RCS prediction method and device based on Transform model

Abstract

The invention discloses an RCS prediction method and device based on a Transform model, and the method comprises the steps: constructing a pre-training network based on self-supervision, training the pre-training network through the RCS sequence training data, obtaining a trained pre-training network, and enabling a Transform-based RCS prediction network in the trained pre-training network to serve as an initial Transform-based RCS prediction network; introducing consideration of an attitude angle into the RCS sequence training data, and training an initial Transform-based RCS prediction network by using the RCS sequence training data to obtain a trained Transform-based RCS prediction network; and introducing consideration of an attitude angle into to-be-predicted RCS sequence data, and inputting the to-be-predicted RCS sequence data into the trained Transform-based RCS prediction network to obtain a prediction result.

Description

Technical Field

[0001] The present invention belongs to the field of radar communication technology, and in particular relates to a method and device for predicting RCS (Radar Cross-Section) based on a Transformer model. Background Art

[0002] In recent years, with the rapid development of stealth technology, highly maneuverable targets, and complex electromagnetic interference in modern battlefield environments, radar systems are facing challenges such as unstable tracking performance caused by fluctuations in the target's radar cross-section. To improve radar's ability to predict the RCS of complex targets, deep learning (DL) has attracted the attention of many researchers.

[0003] In real-world scenarios, factors such as target attitude angle fluctuations, electromagnetic interference, and the application of stealth technology cause RCS to exhibit nonlinear, time-varying characteristics. This dynamic fluctuation directly affects the echo signal-to-noise ratio, leading to increased tracking errors and even target loss, severely limiting the performance of multi-target tracking. Traditional RCS research methods (such as physical optics and geometric diffraction theory) are based on static RCS statistical databases or theoretical calculations, making them difficult to adapt to the demands of highly dynamic battlefield environments.

[0004] Time series prediction algorithms provide a path for modeling dynamic RCS sequences. Their core focus is on exploring how electromagnetic scattering characteristics evolve with target motion and environmental disturbances. Time series prediction is a method for predicting future values ​​based on time series data, which is a sequence of observations arranged in chronological order. The basic principle of time series prediction algorithms is to assume that past trends and patterns will continue to some extent in the future. By analyzing time series data, including trends, seasonality, periodicity, and random fluctuations, they can predict future data.

[0005] However, a key problem faced by existing technologies such as ARIMA (Autoregressive Integrated Moving Average Model) and exponential smoothing is that they rely heavily on stationarity and linearity assumptions, making it difficult to handle the non-stationarity of RCS (such as mutations, periodicity loss) and nonlinear complex time patterns.

[0006] Therefore, there is an urgent need to provide a high-precision radar RCS prediction method based on deep learning to improve the performance stability of multi-target tracking. Summary of the Invention

[0007] In order to solve the above problems existing in the prior art, the present invention provides an RCS prediction method and device based on the Transformer model. The technical problem to be solved by the present invention is achieved through the following technical solutions: In a first aspect, an embodiment of the present invention provides an RCS prediction method based on a Transformer model, the method comprising: Construct RCS sequence training data; A self-supervised pre-training network is constructed, and the self-supervised pre-training network is trained using the RCS sequence training data to obtain a trained self-supervised pre-training network, and the Transformer-based RCS prediction network in the trained self-supervised pre-training network is used as the initial Transformer-based RCS prediction network; wherein, a random mask layer is introduced into the self-supervised pre-training network, and a random masking strategy is implemented on the input RCS sequence training data, so that the self-supervised pre-training network reconstructs the masked content through context perception to complete pre-training; a multi-head autocorrelation mechanism and a cascade structure of a feedforward neural network are adopted in the Transformer-based RCS prediction network, so that different feature spaces are captured in parallel through multiple autocorrelation heads, dependency discovery and information aggregation are realized at the sequence level, and the ability to model complex nonlinear patterns in the RCS sequence training data is enhanced through the feedforward neural network; Introducing attitude angle considerations into the RCS sequence training data to generate new RCS sequence training data, and further training the initial Transformer-based RCS prediction network using the new RCS sequence training data to obtain a trained Transformer-based RCS prediction network; The attitude angle is introduced into the RCS sequence data to be predicted to generate new RCS sequence data to be predicted, and the new RCS sequence data to be predicted is input into the trained Transformer-based RCS prediction network to obtain the prediction result.

[0008] In one embodiment of the present invention, the RCS sequence training data and the RCS sequence data to be predicted both include a stationary RCS sequence satisfying a specific distribution probability and a non-stationary RCS sequence of a high-maneuverability target.

[0009] In one embodiment of the present invention, the self-supervised pre-training network includes an input layer, a sequence block and normalization layer, a random mask layer, a position encoding and high-dimensional projection layer, a Transformer-based RCS prediction network, and a linear output layer connected in sequence.

