: Establish a dynamic position prediction model for offshore parking platforms based on Bayesian deep learning. The dynamic position prediction model for offshore parking platforms based on Bayesian deep learning consists of 1 input layer, 3 convolutional pooling layers, 2 convolutional layers, and 1 output layer. Each convolutional layer is followed by Rule as an activation function, and the results of the output layer are passed to the SSoftmax layer to obtain the dynamic position prediction model for offshore parking platforms.

[0110] Among them, the prior probability is randomly obtained using Gaussian distribution, and the posterior probability is approximated using MCO_Dropout regularization method. The specific method is: add a Dropout before each layer, and then sample the posterior probability of the weight. In this way, the posterior distribution of the probability of the SSoftmax layer is obtained.

[0111] The goal is to minimize the loss function of the Bayesian deep convolutional network:

[0112]

[0113] D represents the number of samples, and i is the training sample number. The above loss consists of two parts: one is the residual of the model, which is used to capture the cognitive uncertainty of the model parameters; the other is the accidental uncertainty of the prediction, which acts as a regularization term. After training, the following output results are obtained:

[0114] in To predict the pixel coordinate of the parking platform relative to the midpoint of the image, To predict the pixel ordinate of the parking platform relative to the midpoint of the image, Indicates the percentage of the downtime platform in the image size. represents the uncertainty of the predicted position. Reflects the height of the parking platform from the drone, from which the distance between the parking platform and the drone can be estimated as: K is the distance between the parking platform and the drone when the parking platform occupies the entire image. Represents the weight parameter.

[0115] Combine the above values ​​into As output, the predicted position of the offshore landing platform relative to the UAV and uncertainty

[0116] Step 3: Train the Bayesian deep learning model, such as Figure 3 As shown, the process training process is:

[0117] 3-1) Define a Bayesian deep convolutional network, including convolutional layers, pooling layers, and fully connected layers, and add a Dropout layer to introduce randomness.

[0118] 3-2) In each training iteration, the Dropout layer is turned on to randomly discard some neurons, and the Monte Carlo algorithm is used to perform multiple forward propagations to calculate the output under different sampling conditions.

[0119] 3-3) Calculate the loss function according to formula (1) and update the model parameters.

[0120] 3-4) Terminate the training process according to the number of training rounds or convergence conditions.

[0121] Step 4: Analyze and process the images of the drone during landing according to the trained dynamic position prediction model of the offshore parking platform to obtain the dynamic position and uncertainty of the offshore parking platform. The prediction effect is as follows: Figure 5 , which comprises the following steps:

[0122] 4-1) Use drones to collect images of the sea surface.

[0123] 4-2) The image data passes through multiple convolutional layers and maximum pooling layers, and finally passes through the SSoftmax layer to obtain the prediction result.

[0124] 4-3) Perform multiple forward propagations through the Monte Carlo sampling method to obtain multiple prediction results.

[0125] 4-4) Take the average of multiple predictions to form the final prediction result And calculate the variance of multiple forecasts to estimate forecast uncertainty.

[0126] The specific process is: input the collected drone shooting images into the model trained by S3 to obtain the position probability distribution of the offshore parking platform relative to the drone, providing data support for the next movement of the drone.

[0127] The description and practice disclosed in the present invention are easy to think and understand for ordinary technicians in the technical field, and several improvements and modifications can be made without departing from the principles of the present invention. Therefore, modifications or improvements made without departing from the spirit of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A method for predicting the dynamic position of an offshore parking platform based on Bayesian deep learning, characterized in that: The following steps are involved: S1. Establish an image dataset of offshore drone parking platforms; S2. Establish a dynamic position prediction model for offshore parking platforms based on Bayesian deep learning; S3. Use the constructed data set to perform model training to obtain a dynamic position prediction model for an offshore shutdown platform based on Bayesian deep learning; S4. Analyze and process the images of the UAV during landing according to the trained dynamic position prediction model of the offshore parking platform to obtain the position and uncertainty of the offshore parking platform relative to the UAV.

