CN-121996944-A - One-dimensional range profile target prediction method based on deep learning

CN121996944ACN 121996944 ACN121996944 ACN 121996944ACN-121996944-A

Abstract

The invention belongs to the technical field of signal processing, and particularly discloses a one-dimensional range profile target prediction method based on deep learning, which comprises the steps of converting sparse feature vectors of strong scattering center dimension, target feature size and scattering center distribution entropy into corresponding dense feature vectors; the method comprises the steps of establishing a parallel FM model and an MLP model, wherein the FM model is used for learning feature interaction of strong scattering center dimension dense feature vectors, target feature size dense feature vectors and scattering center distribution entropy dense feature vectors, the MLP model is used for learning nonlinear relations among the strong scattering center dimension dense feature vectors, the target feature size dense feature vectors, the scattering center distribution entropy dense feature vectors and targets, and under the condition that model parameters of the FM model and the MLP model are determined through training, output of the FM model and output of the MLP model are fused to obtain a target prediction result. The technical problem that the existing one-dimensional range profile target prediction method is low in accuracy under the data sparse scene is solved.

Inventors

PAN MAOSHENG
KUANG WEIKANG
GONG JUNHAO

Assignees

北京遥感设备研究所

Dates

Publication Date: 20260508
Application Date: 20251223

Claims (10)

1. A one-dimensional range profile target prediction method based on deep learning is characterized by comprising the following steps: the sparse feature vectors of the strong scattering center dimension, the target feature size and the scattering center distribution entropy are converted into corresponding dense feature vectors; Establishing a parallel FM model and an MLP model, wherein the FM model is used for learning feature interactions of strong scattering center dimension dense feature vectors, target feature dimension dense feature vectors and scattering center distribution entropy dense feature vectors, and the MLP model is used for learning nonlinear relations between the strong scattering center dimension dense feature vectors, the target feature dimension dense feature vectors, the scattering center distribution entropy dense feature vectors and targets; under the condition that the model parameters of the FM model and the MLP model are determined through training, the output of the FM model and the output of the MLP model are fused, and a target prediction result is obtained.
2. The depth learning-based one-dimensional range profile target prediction method of claim 1, wherein converting sparse feature vectors of strong scattering center dimension, target feature size, scattering center distribution entropy into corresponding dense feature vectors comprises: And (3) converting the binary sparse feature vectors of the strong scattering center dimension, the target feature size and the scattering center distribution entropy into corresponding dense feature directions by utilizing feature embedding.
3. The depth learning-based one-dimensional range profile target prediction method of claim 2, wherein converting sparse feature vectors of binary strong scattering center dimension, target feature size, scattering center distribution entropy into corresponding dense feature vectors using feature embedding comprises: and mapping the sparse feature vector of each binary strong scattering center dimension, the target feature size and the scattering center distribution entropy through a corresponding embedded matrix to obtain a corresponding dense feature vector.
4. The depth learning-based one-dimensional range profile target prediction method of claim 1, wherein establishing the FM model comprises: Establishing a first-order module related to a prediction result of each single feature; establishing a high-order module related to the prediction result of the pairwise feature interaction; and associating the first-order module with the high-order module to obtain an FM model.
5. The depth learning-based one-dimensional range profile target prediction method of claim 1, wherein establishing the MLP model comprises: Establishing a nonlinear relation between a strong scattering center dimension dense feature vector and a target by using a neural network to learn a target feature dimension dense feature vector and a scattering center distribution entropy dense feature vector; An activation function of the MLP model is determined.
6. The depth learning based one-dimensional range profile target prediction method of claim 1, wherein training determines model parameters of each of the FM model and the MLP model, comprising: Defining an objective function corresponding to training according to the model parameters to be determined; the stochastic gradient descent optimizer is used to train in an end-to-end fashion to determine a locally optimal solution for the objective function.
7. The depth learning-based one-dimensional range profile target prediction method of claim 6, wherein defining training corresponding target functions according to model parameters to be determined comprises: the square loss of the target predicted value and the target actual value is determined as an objective function of the model parameters.
8. The depth learning based one-dimensional range profile target prediction method of claim 6, wherein training in an end-to-end manner with a stochastic gradient descent optimizer comprises: In each iteration training, the model parameters are updated according to the learning rate and the gradient until the iteration is finished.
9. The depth learning-based one-dimensional range profile target prediction method according to claim 6 or 8, wherein determining a locally optimal solution of the target function comprises: defining a loss function of the objective function; and under the condition that the convergence of the loss function or the iteration times reach a preset threshold value, acquiring the current solution of the objective function.
10. A one-dimensional range profile target prediction system based on deep learning, comprising: The transformation module is used for transforming the sparse feature vectors of the strong scattering center dimension, the target feature size and the scattering center distribution entropy into corresponding dense feature vectors; The model building module is used for building a parallel FM model and an MLP model, wherein the FM model is used for learning feature interactions of strong scattering center dimension dense feature vectors, target feature size dense feature vectors and scattering center distribution entropy dense feature vectors, and the MLP model is used for learning nonlinear relations among the strong scattering center dimension dense feature vectors, the target feature size dense feature vectors, the scattering center distribution entropy dense feature vectors and targets; And the prediction module is used for fusing the output of the FM model and the output of the MLP model under the condition of training and determining the model parameters of each of the FM model and the MLP model to obtain a target prediction result.

