Space target identification method

Abstract

The invention discloses a space target identification method. The method comprises the following steps: acquiring one-dimensional range profile data of a to-be-identified space target; converting the one-dimensional range profile data of the to-be-identified space target into a time-frequency graph; constructing a space target recognition model by using a differentiable neural network architecture search method; and inputting the time-frequency graph of the to-be-identified space target into the space target identification model so as to identify the category of the to-be-identified space target. According to the method, the category of the space target can be identified by using the one-dimensional range profile data, and meanwhile, relatively low calculation cost and relatively high identification precision can be kept.

Description

Technical Field

[0001] The present invention relates to the technical field of radar signal processing, and in particular to a space target recognition method. Background Art

[0002] With the continuous development of aerospace technology, more and more spacecraft are launched into space, and the space environment is becoming increasingly complex. Space targets generally refer to artificial targets located in the Earth's orbit and outer space that have been in orbit for a long time, including various types of spacecraft and space debris such as failed satellites, rocket bodies, and debris. Space target recognition technology can detect and respond to potential security threats in a timely manner, and can provide important support and services in civilian fields such as meteorological observation, resource exploration, and environmental monitoring. Therefore, the study of space target recognition is of great value.

[0003] Radar has the advantages of all-day, all-weather, and strong penetration, and is an important target detection device. The radar's one-dimensional range profile (HRRP, Height-Resolved Range Profile) is the vector sum of the target scattering point echoes projected in the direction of the radar ray. It contains rich physical structure information such as the size of the space target, echo amplitude, and scattering point distribution, which can achieve precise target type identification. In addition, the one-dimensional range profile is also easy to process and obtain. Therefore, how to identify space targets based on the radar's one-dimensional range profile is of great research value. Summary of the invention

[0004] The object of the present invention is to provide a space target recognition method, which can use one-dimensional range image data to identify the category of space targets while maintaining low calculation cost and high recognition accuracy.

[0005] In order to achieve the above object, the present invention is implemented by the following technical solutions:

[0006] A space target recognition method, comprising:

[0007] Obtain one-dimensional range image data of the space target to be identified;

[0008] Converting the one-dimensional range image data of the space target to be identified into a time-frequency graph;

[0009] Use differentiable neural network architecture search method to build a spatial target recognition model;

[0010] The time-frequency diagram of the space target to be identified is input into the space target identification model to identify the category of the space target to be identified.

[0011] Optionally, the step of converting the one-dimensional range image data of the space target to be identified into a time-frequency diagram includes:

[0012] Multiplying the one-dimensional range image data of the space target to be identified by a sliding window function to divide the one-dimensional range image data of the space target to be identified into a plurality of time segment data;

[0013] Performing Fourier transform on each of the time segment data to calculate the frequency component of the one-dimensional range image data in the time segment data;

[0014] The frequency components of the one-dimensional range image data in all the time segment data are orderly combined to obtain a time-frequency diagram of the space target to be identified.

[0015] Optionally, before executing the step of building a spatial target recognition model using a differentiable neural network architecture search method, the step further includes:

[0016] Obtain one-dimensional range image data of identified space targets;

[0017] Converting the one-dimensional range image data of the identified space target into a time-frequency graph;

[0018] The time-frequency diagram of the identified space target is divided into a training set, a validation set and a test set.

[0019] Optionally, the ratio of the number of the training set, the validation set and the test set is 6:2:2.

[0020] Optionally, the steps of constructing a spatial target recognition model using a differentiable neural network architecture search method include:

[0021] Constructing a neural network model architecture; the neural network model architecture includes at least four basic units, each of which includes an input node, a plurality of intermediate nodes connected to the input node, and an output node connected to the plurality of intermediate nodes; and the output node of the former of two adjacent basic units is connected to the input node of the latter;

[0022] Constructing a structural search space and a weight parameter search space of the neural network model; the structural search space of the neural network model includes a convolution operation, a depth-separable convolution operation, a maximum pooling operation, an average pooling operation, a skip connection operation, and a direct connection operation; the weight parameter search space includes the height, width, and output depth of the convolution kernel, and the size of the pooling area;

[0023] Using a Softmax function to relax the structure search space in the neural network model so that the relaxed structure search space is continuous;

[0024] The parameters of the neural network model are optimized using the gradient descent method and the Adam algorithm to obtain the optimal network model and use it as the spatial target recognition model.

