Non-cooperative target recognition and micro-motion parameter extraction method based on consistent VAE and transformer

Abstract

The application provides a non-cooperative target recognition and micro-motion parameter extraction method based on consistency VAE and a transformer, and belongs to the technical field of radar signal processing. The application is used to solve the problems that the existing CNN-based privileged information guidance network is easy to distort features, lacks global timing modeling capability and the multi-task guidance mechanism is rigid in strong noise. The steps of the application include: S1: acquiring a time-frequency graph; S2: inputting the time-frequency graph into a consistency variational autoencoder to perform denoising and feature dimension reduction; S3: global timing modeling and noise-resistant coding based on a transformer; and S4: designing a cascaded transformer structure to obtain high-precision target type recognition results and micro-motion parameter regression results.

Description

Technical Field

[0001] This invention relates to a method for non-cooperative target recognition and micro-motion parameter extraction based on consistent VAE and Transformer, belonging to the field of radar signal processing technology. Background Technology

[0002] Identification and micro-motion parameter extraction of non-cooperative targets (such as ballistic missile decoys and space debris) are key challenges in radar signal processing. The micro-Doppler effect carries information about the target's fine structure and motion attitude, which is typically processed by converting radar echoes into time-frequency maps through time-frequency analysis. With the development of artificial intelligence technology, deep learning-based methods have gradually replaced traditional manual feature extraction methods, becoming the mainstream research direction in this field.

[0003] Current technologies for signal denoising and feature reduction, in order to cope with noise interference in radar echoes, have widely adopted traditional denoising algorithms such as digital filtering and wavelet transform, as well as deep denoising methods based on convolutional neural networks (CNNs) and variational autoencoders (VAEs). However, these existing technologies have significant drawbacks: First, most deep denoising models fall into the category of "blind denoising" or unsupervised learning, failing to fully utilize the rich "noise level label" information in simulation data. In practical applications, signals with different signal-to-noise ratios require different denoising strategies. Existing models cannot adaptively adjust the denoising intensity according to noise level conditions, resulting in severe feature loss under strong noise or incomplete denoising under weak noise. Second, standard VAEs and their simple variants lack constraints on the consistency of the latent space. Under noise interference, the latent coding distribution of the same target varies greatly under different noise levels, resulting in poor stability of the reconstructed features and difficulty in meeting the requirements for high-precision parameter extraction.

[0004] Regarding existing network architectures for target recognition and parameter extraction, in the feature extraction and task execution stages, current technologies mainly employ convolutional neural networks, recurrent neural networks (RNN / LSTM), or the basic Transformer architecture. In recent years, some research has attempted to introduce the Transformer into radar signal processing, utilizing its self-attention mechanism to capture global features. In terms of multi-task processing, current mainstream solutions typically employ parallel multi-task learning frameworks, sharing a low-level feature extractor and connecting a classification head and a regression head in parallel at the backend, simultaneously outputting target type and micro-motion parameters. However, the aforementioned existing technologies still have the following shortcomings when processing micro-motion parameter extraction for non-cooperative targets: First, although the Transformer possesses powerful temporal modeling capabilities, existing applications are mostly direct end-to-end mappings, without architectural optimization based on the physical mechanisms of radar micro-motion characteristics. Especially for parameters such as precession angle, which are difficult to observe directly from time-frequency diagrams, relying solely on data-driven black-box mapping results in poor generalization ability at low signal-to-noise ratios. Second, existing parallel multi-task frameworks sever the inherent physical logic between "target type" and "micro-motion parameters." In physical reality, the type of a target determines its possible micro-motion patterns and parameter ranges. Existing technologies allow classification and regression tasks to operate independently, failing to utilize the relatively easily and accurately obtained "target type" information to guide and constrain the solution space of the difficult-to-extract "micro-motion parameters." This lack of prior knowledge guidance results in limited parameter estimation accuracy in complex environments and a tendency to produce erroneous predictions that contradict physical common sense. Summary of the Invention

[0005] To address the problems of existing privileged information guidance networks based on CNNs, such as feature distortion under strong noise, lack of global temporal modeling capabilities, and rigid multi-task guidance mechanisms, this invention proposes a non-cooperative target recognition and micro-motion parameter extraction method based on consistent VAE and Transformer.