[0010] In one embodiment of the present invention, a Transformer-based RCS prediction network is composed of two symmetrical modules, an encoder and a decoder, stacked together; the encoder includes N sub-encoders connected in sequence, and the decoder includes N sub-decoders connected in sequence. The output of the Nth sub-encoder is connected to each sub-decoder, and the output of the Nth sub-decoder is used as the final output of the Transformer-based RCS prediction network after passing through a linear output layer, where N is an integer greater than 0; wherein, Each sub-encoder includes a multi-head autocorrelation module, a residual connection layer, a feedforward neural network layer, and a residual connection layer connected in sequence. The input of the multi-head autocorrelation module in the first sub-encoder includes word embedding, position encoding, and angle encoding; the first residual connection layer in the n-th sub-encoder is also connected to the input of the multi-head autocorrelation module in the n-th sub-encoder, and the second residual connection layer in the n-th sub-encoder is also connected to the first residual connection layer in the n-th sub-encoder, where n ranges from 0 to N; Each sub-decoder includes a multi-head autocorrelation module, a residual connection layer, a multi-head autocorrelation module, a residual connection layer, a feedforward neural network layer, a residual connection layer, and a linear output layer connected in sequence. The input of the first sub-decoder includes word embedding and position encoding. The first residual connection layer in the n-th sub-decoder is also connected to the input of the multi-head autocorrelation module in the n-th sub-decoder, the second residual connection layer in the n-th sub-decoder is also connected to the first residual connection layer in the n-th sub-decoder, and the third residual connection layer in the n-th sub-decoder is also connected to the second residual connection layer in the n-th sub-decoder. The input of the second multi-head autocorrelation module in the nth sub-decoder is also connected to the output of the Nth sub-encoder.

[0011] In one embodiment of the present invention, the implementation mechanism of the multi-head autocorrelation module is expressed as follows: ; ; ; ; ; in, Represents a query for the multi-head autocorrelation module, represents the key of the multi-head autocorrelation module, represents the value of the long-term autocorrelation module, represents the weight of each head in the multi-head autocorrelation module, Represents a splicing operation, represents the number of heads in the multi-head autocorrelation module, Indicates the first The output of the head, Indicates the first Individual queries, Indicates the first The key of the head, Indicates the first The value of the head, Express Perform delay operation and the delay is , Indicates the number of delays, , is the empirical hyperparameter, Indicates the length of the input RCS sequence, Indicates a round-down operation. express and In the delay The autocorrelation value at express and In the delay The autocorrelation value at represents the normalization operation, Express The normalized autocorrelation value is Express The normalized autocorrelation value is Indicates delay The value is 1~ L When, from L indivual and In the delay Get the maximum autocorrelation value when k The time delay corresponding to the autocorrelation value.

[0012] In one embodiment of the present invention, the position code of the multi-head autocorrelation module input in the first sub-encoder and the position code of the first sub-decoder input are both obtained by encoding the input RCS sequence using sine and cosine functions; wherein, In the training process of the self-supervised pre-trained network, the input RCS sequence is the RCS sequence output after position encoding and high-dimensional projection layer; During the initial training process of the Transformer-based RCS prediction network, the input RCS sequence is the new RCS sequence training data.

[0013] In one embodiment of the present invention, the attitude angle is introduced into the RCS sequence training data to generate new RCS sequence training data, including: The angle code obtained by converting the posture angle corresponding to the RCS sequence training data through sine / cosine angle coding is tensor-added with the position code and data feature code to generate new RCS sequence training data.

[0014] In one embodiment of the present invention, the attitude angle is introduced into the RCS sequence data to be predicted to generate new RCS sequence data to be predicted, including: The attitude angle corresponding to the RCS sequence data to be predicted is converted through sine / cosine angle coding, and then tensor-added with the position coding and data feature coding to generate new RCS sequence data to be predicted.

[0015] In a second aspect, an embodiment of the present invention provides an RCS prediction device based on a Transformer model, the RCS prediction device comprising: Data construction module, used to construct RCS sequence training data; The first network training module is used to construct a self-supervised pre-training network, use the RCS sequence training data to train the self-supervised pre-training network to obtain a trained self-supervised pre-training network, and use the Transformer-based RCS prediction network in the trained self-supervised pre-training network as the initial Transformer-based RCS prediction network; wherein, a random mask layer is introduced into the self-supervised pre-training network, and a random masking strategy is implemented on the input RCS sequence training data, so that the self-supervised pre-training network reconstructs the masked content through context perception to complete pre-training; the Transformer-based RCS prediction network adopts a cascade structure of a multi-head autocorrelation mechanism and a feedforward neural network to capture different feature spaces in parallel through multiple autocorrelation heads, realize dependency discovery and information aggregation at the sequence level, and enhance the ability to model complex nonlinear patterns in the RCS sequence training data through the feedforward neural network; a second network training module, configured to introduce attitude angle considerations into the RCS sequence training data, generate new RCS sequence training data, and further train the initial Transformer-based RCS prediction network using the new RCS sequence training data to obtain a trained Transformer-based RCS prediction network; The prediction module is used to introduce the consideration of attitude angle into the RCS sequence data to be predicted, generate new RCS sequence data to be predicted, and input the new RCS sequence data to be predicted into the trained Transformer-based RCS prediction network to obtain the prediction result.

[0016] Beneficial effects of the present invention: The RCS prediction method based on the Transformer model proposed in the present invention introduces a random mask layer in the pre-training network based on self-supervision. By implementing a random masking strategy on the input RCS sequence data, the model reconstructs the masked content through context perception to complete pre-training, thereby avoiding the defect of weak generalization ability of the model, improving the prediction accuracy of downstream RCS prediction tasks, and the generalization ability of different RCS sequences; the RCS prediction network based on the Transformer model proposed in the present invention adopts a multi-head autocorrelation mechanism and a cascade structure of a feedforward neural network, and its multi-head autocorrelation submodule can capture different characteristics in parallel through multiple autocorrelation heads The feedforward neural network enhances the ability to model complex nonlinear patterns in the RCS sequence, significantly improving the RCS prediction capability. The present invention takes into account the strong correlation between the radar target scattering characteristics and the target motion attitude angle, establishes an attitude angle-RCS correlation mapping, adds angle stamps to the original RCS sequence data, improves the prediction capability of complex time series, and provides a more robust time series prediction model for radar multi-target tracking. The present invention is highly compatible with the existing technology, abandons the traditional sequential processing methods of recurrent neural networks (RNNs) and convolutional neural networks (CNNs), and can be calculated in parallel, greatly improving the training speed, thus overcoming the shortcomings of the existing technology.