[0025] Optionally, in each of the basic units, the input node and each of the intermediate nodes, as well as each of the intermediate nodes and the output node are connected by edges, each edge represents an operation; and the operation is selected from the structural search space.

[0026] Optionally, before executing the step of relaxing the structural search space in the neural network model using the Softmax function, the method further includes: assigning an architectural parameter to each edge.

[0027] Optionally, it is characterized in that the step of optimizing the parameters of the neural network model by using the gradient descent method and the Adam algorithm to obtain the optimal network model comprises:

[0028] Optimizing the weight parameters in the weight parameter search space using the training set and a gradient descent method to obtain an optimal weight parameter value;

[0029] Optimizing the architecture parameters in the relaxed architecture search space using the validation set and the Adam algorithm to obtain optimal architecture parameter values;

[0030] Determine the operation corresponding to each edge according to the optimal architecture parameter value to obtain the optimal network model architecture;

[0031] The optimal network model architecture is combined with the optimal weight parameter value to obtain the optimal network model.

[0032] Compared with the prior art, the present invention has at least one of the following advantages:

[0033] The present invention provides a method for identifying a space target, which converts the one-dimensional range image data of the space target to be identified into a time-frequency diagram, and constructs a space target recognition model using a differentiable neural network architecture search method; the time-frequency diagram of the space target to be identified is input into the space target recognition model, so that the category of the space target to be identified can be identified.

[0034] When the present invention uses a differentiable neural network architecture search method to construct a space target recognition model, it can automatically search for the optimal neural network and automatically design a suitable network structure based on the time-frequency diagram of the identified space target and the recognition task, without relying on manual experience and intuition. At the same time, it can also significantly improve the construction efficiency and recognition accuracy of the space target recognition model while maintaining a low computing cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 is a flow chart of a method for identifying a space target provided by an embodiment of the present invention;

[0036] Figure 2 is one-dimensional range image data of an identified space target provided by an embodiment of the present invention;

[0037] Figure 3 is a time-frequency diagram of an identified space target provided by an embodiment of the present invention;

[0038] Figure 4 is a schematic diagram of the architecture of a neural network model provided by an embodiment of the present invention;

[0039] Figure 5 It is a schematic diagram of the structure of each basic unit in the neural network model provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0040] The following is a further detailed description of a method for identifying a space target proposed in the present invention in combination with the accompanying drawings and specific implementation methods. According to the following description, the advantages and features of the present invention will be clearer. It should be noted that the drawings are in a very simplified form and use non-precise proportions, which are only used to conveniently and clearly assist in explaining the purpose of the implementation method of the present invention. In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, please refer to the accompanying drawings. It should be noted that the structure, proportion, size, etc. illustrated in the drawings of this specification are only used to match the content disclosed in the specification for people familiar with this technology to understand and read, and are not used to limit the limiting conditions for the implementation of the present invention, so it has no technical substantive significance. Any modification of the structure, change in the proportional relationship or adjustment of the size, without affecting the effect that the present invention can produce and the purpose that can be achieved, should still fall within the scope of the technical content disclosed by the present invention.

[0041] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0042] In recent years, artificial intelligence technology represented by deep learning has developed rapidly. As a machine learning based on deep neural network models and methods, deep learning has achieved results far exceeding traditional methods in various fields. Therefore, it is of great research significance to use deep learning to realize space target recognition technology based on radar one-dimensional range profile.