[0006] The technical solution adopted by the present invention to solve the above problems is as follows: The present invention includes the following steps: Step 1: Obtain the time-frequency diagram based on the radar echo signal; Step 2: Perform phased training on the consistent variational autoencoder and the Transformer encoder; Step 3: Input the time-frequency graph into the consistency variational autoencoder. The encoder maps the time-frequency graph to the latent space and uses the consistency loss constraint to make the latent feature vectors of the same target at different noise levels tend to be consistent, thus obtaining the denoised low-dimensional latent feature vectors. Step 4: Divide the denoised low-dimensional latent feature vector into a feature slice sequence along the time axis, and inject position encoding into each feature slice to obtain a token sequence with temporal information; Step 5: Concatenate the token sequence with the learnable target type-specific token and the micro-motion parameter-specific token to construct a joint input sequence; Step 6: Input the joint input sequence into the Transformer encoder, and use the self-attention mechanism to enable the micro-motion parameter-specific token to interact with the target type-specific token and feature slices to obtain the parameter feature representation of the fused category prior information; Step 7: Obtain the target type recognition result based on the output corresponding to the target type-specific token; Step 8: Obtain the regression results of the micro-motion parameters based on the parameter characteristics.

[0007] Furthermore, step 1 includes: The radar echo signal is transformed into a time-frequency domain representation by performing time-frequency transformation, thereby obtaining a time-frequency map.

[0008] Furthermore, step 3 specifically includes: The clean signal sample and the noisy signal sample are input into the unified encoder with shared weights to obtain the first latent feature vector and the second latent feature vector, respectively. The consistency loss is obtained based on the mean squared error between the first and second latent feature vectors. The first latent feature vector and the second latent feature vector are respectively input into a unified decoder with shared weights to obtain the first reconstructed signal and the second reconstructed signal; The reconstruction loss is obtained based on the differences between the first and second reconstructed signals and the clean signal samples. Based on the consistency loss and reconstruction loss, obtain the total loss and update the weight parameters of the unified encoder and unified decoder according to the total loss; The trained unified encoder is used as the encoder of the consistent variational autoencoder. The time-frequency image to be tested is input into the encoder to obtain the denoised low-dimensional latent feature vector.

[0009] Further, obtaining the first latent feature vector and the second latent feature vector includes: Noisy signal samples and clean signal samples are mapped to the internal vector space respectively, and feature recombination and nonlinear transformation are performed in sequence. The features after nonlinear transformation are mapped to mean vector and variance vector respectively, and the first latent feature vector and the second latent feature vector are obtained through reparameterized sampling.

[0010] Furthermore, step 4 specifically includes: The low-dimensional latent feature vector is transformed from an image domain representation into a series of feature slices arranged along the time axis. Location encoding information is added to each feature slice to preserve the temporal physical evolution of radar signals and obtain a token sequence with temporal information.

[0011] Furthermore, step 5 specifically includes: By placing the target type-specific token and the micro-motion parameter-specific token at the beginning of the token sequence, the target type-specific token and the micro-motion parameter-specific token are preferentially associated with the feature slices in subsequent attention calculations to obtain the joint input sequence.

[0012] Furthermore, step 6 specifically includes: The joint input sequence is fed into a multi-layer Transformer encoder, and the attention weight between the micro-motion parameter-specific token and the target type-specific token is calculated through a multi-head self-attention mechanism, so that the category semantic information carried by the target type-specific token is transferred to the micro-motion parameter-specific token. Based on the attention weights, obtain the parametric feature representation of the prior information of the fusion category.