[0017] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 1 is a flow chart of an RCS prediction method based on a Transformer model provided by an embodiment of the present invention; Figure 2 is a schematic diagram of the structure of a self-supervised pre-training network provided by an embodiment of the present invention; Figure 3 2 is a schematic diagram of the structure of a Transformer-based RCS prediction network provided by an embodiment of the present invention; Figure 4 This is a more detailed structural diagram of the Transformer-based RCS prediction network provided by an embodiment of the present invention; Figure 5 This is a schematic diagram of an implementation process of an autocorrelation mechanism provided by an embodiment of the present invention; Figure 6 Schematic diagram of the input data encoding style of the encoder in the Transformer-based RCS prediction network provided by an embodiment of the present invention; Figure 7(a) to Figure 7(b) This is a visualization diagram of the prediction experimental results of the Transformer-based RCS prediction network on RCS sequence 1 provided by an embodiment of the present invention; Figure 8(a) to Figure 8(b) This is a visualization diagram of the prediction experimental results of the Transformer-based RCS prediction network on RCS sequence 2 provided by an embodiment of the present invention; Figure 9 It is a structural diagram of an RCS prediction device based on a Transformer model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The present invention will be further described in detail below with reference to specific examples, but the embodiments of the present invention are not limited thereto.

[0020] First, see Figure 1 The embodiment of the present invention provides an RCS prediction method based on the Transformer model, which specifically includes the following steps: S10. Construct RCS sequence training data.

[0021] In this embodiment of the present invention, RCS sequence training data is composed of K segments of RCS sequences of length M. Each RCS sequence includes different categories of radar targets, including stationary RCS sequences that satisfy specific distribution probabilities (Sweelin model, Rice distribution, and chi-square distribution) and non-stationary RCS sequences of highly maneuverable targets. K ≥ 2, M ≥ 5000.

[0022] S20. Construct a pre-training network based on self-supervision, use the RCS sequence training data to train the pre-training network based on self-supervision to obtain a trained pre-training network based on self-supervision, and use the Transformer-based RCS prediction network in the pre-training network based on self-supervision as the initial Transformer-based RCS prediction network; wherein, a random mask layer is introduced into the pre-training network based on self-supervision, and a random mask strategy is implemented on the input RCS sequence training data, so that the pre-training network based on self-supervision reconstructs the masked content through context perception to complete pre-training; the cascade structure of the multi-head autocorrelation mechanism and the feedforward neural network is adopted in the Transformer-based RCS prediction network to capture different feature spaces in parallel through multiple autocorrelation heads, realize dependency discovery and information aggregation at the sequence level, and enhance the ability of modeling complex nonlinear patterns in the RCS sequence training data through the feedforward neural network.

[0023] In the embodiment of the present invention, the self-supervised pre-training network is as follows: Figure 2 As shown in the figure, it includes the sequentially connected input layer, sequence block and normalization layer, random mask layer, position encoding and high-dimensional projection layer, Transformer-based RCS prediction network, and linear output layer. The functions of each layer of the self-supervised pre-training network are as follows: Input layer: The input is the RCS information of a target in radar tracking mode, that is, the multi-dimensional RCS sequence training data, and the sequence length is L; Sequence block and normalization layer: First, the input RCS sequence training data is normalized to a standard sequence of 0 to 1, and then blocked according to the window size P = 16 and the step size S = 16 to form N sequence blocks; Random mask layer: The N sequence blocks generated by the sequence block and normalization layer are randomly set to zero with a probability of 30%, forming 0.3*N zero-value sequence blocks; Position encoding and high-dimensional projection layer: First, position encoding is performed on each sequence block output by the random mask layer. Specifically, the encoding matrix PE is generated using sine / cosine encoding. Then, the sequence block is mapped to the Transformer input latent space with a dimension of 256 through a trainable linear parameter matrix Wp. The matrix is ​​bit-wise added to the position encoding to form the input data of the Transformer-based RCS prediction network, resulting in N = L / 16 256-dimensional sequence blocks. Transformer-based RCS prediction network: This network has an encoder-decoder structure. The encoder contains N encoders and the decoder contains N decoders, for example, N=6. Each encoder contains one multi-head autocorrelation mechanism and one feedforward neural network, and each decoder contains two multi-head autocorrelation mechanisms and one feedforward neural network. The multi-head autocorrelation mechanism can be, for example, an 8-head autocorrelation mechanism.

[0024] Linear output layer: The output dimension of the Transformer-based RCS prediction network is 256*N, and the final linear projection is a 256*16 linear layer. Therefore, the self-supervised pre-trained network will ultimately output 16*N sequence reconstruction content.