[0043] The representation ability of a neural network model is closely related to its architecture, so it is very important to design a suitable neural network architecture for a specific task. In the past, the design of the neural network model architecture was mainly done manually, relying on the past experience of experts in the field of deep learning. Every time a new research task is encountered, a special network architecture, network depth, number of channels, etc. need to be designed. For example, classic models such as ResNet and GoogleNet are carefully designed by deep learning experts with rich experience. In order to improve the performance of deep learning models, the neural network model architecture designed by experts has become more and more complex, the number of network layers of the model has become deeper and deeper, the network architecture has become more and more variable, and the parameters of the model have also increased exponentially. In addition, the increase in the complexity of any aspect of the neural network model will cause the number of candidate architectures to grow exponentially, which requires more computing resources for model evaluation, and these expansion tasks are very time-consuming and laborious, and many of them may even be repetitive. Therefore, how to replace the manual design of the network structure with automated search and design a neural network model with fewer parameters and strong recognition ability is a key issue for realizing space target recognition based on radar one-dimensional range images.

[0044] Combined with Figures 1 to 5 As shown, this embodiment provides a space target recognition method, including: step S1, obtaining one-dimensional range image data of the space target to be recognized; step S2, converting the one-dimensional range image data of the space target to be recognized into a time-frequency graph; step S3, constructing a space target recognition model using a differentiable neural network architecture search method; step S4, inputting the time-frequency graph of the space target to be recognized into the space target recognition model to identify the category of the space target to be recognized.

[0045] Specifically, in step S1, a radar may be used to acquire one-dimensional range profile data of the space target to be identified.

[0046] The step S2 includes: step S21, multiplying the one-dimensional range image data of the space target to be identified by a sliding window function (such as a rolling function, a Window Functions function, etc.) to divide the one-dimensional range image data of the space target to be identified into multiple shorter time segment data. Step S22, performing Fourier transform on each of the time segment data to calculate the frequency components of the one-dimensional range image data in the time segment data; and each of the time segment data corresponds to a spectrum. Step S23, orderly combining the frequency components of the one-dimensional range image data in all the time segment data (i.e., the Fourier transform results) to obtain the time-frequency diagram of the space target to be identified. The time-frequency diagram is a time-frequency representation of the one-dimensional range image data, that is, the time-frequency diagram is represented as a two-dimensional array, in which one dimension represents time and the other dimension represents frequency.

[0047] Specifically, before executing step S3, the method also includes: acquiring one-dimensional range image data of the identified space target; converting the one-dimensional range image data of the identified space target into a time-frequency graph; and dividing the time-frequency graph of the identified space target into a training set, a verification set and a test set.

[0048] In this embodiment, the one-dimensional range image data of the identified space target is the one-dimensional range image data obtained by electromagnetic calculation simulation of five satellite targets. Specifically, in the electromagnetic calculation simulation of five satellite targets, the observation distance is 10km, the simulated azimuth range is 0-360°, the elevation range is -25°-+25°, and the electromagnetic scattering simulation calculation is performed every 5°; the center frequency of the simulation is 16GHz, the bandwidth is 150M, and the step frequency is 5M, that is, each angle provides 501 frequency points of sweep frequency data; among them, the one-dimensional range image data of a satellite with an azimuth of 120° and an elevation of 25° is as follows: Figure 2 shown.

[0049] Furthermore, the step of converting the one-dimensional range image data of the identified space target into a time-frequency diagram is similar to the above step S2, and specifically includes: multiplying the one-dimensional range image data of the identified space target by a sliding window function to also divide the one-dimensional range image data of the identified space target into a plurality of shorter time segment data; performing Fourier transform on each of the time segment data to calculate the frequency components of the one-dimensional range image data in the time segment data; and each of the time segment data corresponds to a spectrum; and orderly combining the frequency components of the one-dimensional range image data in all the time segment data (i.e., Fourier transform results) to obtain the time-frequency diagram of the identified space target (such as Figure 3 shown).