[0013] Furthermore, step 7 specifically includes: Extract the vector corresponding to the position of the target type-specific token in the output sequence of the Transformer encoder; The vector corresponding to the target type's unique token is input into the classification task head, and the target type recognition result is obtained through a fully connected layer and a Softmax activation function.

[0014] Furthermore, based on the parameter feature representation, the regression results of the micro-motion parameters are obtained, including: Extract the vector corresponding to the position of the micro-motion parameter-specific token in the output sequence of the Transformer encoder as the parameter feature representation; The parameter feature representation is input into the parameter regression head, and the regression results of the micro-motion parameters of precession frequency and precession angle are obtained through a fully connected regression layer.

[0015] Furthermore, the phased training in step 2 specifically includes: In the first stage, the parameters of the Transformer encoder are frozen, an initial learning rate is set, and the weights of the encoder and decoder in the consistent variational autoencoder are updated using the optimizer until convergence, based on the reconstruction loss. KL divergence loss and consistency loss Total loss of VAE construction ; In the second stage, the weights of the consistency variational autoencoder are frozen, and the noisy time-frequency map is input into the encoder to obtain latent denoising features, based on the classification loss. and regression loss Total loss of the construction task The Adam optimizer is used to update the embedding weights of the Transformer encoder, classification task head, parameter regression head, target type-specific token and micro-motion parameter-specific token, and gradient truncation is performed when backpropagating to the encoder output. VAE total loss The expression is: (1); Total mission losses The expression is: (2).

[0016] The beneficial effects of this invention are: 1. This invention significantly improves feature robustness and denoising performance in low signal-to-noise ratio (SNR) environments. Addressing the issues of weak echoes and strong noise interference from non-cooperative targets in space, this invention utilizes a consistent VAE module, leveraging rich noise level labels and paired samples in the simulation data to force latent feature vectors to maintain consistency under noise perturbation. Compared to traditional wavelet denoising and ordinary VAE methods, this invention effectively filters out background noise and fully preserves the fine structure of micro-Doppler features (such as periodic modulation fringes generated by precession) even in extreme environments with extremely low SNR, significantly reducing feature reconstruction errors and laying a solid feature foundation for subsequent high-precision parameter extraction.

[0017] 2. This invention overcomes the accuracy bottleneck in micro-motion parameter extraction, particularly addressing the problems of existing methods that separate classification and regression tasks and lack physical guidance for difficult-to-extract parameters. This invention proposes a hierarchical guidance mechanism. By transforming target type identification results into context vectors to guide parameter regression, the model can automatically utilize the physical correlation between "target type" and "micro-motion parameter" (e.g., after identifying it as a "cone-cylinder," automatically constraining the solution space of the precession angle). This substantially improves the extraction accuracy of noise-sensitive and difficult-to-observe parameters such as the precession angle, significantly reducing the root mean square error (RMSE) compared to traditional parallel multi-task networks, and effectively avoiding the generation of parameter predictions that violate physical principles.

[0018] 3. This invention achieves a balance between computational efficiency and long-range temporal modeling capabilities. Addressing the challenges of traditional CNNs in capturing long-term dependencies and their high computational cost, this invention employs a VAE dimensionality reduction + Transformer architecture. First, the high-dimensional time-frequency graph is compressed into a low-dimensional tensor using a VAE, significantly reducing the amount of data input to the Transformer and thus effectively lowering the overall computational complexity of the model. Second, utilizing the Transformer's self-attention mechanism, it can excellently capture the evolution patterns and global correlations of micro-motion signals over long time spans. This method ensures the feasibility of real-time system processing and improves the generalization ability to recognize complex micro-motion patterns (such as compound spin precession), making the model more compatible with input data of different sizes.