[0025] Furthermore, in order to adapt to the radar target RCS prediction process and solve the problem that the self-attention mechanism of the standard Transformer is too computationally expensive in long-term predictions due to the quadratic complexity of the sequence length, the present invention replaces the self-attention mechanism with an autocorrelation mechanism to perform dependency discovery and information aggregation at the sequence level. This mechanism avoids the problem of low information utilization due to the sparse attention mechanism, and can improve both computational efficiency and information utilization. Its core structure is composed of two symmetrical modules, an encoder and a decoder. The encoder is responsible for abstracting the input sequence (RCS sequence training data) layer by layer into a high-dimensional representation containing global semantics, while the decoder converts the high-dimensional representation into a target sequence through a dynamic interaction mechanism. Both the encoder and the decoder are composed of, for example, a stack of 6 layers with the same structure but independent parameters. The specific network structure is as follows: The Transformer-based RCS prediction network is composed of two symmetrical modules, an encoder and a decoder, stacked together; the encoder includes N sub-encoders connected in sequence, and the decoder includes N sub-decoders connected in sequence. The output of the Nth sub-encoder is connected to each sub-decoder, and the output of the Nth sub-decoder is used as the final output of the Transformer-based RCS prediction network after passing through the linear output layer, where N is an integer greater than 0; wherein each sub-encoder includes a multi-head autocorrelation module, a residual connection layer, a feedforward neural network layer, and a residual connection layer connected in sequence. The input of the multi-head autocorrelation module in the first sub-encoder includes word embedding, position encoding, and angle encoding; the first residual connection layer in the nth sub-encoder is also connected to the input of the multi-head autocorrelation module in the nth sub-encoder, and the nth The second residual connection layer in the nth sub-encoder is also connected to the first residual connection layer in the nth sub-encoder, where n ranges from 0 to N; each sub-decoder includes a multi-head autocorrelation module, a residual connection layer, a multi-head autocorrelation module, a residual connection layer, a feedforward neural network layer, a residual connection layer and a linear output layer connected in sequence, and the input of the first sub-decoder includes word embedding and position encoding; the first residual connection layer in the nth sub-decoder is also connected to the input of the multi-head autocorrelation module in the nth sub-decoder, the second residual connection layer in the nth sub-decoder is also connected to the first residual connection layer in the nth sub-decoder, and the third residual connection layer in the nth sub-decoder is also connected to the second residual connection layer in the nth sub-decoder; the input of the second multi-head autocorrelation module in the nth sub-decoder is also connected to the output of the Nth sub-encoder. Figure 3 and Figure 4 The network structure of the Transformer-based RCS prediction network when N=6 is shown. (1) Position coding Typically, the input of a word in a Transformer model is obtained by adding word embedding and position encoding, where word embedding maps the input symbol from a low dimension to a high-dimensional vector. The position encoding of the Transformer model is the core mechanism for solving the disorder of the autocorrelation mechanism. Its function is to give each position in the input sequence a unique vector representation, so that the Transformer model can perceive the sequential relationship of the elements. Since the autocorrelation mechanism itself does not contain temporal information, the position encoding enables the Transformer model to distinguish the differences between different time steps or positions by explicitly injecting position features. In the Transformer model, the commonly used position encoding method is to use a combination of sine and cosine functions to generate the encoding matrix PE. The specific position encoding calculation method is as follows: ; in, Indicates The position code at Indicates The position code at Indicates the relative position of the input symbol in the input sequence, represents the dimension of the high-dimensional mapping, Represents the absolute position of the vector after mapping, represents the sine function, In the embodiment of the present invention, the position code of the multi-head autocorrelation module input in the first sub-encoder and the position code of the first sub-decoder input are both obtained by encoding and calculating the input RCS sequence using sine and cosine functions.

[0026] (2) Autocorrelation mechanism Figure 5 This is a schematic diagram of the implementation process of an autocorrelation mechanism provided by an embodiment of the present invention. According to random process theory, similar phases of different periods usually exhibit similar sub-processes. Therefore, the inherent periodicity of this sequence can be used to design an autocorrelation mechanism. This includes period-based dependency discovery and time delay information aggregation. For period-based dependency, according to random process theory, for real discrete time series, , its autocorrelation coefficient It can be expressed as follows: ; Among them, the autocorrelation coefficient Represents a discrete time series and its delayed discrete time series The similarity between As the estimated cycle length The unnormalized confidence level is selected, and the one with the largest autocorrelation coefficient is selected. Cycle length Therefore, the period-based dependencies can be derived from these estimated periods and weighted by the corresponding autocorrelation values.

[0027] In order to connect the subsequences between estimation periods, the autocorrelation mechanism aggregates similar subsequence information through the time delay aggregation module. The time delay aggregation module aggregates similar subsequence information based on the selected time delay. The value information is cyclically operated, which can align similar subsequences with the same phase position in the estimation cycle. Finally, Normalized confidence aggregate subsequence. In the single-head scenario, for a length of Time series , after projection, we get the query ,key Sum Therefore, this mechanism can seamlessly replace the self-attention mechanism. The single-head autocorrelation mechanism can be expressed as: ; ; ; in, Indicates the first Individual queries, Indicates the first The key of the head, Indicates the first The value of the head, Express Perform delay operation and the delay is , that is, the part beyond the first position is looped to the end, Indicates the number of delays, , is the empirical hyperparameter, Indicates the length of the input RCS sequence, Indicates a round-down operation. express and In the delay The autocorrelation value at express and In the delay The autocorrelation value at represents the normalization operation, Express The normalized autocorrelation value is Express The normalized autocorrelation value is Indicates delay The value is 1~ L When, from L indivual and In the delay Get the maximum autocorrelation value when k The time delay corresponding to each autocorrelation value is used to avoid selecting irrelevant or even opposite phases.

[0028] When the single-head autocorrelation mechanism is extended to the multi-head autocorrelation mechanism, assuming that the number of latent variable channels is And the number of heads is , No. The query, key and value of each header are The implementation mechanism of the multi-head autocorrelation module is expressed as: ; ; in, Represents a query for the multi-head autocorrelation module, represents the key of the multi-head autocorrelation module, represents the value of the long-term autocorrelation module, represents the weight of each head in the multi-head autocorrelation module, Represents a splicing operation, Indicates the first The output of the head, Indicates the number of heads of the multi-head autocorrelation module.

