A point cloud partitioning method based on historical scheme learning

Abstract

The application belongs to the field of satellite management and control, and discloses a point group division method based on historical scheme learning, and the specific implementation steps are as follows: modeling historical scheme data; extracting point group target features, constructing label data, obtaining a training sample set and a test sample set; constructing a point group division model based on historical scheme learning; iteratively training the point group division model based on historical scheme learning; and generating a point group target planning scheme. The application learns historical experience at the meta-task level, determines a strip according to a satellite and a corresponding shooting time, and obtains a scheme for dividing point group targets. The application combines the spatial and temporal characteristics of historical point group target planning schemes, learns historical observation schemes, maps observation meta-tasks according to satellite usage and corresponding shooting start times, and thus effectively avoids the problems of forcibly dividing adjacent points into different clusters and listing edge points as separate point targets.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of satellite demand planning, in particular to a point group division method based on historical scheme learning. BACKGROUND

[0002] With the development of satellite technology, the application scenarios of satellites are also expanding. For example, real-time monitoring of multiple key targets in a key area using satellites usually requires satellites to observe a large number of unevenly distributed point targets in the area and generate image information for situation analysis.

[0003] When observing point group targets, the traditional method generates corresponding observation element tasks after analyzing each discrete point target, which can cause waste of satellite resources and low feasibility of planning schemes. Existing methods usually divide the point group targets into regions by grid division, clustering, etc., thereby converting the problem into generating satellite observation schemes for regional targets. These methods have the problem of forcibly dividing adjacent points into different regions or considering points at the edge of the grid as separate point targets due to strict grid restrictions.

[0004] At the same time, for key targets in a key area with high observation frequency, using prior knowledge contained in historical observation schemes can help simplify the generation process of new observation schemes for point group targets, while maintaining high accuracy and high executability.

[0005] To solve the above problems, a more reasonable method is needed to process point group targets to ensure that prior knowledge from historical observation schemes can be combined when planning demands to obtain point group division schemes that cover as many point targets in the point group as possible. SUMMARY

[0006] The technical problem to be solved by the present application is to avoid the shortcomings in the background art and utilize prior knowledge from historical schemes to propose a point group target division method based on historical scheme learning. In this method, first, historical scheme data is modeled, then point group data is feature extracted, label data is designed, and training sample sets and test sample sets are obtained; a network model is constructed, training data is input into the network for iterative training, and a trained network model is obtained; the test sample set is input into the trained model to generate a point group division scheme. Through this method, a point group division scheme that covers as many point targets in the point group as possible can be generated based on historical observation schemes.

[0007] The technical problems solved by the present application are mainly realized by the following technical solutions:

[0008] A point group division method based on historical scheme learning, comprising the following steps:

[0009] Step 1, modeling historical scheme data, including point group targets and meta tasks, each point group containing multiple point targets, each point target represented by longitude and latitude, and each meta task represented by a four-tuple including satellite model, start time, end time and satellite payload type;

[0010] Step 2, extracting point group data feature representation from historical scheme data, calculating discretization indicators as feature representation of the entire point group, and constructing corresponding label data to obtain training sample set I train =(X train ,Y train ) and test sample set I test =(X test ,Y test );

[0011] Step 3, constructing a point group division model H based on historical scheme learning, including a backbone network and a prediction layer, the backbone network including a first linear layer, a first convolutional layer, a Relu activation function, a max pooling layer, a single-layer unidirectional long short-term memory network and a Transformer encoding module connected in sequence, and the prediction layer including a second linear layer, a dropout layer, a Relu activation function and a third linear layer; wherein the Transformer encoding module includes four Transformer encoder layers;

[0012] Step 4, using the training sample set and the test sample set to iteratively train the point group division model H to obtain a trained point group division model H' based on historical scheme learning;

[0013] Step 5, using the test sample set as the input of the point group division model H' based on historical scheme learning to perform forward propagation to obtain the corresponding satellite usage, corresponding start time and satellite payload type for each point group, which are mapped to the point group data to obtain a meta task list and generate a point group division scheme.

[0014] Further, in step 2, N train point groups are selected, and the longitude and latitude of all point targets in each point group are calculated to represent the discretization degree indicators: range, average difference, standard deviation, mean, variance, skewness and kurtosis as the feature representation of the entire point group, obtaining the feature representation of N train point groups Meanwhile, N test point groups are selected, and the feature representation of N test point groups is calculated wherein N train ≥2000, N train and N test point groups are constructed to obtain corresponding label data Wherein, M represents the number of satellites, the label data includes 2M+1 bits, the first M bits take values of 0 and 1, indicating not using or using satellites; the last bit takes values of -1 or 1, wherein -1 represents optical, 1 represents SAR; the remaining M bits take values of T i represents the regression period of satellite operation, t i represents the time length passed in the current orbit regression period corresponding to the starting time of satellite shooting.