Description

One-dimensional range profile target prediction method based on deep learning Technical Field The invention belongs to the technical field of signal processing, and particularly relates to a one-dimensional range profile target prediction method based on deep learning. Background In the optical area, when the frequency spectrum of the radar emission signal is wide enough, the radial distance of the target occupies a plurality of radar distance resolution units, and the backward high-frequency electromagnetic scattering of the target presents continuous fluctuation characteristics in the time domain, so that a target amplitude image along the radar sight distance, namely a one-dimensional range profile, is formed. The current common practice of target prediction is to predict a target by calculating the matching degree of a target one-dimensional distance image and a corresponding attitude angle distance image in a library after the one-dimensional distance image is correlated and aligned. However, the related alignment of the range profile is involved, so that a large amount of calculation is caused, and meanwhile, the range profile is stored in the library, so that the storage capacity is relatively large. Over the past few years, some methods have been proposed by the scholars to improve the accuracy of target prediction, and these methods have been widely used. While having great success, these methods still have the following problems: (1) The existing target prediction method lacks qualitative analysis between target data and strong scattering center dimension, target feature size and scattering center distribution entropy. Studies have demonstrated the importance of strong scattering center dimensions, target feature sizes, scattering center distribution entropy in target prediction. However, the relation between the dimension of the target and the strong scattering center, the characteristic dimension of the target and the distribution entropy of the scattering center is not established explicitly, the dimension of the strong scattering center, the characteristic dimension of the target and the distribution entropy of the scattering center only indirectly participate in the prediction process, and the dimension of the strong scattering center, the characteristic dimension of the target and the distribution entropy of the scattering center are not involved in the subsequent prediction process, so that the perceptibility of the model on the dimension of the strong scattering center, the characteristic dimension of the target and the distribution entropy of the scattering center is further limited. (2) A commonly employed method for data prediction is a collaborative filtering method. The core idea of collaborative filtering is to make predictions using historical data of similar users. Although existing collaborative filtering methods improve accuracy of target prediction by improving similarity, improving reliability of neighbors, and mitigating the impact of target data ranges, these collaborative filtering methods only use historical target data of some similar neighbors to predict, and cannot make more accurate predictions with all data. The accuracy of the collaborative filtering method is closely related to the available data, and when the available target data is very thin, the accuracy of the collaborative filtering method is greatly reduced. Furthermore, in the case of sparse data, collaborative filtering methods may not work due to lack of commonly related targets between one-dimensional range profile features. (3) The existing prediction method ignores the influence of nonlinear interaction. The matrix decomposition method can learn one-dimensional range profile features and potential interaction feature vectors of the target from original target interaction data, and can use simple linear inner product operation of the one-dimensional range profile features and the target potential feature vectors to realize prediction of the target. The factorizer approach represents each feature as a potential vector, with the inner product of the feature potential vectors representing the interaction between them. Although the matrix decomposition and factorization machine method can improve the prediction accuracy to a certain extent by capturing the features and the potential features of the target, the simple linear model is difficult to capture the nonlinear relation between the one-dimensional range profile features and the target interaction, and the improvement of the target prediction accuracy is limited. Disclosure of Invention The invention aims to provide a one-dimensional range profile target prediction method based on deep learning, which aims to solve the technical problem that the existing one-dimensional range profile target prediction method is low in accuracy under a data sparse scene. In order to achieve the above purpose, the invention adopts the following technical scheme: A one-dime