[0029] For period-based dependencies, which point to subprocesses with the same phase position in the underlying period, which are sparse in nature, the time delay with the largest autocorrelation coefficient is selected to avoid selecting opposite phases. The length is The complexity of the multi-head autocorrelation mechanism is For the calculation of autocorrelation coefficient, given the input sequence In the case of , the autocorrelation coefficient can be calculated by fast Fourier transform (FFT) based on the Wiener-Schinchin theorem: ; Among them, the delay , Indicates FFT operation, express The inverse FFT transform, express The conjugate operation of Represented as an input sequence The frequency spectrum of Represents the input sequence The corresponding frequency. Because all delays The autocorrelation coefficient All can be calculated by FFT at one time, so the autocorrelation mechanism is realized Complexity.

[0030] Unlike the point-by-point self-attention mechanism, the autocorrelation mechanism presents sequence-level connections. Specifically, in terms of temporal dependency discovery, the autocorrelation mechanism discovers dependencies between subsequences based on periodicity, while the self-attention mechanism only calculates the correlation between scattered points.

[0031] (3) Feedforward supplement mechanism The Transformer model introduces a feedforward neural network layer to perform nonlinear transformation on the representation of each position to supplement the global interaction information of the multi-head autocorrelation module and enhance the local feature expression ability. The feedforward neural network layer consists of two fully connected layers, and its mathematical form is: ; in, 、 is the learnable weight matrix, 、 is a learnable bias parameter. The first fully connected layer uses The activation function introduces nonlinearity, and the second fully connected layer is a linear projection. is the input of the feedforward neural network layer, and the output matrix of the feedforward neural network layer is Dimensions and input consistent.

[0032] (4) Residual connection mechanism The residual connection layer consists of two parts: the residual connection layer and the normalization layer. Each sub-encoder and each sub-decoder contains the residual connection after the multi-head autocorrelation module and the residual connection after the feedforward neural network layer. The calculation process is as follows: ; ; in, is the input of the multi-head autocorrelation module, is the output of the feedforward neural network layer, The input to the multi-head autocorrelation module is directly added to the feedforward neural network layer because its outputs and input dimensions are consistent. The residual connection mechanism constructs an identity mapping path by directly adding the sub-layer inputs and outputs, enabling the network to focus on the residual information difference between input and output, effectively alleviating the vanishing gradient problem. The normalization layer standardizes the superimposed features (mean and variance normalization), accelerating model convergence and improving training stability.

[0033] The embodiment of the present invention adopts self-supervised learning. Self-supervised learning is the core method of learning unlabeled data representation. The present invention extends this method to the RCS prediction task. The present invention implements a random masking strategy on the input RCS sequence using a random masking layer based on a self-supervised pre-training network, so that the self-supervised pre-training network reconstructs the masked content through context perception to complete pre-training. Finally, the Transformer-based RCS prediction network in the trained self-supervised pre-training network is used as the initial Transformer-based RCS prediction network. Through self-supervised pre-training, the generalization ability of the Transformer-based RCS prediction network can be improved.

[0034] S30. Introducing the consideration of attitude angle into the RCS sequence training data, generating new RCS sequence training data, and further training the initial Transformer-based RCS prediction network using the new RCS sequence training data to obtain a trained Transformer-based RCS prediction network.

[0035] It can be seen from the RCS characteristics that the RCS value shows a significant correlation with the attitude angle. Based on this characteristic, the present invention adopts a data enhancement strategy: the attitude angle is introduced into the RCS sequence training data, and the attitude angle-RCS association mapping is established by discretizing the attitude angle quantization features (azimuth angle resolution 0.1°), and the angle stamp is added to the original input RCS sequence training data. According to the 180° change range of the target azimuth, an 1800-dimensional attitude angle quantization label is generated (1~1800 index corresponds to 0.1°-180.0° azimuth change). In the data preprocessing stage, the time series feature processing method in time series prediction is used for reference, and the attitude angle corresponding to the RCS sequence training data is converted through sine / cosine angle coding to obtain the angle coding, and the position coding and data feature coding (word embedding) are tensor-added to generate new RCS sequence training data. Figure 6 As shown, this serves as the multidimensional input vector for the first sub-encoder in the Transformer-based RCS prediction network. By integrating multiple features such as attitude angles, position encoding, and data feature encoding, this embodiment of the present invention can effectively improve the prediction capability of complex time series and provide a more robust time series prediction model for radar multi-target tracking.

[0036] The new RCS sequence training data was then used to further train the initial Transformer-based RCS prediction network obtained in S20 to obtain a trained Transformer-based RCS prediction network for RCS prediction during testing. The Transformer model not only overcomes the linearity and stationarity assumptions of traditional methods but also allows for flexible adaptation to RCS prediction requirements in different environments by adjusting the number of autocorrelation heads and stacking layers.

[0037] This embodiment of the present invention uses a Transformer-based RCS prediction network to perform RCS prediction. First, to address non-stationarity, the Transformer's position encoding (PE) explicitly injects temporal position information, addressing the shortcomings of traditional models that rely solely on time lags. This allows the autocorrelation mechanism to capture long-term dependencies and dynamic changes in the sequence. Second, to handle nonlinear relationships, the Transformer, through a multi-head autocorrelation mechanism and a cascaded structure of a feedforward neural network (FFN), can model complex nonlinear patterns in RCS sequences, such as periodicity or sudden events. Specifically, the multi-head autocorrelation submodule uses multiple autocorrelation heads to capture patterns in different feature spaces in parallel. For example, some heads can learn periodic RCS fluctuations (such as periodic reflection enhancement caused by target rotation), while other heads can identify sudden interference (such as instantaneous scattering changes caused by enemy chaff decoys). Finally, the Transformer can process external variables, such as attitude angles, by inputting multivariate time series data. These variables can be integrated into the model through embedding or splicing to enhance the prediction capabilities of complex time series.