[0019] 4. This invention addresses the poor adaptability problem caused by the lack of prior templates for non-cooperative targets. To overcome the challenge of insufficient prior templates for non-cooperative targets, this invention constructs a data-driven end-to-end learning framework. Without the need to pre-construct precise target electromagnetic scattering templates, the model can automatically learn the distribution of micro-motion features of various targets from the data. Combined with consistent denoising and hierarchical guidance strategies, the model still exhibits excellent generalization performance when facing unknown complex noise environments or target structures with slight deformations, effectively improving the overall recognition rate and parameter estimation reliability of space situational awareness systems for unknown non-cooperative targets. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a non-cooperative target recognition and micro-motion parameter extraction method based on consistency VAE and Transformer. Figure 2 Schematic diagram of VAE noise reduction design; Figure 3 For consistent VAE pre-training flow graph; Figure 4 This is a schematic diagram of the Transformer's internal structure and noise reduction mechanism. Figure 5 This is a schematic diagram of time-series slicing and cross-domain bridging; Figure 6 This is a schematic diagram of time-series slicing and cross-domain bridging; Figure 7 This is a diagram illustrating joint mission training. Detailed Implementation

[0021] The overall model of this invention consists of three core modules connected in series: first, a consistent variational autoencoder module responsible for purifying low-level features; second, a Transformer module responsible for global temporal modeling; and finally, two cascaded task heads that implement "privileged information parameter guidance." The radar echo signal, after undergoing time-frequency transformation to obtain a time-frequency map, passes through the above modules sequentially, ultimately outputting high-precision target type and micro-motion parameters.

[0022] like Figure 1 As shown, the steps of the non-cooperative target recognition and micro-motion parameter extraction method based on consistent VAE and Transformer described in this embodiment include: S1: Obtain the time-frequency diagram; The radar echo signal is transformed into a time-frequency domain representation by performing time-frequency transformation, thereby obtaining a time-frequency map.

[0023] S2: Input the time-frequency graph into the consistency variational autoencoder for denoising and feature dimensionality reduction; To address the problems of overfitting and loss of micro-motion information due to rigid scaling when processing high-dimensional noisy time-frequency graphs using traditional CNNs, this invention designs a VAE module based on consistency constraints. For example... Figure 2 The diagram shown illustrates the noise reduction design concept of a consistent VAE.

[0024] This module is used to process the original radar time-frequency map to achieve noise suppression and feature compression.

[0025] Input / output definition: The input is a single-channel radar target time-frequency map with a resolution of 512×512. After processing by the encoder network, the output is a low-dimensional latent feature vector.

[0026] Encoder Structure: The encoder adopts a convolutional neural network architecture, which contains multiple convolutional downsampling blocks. Specifically, the feature map size is halved layer by layer through convolutional layers with a stride of 2, and the number of channels is doubled layer by layer, ultimately compressing the 512×512 input into a latent feature vector z of size 1×256.

[0027] Internal vector design: At the bottleneck layer of the encoder and decoder, an internal vector dimension of 512 is designed. The features output by the encoder are first mapped to this 512-dimensional internal space for feature reconstruction and nonlinear transformation, and then mapped to a 256-dimensional mean vector and variance vector through a fully connected layer for reparameterization sampling.

[0028] In the training phase of the model, this invention employs a phased training strategy. This step first introduces the training process of the VAE. The feature compression and denoising capabilities of the variational autoencoder are trained. The optimizer chosen is Adam, with an initial learning rate of 10. -4Freeze the parameters of the Transformer module and update only the encoder and decoder weights of the VAE. Minimize the VAE loss using noise labels from the simulation dataset. Through this stage, the model learns to map a noisy 512×512 time-frequency graph into a denoised 256-dimensional latent vector.

[0029] Consistency constraints of high-dimensional latent spaces: Figure 2 The left side shows time-frequency graph sequences of the same target at different noise levels and signal-to-noise ratios (SNRs). These time-frequency graphs with varying degrees of background noise are mapped into a high-dimensional latent space at the center of the graph by the encoder, forming multiple latent vectors. This invention introduces a consistency loss here, forcing all latent vectors of noisy signals to converge in the latent space towards the latent vectors of clean signals with high SNR.