[0015] Further, the point group division model H based on historical scheme learning in step 3, wherein the parameter settings of each network layer are:

[0016] The input and output dimensions of the first linear layer are 7 and 32 respectively;

[0017] The input signal channel of the first convolutional layer is 32, the channel generated by convolution is set to 64, and the convolution kernel size is set to 1;

[0018] The convolution kernel size of the max pooling layer is set to 2, and the step is 2;

[0019] The number of network layers of the single-layer one-way long short-term memory network is 1, the input dimension is 64, the dimension of the hidden node is 32, and the output dimension is 16;

[0020] The number of multi-head attention heads in each Transformer encoder layer in the Transformer encoding module is 2, and the scaling factor d_model is 16;

[0021] The input and output dimensions of the second linear layer are 16 and 16 respectively;

[0022] The input and output dimensions of the third linear layer are 16 and 2M+1 respectively.

[0023] Further, the specific process of step 4 is:

[0024] (401) Initialize the number of iterations k, the maximum number of iterations K, K≥30, and the point group division model based on historical scheme learning in the kth iteration is H k , and let k=1, H k =H;

[0025] (402) Take I train as the input of the point group division model H k based on historical scheme learning, and perform forward propagation, the first convolutional layer and the single-layer one-way long short-term memory network encode the input point group features into feature representation F mid containing position information, F midThe point cluster feature representation F, which contains spatial and temporal information, is obtained after encoding by the Transformer encoding module. F is then input into the prediction layer to obtain the predicted value.

[0026] (403) Using the formula of Mean Absolute Error (MAE), through the predicted value The loss function value L for satellite usage is calculated based on the first M bits of the actual labeled data y. s The load loss value L is calculated using the last digit. pay The shooting time loss function value L is calculated using the remaining M bits. t The total loss value L is obtained. k =αL s +βL t +γL pay Where α, β, and γ represent weight values; and a backpropagation method is used, through the total loss value L k Calculate H k The gradient dθ of the weight parameters k The stochastic gradient descent method is used to analyze H. k The weight parameters are updated;

[0027] (404) Determine whether k≥K holds true. If yes, then the trained point group partitioning model H' based on historical scheme learning is obtained. Otherwise, let k = k+1 and return to step (402).

[0028] Furthermore, the satellite usage loss function value L mentioned in step (403) s Load loss value L pay Shooting time loss function value L t , gradient of weight parameters dθ k And the use of stochastic gradient descent method for H k The weight parameters are updated using the following formula:

[0029]

[0030] in, This represents the predicted value of the i-th dimension of the nth point group sample. This represents the true value of the i-th dimension in the label of the n-th point cluster sample, where n∈[1,N]. train ];dθ k H represents k The gradient of the weight parameters, This represents the partial derivative operation, θ k Represents the weight parameter, θ * This represents the updated weight parameters, and η represents the learning rate, where η∈[0,1].

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

[0032] 1. This invention combines the spatial and temporal characteristics of historical point cluster target planning schemes, learns from historical observation schemes, and obtains observation meta-tasks based on satellite usage and corresponding shooting start times, thereby effectively avoiding the problem of adjacent points being forcibly divided into different clusters and edge points being listed as separate point targets.

[0033] 2. The point grouping method based on historical scheme learning adopted in this invention also has the characteristics of simple process and high accuracy. Attached Figure Description

[0034] Figure 1 This is a flowchart of the point group partitioning method based on historical scheme learning of the present invention. Detailed Implementation

[0035] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0036] Reference Figure 1 The present invention provides a point cluster partitioning method based on historical scheme learning, comprising the following steps:

[0037] Step 1, Model historical scheme data scheme<Pgroup, Mlist>:

[0038] Modeling and representation of historical scheme data. Where Pgroup∈R N×1×2 Let represent a group of point targets corresponding to a solution. The point group contains N point targets, each represented by longitude and latitude. Mlist represents all meta-tasks in the solution, with the i-th meta-task consisting of a quadruple. It means, s i Satellite model Indicates the start time of filming. Represents the end time of shooting, payload i This indicates the type of satellite payload; a meta-mission can be determined using the satellite model, the start time of the image capture, and the satellite payload type.