[0038] It should be noted that during the training process of the self-supervised pre-trained network, the input RCS sequence is the RCS sequence output after position encoding and high-dimensional projection layers. This RCS sequence also needs to first introduce the consideration of the posture angle to generate a new RCS sequence, and then input the new RCS sequence into the Transformer-based RCS prediction network.

[0039] S40, introducing the consideration of the attitude angle into the RCS sequence data to be predicted, generating new RCS sequence data to be predicted, and inputting the new RCS sequence data to be predicted into the trained Transformer-based RCS prediction network to obtain a prediction result.

[0040] Before the RCS sequence data to be predicted is input into the Transformer-based RCS prediction network, this embodiment of the present invention introduces attitude angle considerations into the RCS sequence data to be predicted, similar to the processing of the RCS sequence training data in S30, to generate new RCS sequence data to be predicted. This includes: converting the attitude angle corresponding to the RCS sequence data to be predicted using sine / cosine angle encoding, performing tensor addition on the attitude angle, the position encoding, and the data feature encoding to generate the new RCS sequence data to be predicted. Next, this new RCS sequence data to be predicted is input into the trained Transformer-based RCS prediction network to obtain a prediction result.

[0041] It should be noted that the training of the network in S20 and S30 can adopt existing training methods, such as using the back propagation algorithm to iteratively update the parameters of the network, which will not be described in detail here.

[0042] In order to verify the effectiveness of the RCS prediction method based on the Transformer model provided by the embodiment of the present invention, the following experiments were conducted for verification.

[0043] 1. Simulation conditions During model training, the dataset was divided into a training set and a validation set in a ratio of 8:2. A teacher-forcing training strategy was employed during model training. The encoder interface of the Transformer-based RCS prediction network inputs a portion of the true RCS values ​​to facilitate rapid network convergence. To prevent overfitting, an early stopping mechanism (EarlyStop) was employed during training. Training was terminated if the loss function in the test set did not decrease after 30 cycles. The main parameters for Transformer model network training are shown in Table 1.

[0044] Table 1 Main parameters of network training of Transformer model

[0045] This experiment performs time series prediction on two RCS sequences with a length of 5000. The first 4000 sequence data are used as training data, and the last 1000 sequence data are used as test data. The first RCS sequence is a stationary sequence with autocorrelation. The probability distribution obeys the Swerling I model and satisfies the RCS mean. The second RCS sequence is a segment of measured RCS data for a certain aircraft type. This data has obvious trends and seasonality, and meets the non-stationary condition.

[0046] 2. Simulation content and result analysis See Figure 7(a) to Figure 7(b) 、 Figure 8(a) to Figure 8(b) : Figure 7(a) shows the true value of RCS sequence 1 and the corresponding visualization of the Transformer-based RCS prediction experimental results. In Figure 7(a), the horizontal axis represents the number of prediction steps and the vertical axis represents the RCS value (unit: m 2 ), Figure 7(b) is a visualization of the RCS prediction relative error between the true value and the predicted value of RCS sequence 1. In Figure 7(b), the horizontal axis represents the number of prediction steps, and the vertical axis represents the RCS prediction relative error (unit %). Figure 8(a) is a visualization of the true value of RCS sequence 2 and the corresponding Transformer-based RCS prediction experimental results. In Figure 8(a), the horizontal axis represents the number of prediction steps, and the vertical axis represents the RCS value (unit m 2), Figure 8(b) shows a visualization of the RCS prediction relative error between the true and predicted values ​​of RCS sequence 2. In Figure 8(b), the horizontal axis represents the number of prediction steps, and the vertical axis represents the RCS prediction relative error (unit: %). To quantitatively describe the RCS prediction error and its distribution, the RCS prediction error was statistically analyzed on the test set. The error statistics for the RCS prediction values ​​are shown in Table 2.

[0047] Table 2 Statistics of error values ​​of RCS prediction

[0048] Simulation results show that the Transformer-based model's predictions are generally able to identify the changing trends of the RCS series and keep the prediction error within a small range, demonstrating the superior performance of the model. For RCS series 1, the prediction results show that the predicted values ​​and the true values ​​follow the same trend, but the amplitudes deviate. This indicates that the model can capture temporal patterns but lacks response to peaks or rapid changes. For RCS series 2, the error plot shows that the error increases significantly in the middle of the series, reflecting the model's inadequate capture of the time series trend. The MSE for RCS series 1 is 0.4755, while the MSE for RCS series 2 is 0.5019. While the two values ​​are similar, RCS series 1 has a slightly better MSE. This indicates that the model fails to accurately predict the significant downward trend in RCS series 2. Due to the larger absolute value of the RCS in RCS series 2, the relative error distribution of RCS series 2 is better (with a higher proportion of low-error intervals), suggesting that the MSE may be affected by the uneven distribution of the series data. In order to further demonstrate the performance advantage of the Transformer model of the present invention, MSE is selected as the prediction error indicator, and the prediction results are compared with those of other traditional models as shown in Table 3.