[0050] Optionally, the time-frequency graph of the identified space target is divided into a training set, a validation set, and a test set in a ratio of 6:2:2, that is, the ratio of the number of the training set, the validation set, and the test set is 6:2:2. In this embodiment, the number of samples in the training set is 2845, the number of samples in the validation set is 950, and the number of samples in the test set is 950.

[0051] Specifically, in step S3, when the space target recognition model is constructed using the differentiable neural network architecture search method, a suitable network structure can be automatically designed according to the time-frequency diagram of the identified space target and the recognition task, and the optimal network model can be automatically searched without relying on manual experience and intuition. At the same time, the performance and construction efficiency of the space target recognition model can be significantly improved while maintaining a low computational cost.

[0052] Please also refer to Figure 4 and Figure 5 , the step S3 includes: step S31, constructing the architecture of the neural network model; the architecture of the neural network model includes at least four basic units, each of which includes an input node, multiple intermediate nodes connected to the input node, and an output node connected to the multiple intermediate nodes; and the output node of the former of two adjacent basic units is connected to the input node of the latter. Step S32, constructing the structural search space and weight parameter search space of the neural network model; the structural search space of the neural network model includes convolution operation, depth-separable convolution operation, maximum pooling operation, average pooling operation, skip connection operation and direct connection operation; the weight parameter search space includes the height, width and output depth of the convolution kernel, and the size of the pooling area. Step S33, using the Softmax function to relax the structural search space in the neural network model so that the relaxed structural search space is continuous. Step S34, using the gradient descent method and the Adam algorithm to optimize the parameters of the neural network model to obtain the optimal network model and use it as the spatial target recognition model.

[0053] Specifically, Figure 4 As shown, four basic units can be stacked to form the architecture of the neural network model, and a node in each basic unit represents a potential representation of a network layer (such as a feature map in a convolutional network). Figure 5 As shown, each of the basic units is composed of 6 nodes, namely Figure 5 The corresponding input nodes, four intermediate nodes and output nodes from left to right in the figure are shown in Figure 1. The output node uses a concatenation operation to concatenate the feature maps of all the intermediate nodes to obtain the output feature map of the basic unit.

[0054] More specifically, in each of the basic units, the input node and each of the intermediate nodes, as well as each of the intermediate nodes and the output node are connected by edges, each edge represents an operation; and the operation is selected from the structure search space in a selective manner. Optionally, the edge between each of the intermediate nodes and the output node represents a direct connection operation.

[0055] Furthermore, in each of the basic units, each intermediate node obtains the feature map of the intermediate node after performing corresponding operations on the feature map of the previous node (i.e., the input node), and then the feature map of all the intermediate nodes is added together to obtain the feature map of the output node, which is the output feature map of the basic unit.

[0056] The characteristic graph of each node in each basic unit can be expressed by the following formula:

[0057]

[0058] Among them, x (j) Represents the feature graph of the jth node in the basic unit; o (i,j) represents the operation from the i-th node to the j-th node in the basic unit; x (i) Represents the feature graph of the i-th node in the basic unit; i and j are both integers.

[0059] Specifically, in the weight parameter search space of the neural network model, the height h and width w of the convolution kernel are both set to {1, 3, ..., 7}, and the depth c is set to {4, 8, ..., 1024}; the size of the pooling area is set to {2×2, 3×3, ..., 5×5}, but the present invention is not limited to this.

[0060] Specifically, before executing step S33, the method further includes: assigning a structure parameter to each edge. In this embodiment, an edge connecting the i-th node and the j-th node is assigned a structure parameter α={α (i,j)}, that is, the task of structure search is simplified to learning continuous variables α = {α (i,j) During the training process, the Softmax function is used to apply continuous relaxation to the structure search space, so that the structure search space changes from discrete to continuous, and the output is redefined as:

[0061]

[0062] Among them, y represents the feature map output after the feature map x is operated; σ(·) represents the activation function used to generate the architecture parameters; o represents the operation; O represents the discrete structure search space; o(x) represents the operation on the feature map x.