2. The method for predicting the dynamic position of an offshore parking platform based on Bayesian deep learning according to claim 1 is characterized in that: In the above step S1, constructing the parking platform image dataset specifically includes the following steps: S11: Collect environmental images of offshore drone parking platforms, including T-shaped parking platforms, H-shaped parking platforms, concentric circular parking platforms, rectangular parking platforms, and offshore unmanned boat platforms; S12: To simulate the various unfavorable factors encountered by drones during the recognition process, the collected images are randomly rotated to different degrees and uncertainty noise is randomly added to improve the image gradient and highlight the main features; S13: Mark and save the offshore parking platform area in the form of (v i ,u i ,d i ), where v i is the downtime platform ROI, u i is the pixel coordinate of the ROI midpoint relative to the image center point, in the form of (x i ,y i ), d i The ratio of the parking platform area to the entire image is calculated by randomly combining the original image and the processed image as the offshore parking platform image dataset.

3. The method for predicting the dynamic position of an offshore shutdown platform based on Bayesian deep learning according to claim 1 is characterized in that: In the step S2, the dynamic position prediction model of the offshore parking platform based on Bayesian deep learning is composed of 1 input layer, 3 convolutional pooling layers, 2 convolutional layers, and 1 output layer. Each convolutional layer is followed by Rule as an activation function, and the result of the output layer is passed into the double exponential normalized SSoftmax layer to obtain the dynamic position prediction model of the offshore parking platform. The prior probability is randomly obtained using a Gaussian distribution, and the posterior probability is approximated using the MCO_Dropout regularization method. The specific method is: a random discarding unit Dropout is introduced before each layer, and then the posterior probability of the weight is sampled to obtain the posterior distribution of the probability of the SSoftmax layer; The goal is to minimize the loss function of the Bayesian deep convolutional network: D represents the number of samples, i is the training sample number, and the above loss consists of two parts: one is the residual of the model, which is used to capture the cognitive uncertainty of the model parameters; the other is the accidental uncertainty of the prediction, which acts as a regularization term. After training, the following output results are obtained: in To predict the pixel coordinate of the parking platform relative to the midpoint of the image, To predict the pixel ordinate of the parking platform relative to the midpoint of the image, Indicates the percentage of the downtime platform in the image size. represents the uncertainty of the predicted position, where Reflects the height of the parking platform from the drone, and the distance between the parking platform and the drone is estimated to be: K is the distance between the parking platform and the drone when the parking platform occupies the entire image. represents the weight parameter; Combine the above values ​​into As output, the predicted position of the offshore landing platform relative to the UAV and uncertainty 4. The method for predicting the dynamic position of an offshore parking platform based on Bayesian deep learning according to claim 1 is characterized in that: In step S3, the model training specifically includes the following steps: S31: Define a dynamic position prediction model for offshore parking platforms based on Bayesian deep learning, including convolutional layers, pooling layers, and fully connected layers, and add a Dropout layer to introduce randomness; S32: In each training iteration, the Dropout layer is turned on to randomly discard some neurons, and the Monte Carlo algorithm is used to perform multiple forward propagations to calculate the output under different sampling conditions; S33: Calculate the loss function according to formula (1) and update the model parameters; S34: Terminate the training process according to the number of training rounds or convergence conditions.

5. The method for predicting the dynamic position of an offshore parking platform based on Bayesian deep learning according to claim 1 is characterized in that: In step S4, the model detection specifically includes the following steps: S41: collecting sea surface pictures taken by drones to obtain sea surface images; S42: The image data passes through multiple convolutional layers and maximum pooling layers, and finally obtains the prediction result through the SSoftmax layer; S43: Perform multiple forward propagations through the Monte Carlo sampling method to obtain multiple prediction results; S44: Take the average of multiple predictions to form the final prediction result And calculate the variance of multiple forecasts to estimate forecast uncertainty; The specific method is as follows: the collected drone shooting images are input into the model trained in step S3 to obtain the position probability distribution of the offshore parking platform relative to the drone, providing data support for the next movement of the drone.

Images