[0049] Table 3 Comparison of MSE of RCS prediction experimental models

[0050] Table 3 shows the MSE performance of different models in the RCS prediction task. The Transformer model achieved MSEs of 0.4755 and 0.5019 on two RCS sequences, significantly outperforming other models and improving by 19.84% and 16.99%, respectively, over the next-best model, the LSTM (Long Short-Term Memory). Compared to traditional autoregressive (AR) models and Markov (a stochastic process model based on Markov properties), the Transformer achieved a 32% to 37% reduction in MSE. Experimental results demonstrate the superiority of the Transformer in RCS prediction tasks. Its core advantage stems from the powerful predictive power of its multi-head self-attention mechanism for complex RCS temporal dynamics.

[0051] In summary, the RCS prediction method based on the Transformer model proposed in the embodiment of the present invention introduces a random mask layer in the pre-training network based on self-supervision, and implements a random mask strategy on the input RCS sequence data, so that the model reconstructs the masked content through context perception to complete pre-training, thereby avoiding the defect of weak generalization ability of the model, improving the prediction accuracy of downstream RCS prediction tasks, and the generalization ability of different RCS sequences; the RCS prediction network based on the Transformer model proposed in the present invention adopts a multi-head autocorrelation mechanism and a cascade structure of a feedforward neural network, and its multi-head autocorrelation submodule can be used through multiple autocorrelations. The present invention captures patterns of different feature spaces in parallel at key points, and the feedforward neural network enhances the ability to model complex nonlinear patterns in the RCS sequence, which significantly improves the RCS prediction ability. The present invention takes into account the strong correlation between the radar target scattering characteristics and the target motion attitude angle, and establishes an attitude angle-RCS correlation mapping. The angle stamp is added to the original RCS sequence data, which improves the prediction ability of complex time series and provides a more robust time series prediction model for radar multi-target tracking. The present invention is highly compatible with the existing technology, abandons the traditional sequential processing methods of recurrent neural networks and convolutional neural networks, can be calculated in parallel, greatly improves the training speed, and makes up for the shortcomings of the existing technology.

[0052] Second, see Figure 9 , an embodiment of the present invention provides an RCS prediction device based on a Transformer model, the RCS prediction device comprising: Data construction module, used to construct RCS sequence training data; The first network training module is used to construct a self-supervised pre-training network, train the self-supervised pre-training network using RCS sequence training data to obtain a trained self-supervised pre-training network, and use the Transformer-based RCS prediction network in the trained self-supervised pre-training network as the initial Transformer-based RCS prediction network; wherein, a random mask layer is introduced into the self-supervised pre-training network, and a random masking strategy is implemented on the input RCS sequence training data, so that the self-supervised pre-training network reconstructs the masked content through context perception to complete pre-training; the Transformer-based RCS prediction network adopts a multi-head autocorrelation mechanism and a cascade structure of a feedforward neural network to capture different feature spaces in parallel through multiple autocorrelation heads, realize dependency discovery and information aggregation at the sequence level, and enhance the ability to model complex nonlinear patterns in RCS sequence training data through the feedforward neural network; The second network training module is used to introduce the consideration of attitude angles into the RCS sequence training data, generate new RCS sequence training data, and use the new RCS sequence training data to further train the initial Transformer-based RCS prediction network to obtain a trained Transformer-based RCS prediction network; The prediction module is used to introduce the consideration of attitude angle into the RCS sequence data to be predicted, generate new RCS sequence data to be predicted, and input the new RCS sequence data to be predicted into the trained Transformer-based RCS prediction network to obtain the prediction result.

[0053] As for the device embodiment of the second aspect, since it is basically similar to the method embodiment of the first aspect, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment of the first aspect.

[0054] In the description of the present invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.

[0055] Although the present invention is described herein in conjunction with various embodiments, those skilled in the art may understand and implement other variations of the disclosed embodiments by reviewing the specification and accompanying drawings in the process of implementing the claimed invention. In the specification, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple components or steps. The fact that certain measures are described in different embodiments does not mean that these measures cannot be combined to produce good results.

[0056] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. A RCS prediction method based on a Transformer model, characterized in that: The method comprises: Construct RCS sequence training data; A self-supervised pre-training network is constructed, and the self-supervised pre-training network is trained using the RCS sequence training data to obtain a trained self-supervised pre-training network, and the Transformer-based RCS prediction network in the trained self-supervised pre-training network is used as the initial Transformer-based RCS prediction network; wherein, a random mask layer is introduced into the self-supervised pre-training network, and a random masking strategy is implemented on the input RCS sequence training data, so that the self-supervised pre-training network reconstructs the masked content through context perception to complete pre-training; a multi-head autocorrelation mechanism and a cascade structure of a feedforward neural network are adopted in the Transformer-based RCS prediction network, so that different feature spaces are captured in parallel through multiple autocorrelation heads, dependency discovery and information aggregation are realized at the sequence level, and the ability to model complex nonlinear patterns in the RCS sequence training data is enhanced through the feedforward neural network; Introducing attitude angle considerations into the RCS sequence training data to generate new RCS sequence training data, and further training the initial Transformer-based RCS prediction network using the new RCS sequence training data to obtain a trained Transformer-based RCS prediction network; The attitude angle is introduced into the RCS sequence data to be predicted to generate new RCS sequence data to be predicted, and the new RCS sequence data to be predicted is input into the trained Transformer-based RCS prediction network to obtain the prediction result.

2. The RCS prediction method based on the Transformer model according to claim 1, characterized in that: The RCS sequence training data and the RCS sequence data to be predicted both include a stationary RCS sequence that satisfies a specific distribution probability and a non-stationary RCS sequence of a high-maneuverability target.