[0063] Furthermore, the expression of the relaxed structure search space is:

[0064]

[0065] Among them, o represents operation; o (i,j) (·) represents the operation from the i-th node to the j-th node; x represents the feature graph of each node; o' represents the operation other than o; Represents the architectural parameters assigned to the operation taken from the i-th node to the j-th node; In addition to Operations other than those specified are given schema parameters.

[0066] Specifically, the step S34 includes: step S341, using the training set and adopting the gradient descent method to optimize the weight parameters in the weight parameter search space to obtain the optimal weight parameter value; step S342, using the verification set and adopting the Adam algorithm to optimize the architecture parameters in the relaxed structure search space to obtain the optimal architecture parameter value; step S343, determining the operation corresponding to each edge according to the optimal architecture parameter value to obtain the optimal network model architecture; step S344, combining the optimal network model architecture with the optimal weight parameter value to obtain the optimal network model.

[0067] In this embodiment, after the architecture parameters are introduced, in the training of the neural network model, not only the traditional weight parameters need to be optimized, but also the architecture parameters need to be optimized; and the weight parameters in the weight parameter search space are denoted as ω. trian and L val is the training loss and validation loss of the neural network model. The goal is to find the optimal architecture parameter value α * , so that it is optimal on the validation set, that is, minimizing L val (ω * (α),α); and find the optimal parameter weight value ω * , which can be achieved at the optimal architecture parameter value α * The optimal performance is obtained on the corresponding specific structure, that is, minimizing the training set ω * (α) = argmin ω L trian (ω,α * ), that is,

[0068]

[0069] Define the learning rate ξ and approximate formula (4) as

[0070]

[0071] The optimal result is obtained by alternately optimizing the weight parameter ω and the architecture parameter α. The specific optimization process is: set the upper limit of the number of iterations to N. When the number of iterations is less than N, fix the architecture parameter α, and use the samples in the training set to train the weight parameter ω to obtain the optimal weight parameter value ω. * ; Fixed weight parameter ω, use the samples in the validation set to train the architecture parameter α, the optimal architecture parameter value α * .

[0072] Specifically, in step S343, the argmax function is used to select the operation with the largest architecture parameter on each edge to replace the mixed operation to form the final search unit topology structure, that is, the following formula:

[0073]

[0074] After replacement, the optimal network model architecture can be obtained.

[0075] In step S344, the optimal network model architecture and the optimal weight parameter value are combined to obtain the optimal network model. At this time, the optimal network model is the space target recognition model; and the accuracy of the space target recognition model can be verified through the test set.

[0076] In this embodiment, the test set is input into the constructed space target recognition model, and the recognition accuracy of the space target recognition model is 91.58%, and the recognition accuracy of each type of target is shown in Table 1; and the parameter amount of the space target recognition model is 6.1M. Optionally, the size of the time-frequency graph of the identified space target in the sample set is 600×600.

[0077] Table 1 Recognition accuracy of spatial target recognition model

[0078] Goal 1 Goal 2 Goal 3 Goal 4 Goal 5 Recognition accuracy 80％ 95.79％ 98.42％ 90.53％ 93.16％

[0079] In summary, the present embodiment provides a method for identifying a space target, which converts the one-dimensional range image data of the space target to be identified into a time-frequency diagram, and uses a differentiable neural network architecture search method to construct a space target recognition model; the time-frequency diagram of the space target to be identified is input into the space target recognition model, and the category of the space target to be identified can be identified. In the present embodiment, when the space target recognition model is constructed using a differentiable neural network architecture search method, the optimal neural network can be automatically searched, and a suitable network structure can be automatically designed according to the time-frequency diagram of the identified space target and the recognition task, without relying on manual experience and intuition. At the same time, it can also significantly improve the construction efficiency and recognition accuracy of the space target recognition model, while maintaining a low computational cost.