[0039] Step 2, Obtain the training sample set I train and test sample set I test :

[0040] (201) Feature extraction from point cluster data. Select N from historical scheme data. train For each point group, the longitude and latitude of all points in the group are used to calculate indices representing the degree of discretization. The range, mean deviation, standard deviation, mean, variance, skewness, and kurtosis are calculated as characteristic representations of the entire point group, resulting in N. train Feature representation of a point group Similarly, N is selected. testN test characteristic representation of N N train ≥2000,

[0041] (202) label construction. For N train and N test corresponding label data is constructed for N

[0042] wherein M represents the number of satellites, the label data includes 2M+1 bits, the first M bits take values of 0 and 1, indicating not using or using satellites; the last bit takes values of -1 or 1, wherein -1 represents optical, and 1 represents SAR; the remaining M bits take values of T i representing the regression period of satellite operation, t i representing the time length passed in the current orbit regression period corresponding to the starting time of satellite shooting;

[0043] (203) construction of training sample set I train =(X train ,Y train ), and test sample set I test =(X test ,Y test ).

[0044] Step 3, constructing a point group division model H based on historical scheme learning:

[0045] The point group division model H based on historical scheme learning includes a backbone network and a prediction layer, the backbone network includes a first linear layer, a first convolutional layer, a Relu activation function, a maximum pooling layer, a single-layer one-way long short-term memory network and a Transformer encoding module connected in sequence, and the prediction layer includes a second linear layer, a dropout layer, a Relu activation function and a third linear layer; wherein the Transformer encoding module includes four layers of Transformer encoder layers.

[0046] The parameter settings of each network layer are introduced as follows:

[0047] The input and output dimensions of the first linear layer are 7 and 32 respectively;

[0048] The input signal channel of the first convolutional layer is 32, the channel generated by convolution is set to 64, and the convolution kernel size is set to 1;

[0049] The convolution kernel size of the maximum pooling layer is set to 2, and the step is 2;

[0050] The network layer number of the single-layer one-way long short-term memory network is 1, the input dimension is 64, the dimension of the hidden node is 32, and the output dimension is 16;

[0051] The number of multi-head attention heads in each Transformer encoder layer in the Transformer encoding module is 2, and the scaling factor d_model is 16;

[0052] The input and output dimensions of the second linear layer are 16 and 16, respectively;

[0053] The input and output dimensions of the third linear layer are 16 and 2M+1, respectively.

[0054] Step 4, using the training sample set and the test sample set to iteratively train the point group division model H:

[0055] (401) initialize the iteration number k, the maximum iteration number K, K≥30, and the point group division model based on historical scheme learning in the kth iteration is H k , and let k=1, H k =H;

[0056] (402) take I train as the input of the point group division model H k based on historical scheme learning, perform forward propagation, and the first convolutional layer and the single-layer one-way long short-term memory network encode the input point group features into feature representations F mid containing position information, F mid passes through the Transformer encoding module to obtain point group feature representations F containing spatial and temporal information, and F is input into the prediction layer to obtain the predicted value

[0057] (403) using the formula of mean absolute error MAE, calculate the satellite usage loss function value L s through the first M bits of the predicted value and the real label data y, calculate the load loss value L pay through the last bit, calculate the shooting time loss function value L t through the remaining M bits, and obtain the total loss value L k =αL s +βL t +γL pay , where α, β, γ represent weight values respectively; and using the back propagation method, calculate the weight parameter gradient dθ k of H k through the total loss value L k , and update the weight parameters of H k using the stochastic gradient descent method;

[0058] the satellite usage loss function value L s , the payload loss value L pay , the shooting time loss function value L t , the weight parameter gradient dθ k and the weight parameters of H k are updated by using the stochastic gradient descent method, and the calculation formula is:

[0059]

[0060] In the formula, represents the predicted value of the i-th dimension of the n-th point group sample, represents the true value of the i-th dimension in the label of the n-th point group sample, n∈[1,N train ]; dθ k represents the weight parameter gradient of H k , represents the partial derivative operation, θ k represents the weight parameter, θ * represents the updated weight parameter, and η represents the learning rate, η∈[0,1];

[0061] (404) judges whether k≥K is established, if yes, a trained point group division model H' based on historical scheme learning is obtained, otherwise, k=k+1 is set, and the step (402) is returned.

[0062] Step 5, generating a point group division scheme:

[0063] The test sample set is input into the point group division model H' based on historical scheme learning for forward propagation, the corresponding satellite usage, the corresponding shooting start time and the satellite payload type of each point group are obtained, which are mapped into the point group data to obtain a meta task list, and the point group division scheme is generated.