3. The RCS prediction method based on the Transformer model according to claim 1, characterized in that: The self-supervised pre-training network includes an input layer, a sequence block and normalization layer, a random mask layer, a position encoding and high-dimensional projection layer, a Transformer-based RCS prediction network, and a linear output layer connected in sequence.

4. The RCS prediction method based on the Transformer model according to claim 3, characterized in that: The Transformer-based RCS prediction network consists of two symmetrical modules, an encoder and a decoder. The encoder includes N sub-encoders connected in sequence, and the decoder includes N sub-decoders connected in sequence. The output of the Nth sub-encoder is connected to each sub-decoder. The output of the Nth sub-decoder is passed through a linear output layer as the final output of the Transformer-based RCS prediction network. N is an integer greater than 0. Each sub-encoder includes a multi-head autocorrelation module, a residual connection layer, a feedforward neural network layer, and a residual connection layer connected in sequence. The input of the multi-head autocorrelation module in the first sub-encoder includes word embedding, position encoding, and angle encoding; the first residual connection layer in the n-th sub-encoder is also connected to the input of the multi-head autocorrelation module in the n-th sub-encoder, and the second residual connection layer in the n-th sub-encoder is also connected to the first residual connection layer in the n-th sub-encoder, where n ranges from 0 to N; Each sub-decoder includes a multi-head autocorrelation module, a residual connection layer, a multi-head autocorrelation module, a residual connection layer, a feedforward neural network layer, a residual connection layer, and a linear output layer connected in sequence. The input of the first sub-decoder includes word embedding and position encoding. The first residual connection layer in the n-th sub-decoder is also connected to the input of the multi-head autocorrelation module in the n-th sub-decoder, the second residual connection layer in the n-th sub-decoder is also connected to the first residual connection layer in the n-th sub-decoder, and the third residual connection layer in the n-th sub-decoder is also connected to the second residual connection layer in the n-th sub-decoder. The input of the second multi-head autocorrelation module in the nth sub-decoder is also connected to the output of the Nth sub-encoder.

5. The RCS prediction method based on the Transformer model according to claim 4, characterized in that: The implementation mechanism of the multi-head autocorrelation module is expressed as: ； ； ； ； ； in, Represents a query for the multi-head autocorrelation module, represents the key of the multi-head autocorrelation module, represents the value of the long-term autocorrelation module, represents the weight of each head in the multi-head autocorrelation module, Represents a splicing operation, represents the number of heads in the multi-head autocorrelation module, Indicates the first The output of the head, Indicates the first Individual queries, Indicates the first The key of the head, Indicates the first The value of the head, Express Perform delay operation and the delay is , Indicates the number of delays, , is the empirical hyperparameter, Indicates the length of the input RCS sequence, Indicates a round-down operation. express and In the delay The autocorrelation value at express and In the delay The autocorrelation value at represents the normalization operation, Express The normalized autocorrelation value is Express The normalized autocorrelation value is Indicates delay The value is 1~ L When, from L indivual and In the delay Get the maximum autocorrelation value when k The time delay corresponding to the autocorrelation value.

6. The RCS prediction method based on the Transformer model according to claim 4, characterized in that: The position coding of the multi-head autocorrelation module input in the first sub-encoder and the position coding of the first sub-decoder input are both obtained by encoding the input RCS sequence using sine and cosine functions; wherein, In the training process of the self-supervised pre-trained network, the input RCS sequence is the RCS sequence output after position encoding and high-dimensional projection layer; During the initial training process of the Transformer-based RCS prediction network, the input RCS sequence is the new RCS sequence training data.

7. The RCS prediction method based on the Transformer model according to claim 1, characterized in that: Introducing attitude angle considerations into the RCS sequence training data to generate new RCS sequence training data, including: The posture angle corresponding to the RCS sequence training data is converted through sine / cosine angle coding, and then tensor-added with the position coding and data feature coding to generate new RCS sequence training data.

8. The RCS prediction method based on the Transformer model according to claim 1, characterized in that: The attitude angle is introduced into the RCS sequence data to be predicted to generate new RCS sequence data to be predicted, including: The angle code obtained by converting the posture angle corresponding to the RCS sequence data to be predicted through sine / cosine angle coding is tensor-added with the position code and the data feature code to generate new RCS sequence data to be predicted.

9. A RCS prediction device based on a Transformer model, characterized in that: The RCS prediction device includes: Data construction module, used to construct RCS sequence training data; The first network training module is used to construct a self-supervised pre-training network, use the RCS sequence training data to train the self-supervised pre-training network to obtain a trained self-supervised pre-training network, and use the Transformer-based RCS prediction network in the trained self-supervised pre-training network as the initial Transformer-based RCS prediction network; wherein, a random mask layer is introduced into the self-supervised pre-training network, and a random masking strategy is implemented on the input RCS sequence training data, so that the self-supervised pre-training network reconstructs the masked content through context perception to complete pre-training; the Transformer-based RCS prediction network adopts a cascade structure of a multi-head autocorrelation mechanism and a feedforward neural network to capture different feature spaces in parallel through multiple autocorrelation heads, realize dependency discovery and information aggregation at the sequence level, and enhance the ability to model complex nonlinear patterns in the RCS sequence training data through the feedforward neural network; a second network training module, configured to introduce attitude angle considerations into the RCS sequence training data, generate new RCS sequence training data, and further train the initial Transformer-based RCS prediction network using the new RCS sequence training data to obtain a trained Transformer-based RCS prediction network; The prediction module is used to introduce the consideration of attitude angle into the RCS sequence data to be predicted, generate new RCS sequence data to be predicted, and input the new RCS sequence data to be predicted into the trained Transformer-based RCS prediction network to obtain the prediction result.

Images