[0080] Although the content of the present invention has been described in detail through the above preferred embodiments, it should be appreciated that the above description should not be considered as a limitation of the present invention. After reading the above content, it will be apparent to those skilled in the art that various modifications and substitutions of the present invention will occur. Therefore, the protection scope of the present invention should be limited by the appended claims.

Claims

1. A space target recognition method, characterized in that: include: Obtain one-dimensional range image data of the space target to be identified; Converting the one-dimensional range image data of the space target to be identified into a time-frequency graph; Use differentiable neural network architecture search method to build a spatial target recognition model; The time-frequency diagram of the space target to be identified is input into the space target identification model to identify the category of the space target to be identified.

2. The space target recognition method according to claim 1, characterized in that: The step of converting the one-dimensional range image data of the space target to be identified into a time-frequency graph comprises: Multiplying the one-dimensional range image data of the space target to be identified by a sliding window function to divide the one-dimensional range image data of the space target to be identified into a plurality of time segment data; Performing Fourier transform on each of the time segment data to calculate the frequency component of the one-dimensional range image data in the time segment data; The frequency components of the one-dimensional range image data in all the time segment data are orderly combined to obtain the time-frequency diagram of the space target to be identified.

3. The space target recognition method according to claim 1, characterized in that: Before executing the step of building a spatial target recognition model using a differentiable neural network architecture search method, the step also includes: Obtain one-dimensional range image data of identified space targets; Converting the one-dimensional range image data of the identified space target into a time-frequency graph; The time-frequency diagram of the identified space target is divided into a training set, a validation set and a test set.

4. The space target recognition method according to claim 3, characterized in that: The ratio of the number of the training set, the validation set and the test set is 6:2:

2.

5. The space target recognition method according to claim 3, characterized in that: The steps of building a spatial target recognition model using a differentiable neural network architecture search method include: Constructing a neural network model architecture; the neural network model architecture includes at least four basic units, each of which includes an input node, a plurality of intermediate nodes connected to the input node, and an output node connected to the plurality of intermediate nodes; and the output node of the former of two adjacent basic units is connected to the input node of the latter; Constructing a structural search space and a weight parameter search space of the neural network model; the structural search space of the neural network model includes a convolution operation, a depth-separable convolution operation, a maximum pooling operation, an average pooling operation, a skip connection operation, and a direct connection operation; the weight parameter search space includes the height, width, and output depth of the convolution kernel, and the size of the pooling area; Using a Softmax function to relax the structure search space in the neural network model so that the relaxed structure search space is continuous; The parameters of the neural network model are optimized using the gradient descent method and the Adam algorithm to obtain the optimal network model and use it as the spatial target recognition model.

6. The space target recognition method according to claim 5, characterized in that: In each of the basic units, the input node and each of the intermediate nodes, as well as each of the intermediate nodes and the output node are connected via edges, each edge represents an operation; the operation is selected from the structural search space.

7. The space target recognition method according to claim 6, characterized in that: Before executing the step of relaxing the structural search space in the neural network model using the Softmax function, the method also includes: assigning an architectural parameter to each edge.

8. The space target recognition method according to claim 7, characterized in that: The steps of optimizing the parameters of the neural network model by using the gradient descent method and the Adam algorithm to obtain the optimal network model include: Optimizing the weight parameters in the weight parameter search space using the training set and a gradient descent method to obtain an optimal weight parameter value; Optimizing the architecture parameters in the relaxed architecture search space using the validation set and the Adam algorithm to obtain optimal architecture parameter values; Determine the operation corresponding to each edge according to the optimal architecture parameter value to obtain the optimal network model architecture; The optimal network model architecture is combined with the optimal weight parameter value to obtain the optimal network model.

Images