Claims

1. A point cloud partitioning method based on historical scheme learning, characterized in that, The method comprises the following steps: Step 1, modeling historical scheme data, including point group targets and meta tasks, each point group containing multiple point targets, each point target being represented by longitude and latitude, and each meta task being represented by a four-tuple including satellite model, shooting start time, shooting end time and satellite payload type; Step 2, extract point cloud data feature representation from historical scheme data, calculate discretization index as the feature representation of the entire point cloud, and construct the corresponding label data to obtain the training sample set I train =(X train ,Y train ) and the test sample set I test =(X test ,Y test ). Step 3, constructing a point group division model H based on historical scheme learning, including a backbone network and a prediction layer, the backbone network including a first linear layer, a first convolutional layer, a Relu activation function, a max pooling layer, a single-layer one-way long short-term memory network and a Transformer encoding module connected in sequence, and the prediction layer including a second linear layer, a dropout layer, a Relu activation function and a third linear layer; wherein the Transformer encoding module includes four Transformer encoder layers; Step 4, iteratively training the point group division model H using a training sample set and a test sample set to obtain a trained point group division model H' based on historical scheme learning; Step 5, performing forward propagation on the test sample set as input of the point group division model H' based on historical scheme learning to obtain satellite usage corresponding to each point group, corresponding shooting start time and satellite payload type, mapping to point group data to obtain a meta task list and generating a point group division scheme.

2. The point cloud partitioning method based on historical scheme learning according to claim 1, wherein, In step 2, N train point groups are selected, and the longitude and latitude of all point targets in each point group are calculated to represent the indexes of the degree of discretization, including range, average difference, standard deviation, mean, variance, skewness and kurtosis, as the feature representation of the entire point group, to obtain the feature representation of N train point groups Meanwhile, N test point groups are selected, and the feature representation of N test point groups is obtained Wherein, N train ≥2000, N train and N test point groups are constructed to obtain the corresponding label data Wherein, M represents the number of satellites, and the label data includes 2M+1 bits, the first M bits are 0 and 1, indicating not using or using the satellite; the last bit is-1 or 1, wherein-1 represents optical, and 1 represents SAR; the remaining M bits are T i represents the regression period of satellite operation, and t i represents the time length of the corresponding satellite starting shooting time in the current orbit regression period.

3. The point cloud partitioning method based on historical scheme learning according to claim 2, characterized in that, The point group division model H based on historical scheme learning in step 3, wherein the parameter settings of each network layer are as follows: The input and output dimensions of the first linear layer are 7 and 32 respectively; The input signal channel of the first convolutional layer is 32, the channel generated by convolution is set to 64, and the convolution kernel size is set to 1; The convolution kernel size of the max pooling layer is set to 2, and the step is 2; The number of network layers of the single-layer one-way long short-term memory network is 1, the input dimension is 64, the dimension of the hidden node is 32, and the output dimension is 16; The number of multi-head attention heads in each Transformer encoder layer in the Transformer encoding module is 2, and the scaling factor d_model is 16; The input and output dimensions of the second linear layer are 16 and 16 respectively; The input and output dimensions of the third linear layer are 16 and 2M+1 respectively.

4. The point cloud partitioning method based on historical scheme learning of claim 1, wherein, The specific process of step 4 is as follows: (401)Initialize the iteration number k, the maximum iteration number K, K≥30, and the point group division model based on historical scheme learning of the kth iteration as H k and let k = 1, H k = H; (402)I train As the input of the point cloud partitioning model H based on historical scheme learning, the first convolutional layer and the single-layer one-way long short-term memory network encode the input point cloud features into a feature representation F containing position information k mid mid The point cloud feature representation F containing spatial and temporal information is obtained by encoding through the Transformer encoding module, and F is input into the prediction layer to obtain the prediction value ​​ (403) Using the formula of Mean Absolute Error (MAE), through the predicted value The loss function value L for satellite usage is calculated based on the first M bits of the actual labeled data y. s The load loss value L is calculated using the last digit. pay The shooting time loss function value L is calculated using the remaining M bits. t The total loss value L is obtained. k =αL s +βL t +γL pay Where α, β, and γ represent weight values; and a backpropagation method is used, through the total loss value L k Calculate H k The gradient dθ of the weight parameters k The stochastic gradient descent method is used to analyze H. k The weight parameters are updated; (404) determining whether k >= K is true, if yes, obtaining the trained point group division model H' based on historical scheme learning, otherwise, setting k = k + 1 and returning to step (402).

5. The point cloud partitioning method based on historical scheme learning according to claim 1, wherein, The satellite usage loss function value L in step (403) s , a load loss value L pay , a shooting time loss function value L t , a weight parameter gradient dθ k , and the weight parameter of H k is updated by using a stochastic gradient descent method, and the calculation formula is: wherein, represents the predicted value of the i-th dimension of the n-th point cloud sample, represents the true value of the i-th dimension in the label of the n-th point cloud sample, n ∈ [1, N train ] ; dθ k represents the weight parameter gradient of H k , represents the partial derivative operation, θ k represents the weight parameter, θ * represents the updated weight parameter, and η represents the learning rate, η ∈ [0, 1].

Images