[0030] Pure feature reconstruction: Figure 2 The right side illustrates the process of reconstructing a clean time-frequency map from a high signal-to-noise ratio latent vector that tends to be consistent in the latent space through the decoder. In the actual inference stage, only the trained consistent encoder is retained, and its output low-dimensional feature vectors are naturally free from the amplitude influence of noise, ensuring the purity of the underlying micro-motion features.

[0031] The pre-training process of a consistent VAE is as follows: Figure 3 As shown, the core of the training process lies in using paired data to achieve "weight sharing" and "latent space constraints." The specific steps are as follows: S201: Pairwise data sampling and forward propagation. In each epoch, the model simultaneously samples "clean signal samples" containing the same micro-motion characteristics from the simulation dataset. "and noisy signal samples" Subsequently, these two samples were fed into the shared weight model architecture. It is particularly important to emphasize the "unified encoder" in the diagram. "and "unified decoder" "Physically, they are unique network modules, and they use the exact same parameters." and The two signals are processed separately. The encoder reduces the dimensionality of the signals and outputs latent space vectors for each. and The decoder then reconstructs these two vectors into the output signal. ^ and ^ .

[0032] S202: Goal-guided loss aggregation. This invention breaks away from the conventional blind denoising logic of VAEs and designs a dual supervision mechanism: Guidelines for Pure Signal Reconstruction: This applies to the reconstruction results of pure samples. ^ Or the reconstruction result of noisy samples. ^ When calculating the reconstruction loss, the original, clean signal is always used. As a target benchmark, this forces the decoder to... Learn how to filter out noise and restore the true physical signal.

[0033] Latent space consistency constraint: This is the core innovation of this invention. When calculating the consistency loss, it is directly calculated... and The mean square error between them, i.e. 2-norm. This constraint acts like a "spring," forcing the encoder... Even when faced with strong noise interference, it can still extract low-dimensional features that are highly consistent with the clean signal. S203: Global Optimization and Weight Update. Reconstruction loss and consistency loss are aggregated into the total loss. Then it proceeds to the "Global Optimizer" at the bottom of Figure 3. After calculating the error gradient, the optimizer updates the weight parameters of the encoder and decoder uniformly through error backpropagation. and After continuous iteration of the closed loop described above, the model eventually converges. At this point, Figure 3 Unified encoder in When extracted separately, it possesses the powerful ability to stably extract pure micro-motion features of non-cooperative targets under extremely low signal-to-noise ratios, thus laying a solid foundation for the subsequent accurate parameter extraction of the Transformer.

[0034] S3: Global temporal modeling and noise-resistant coding based on Transformer; To overcome the limitations of the local receptive field in traditional CNNs and seamlessly inherit the low-dimensional features output by VAEs, this invention uses Transformer as the backbone sequence encoder and designs an adaptive noise-resistant attention and cross-domain bridging mechanism.

[0035] Adaptive noise reduction from VAE to Transformer: Figure 4 The presentation highlights the data flow of the time-frequency image sequence after VAE encoding, entering the Transformer's internal block for in-depth processing. Thanks to the global computational nature of the Transformer's multi-head attention mechanism, the network can adaptively assign different weights to different sequence features in the VAE input. For feature segments with significant micro-motion characteristics and high signal-to-noise ratios, the network automatically assigns extremely high attention weights; while for segments severely affected by spatial environmental interference, with low signal-to-noise ratios, or even containing only pure noise, the network automatically assigns extremely low weights. This mechanism closely integrates with the feature extraction results of the VAE, effectively overcoming the shortcomings of traditional models that are extremely sensitive to missing signal segments or strong noise occlusion.

[0036] Input sequence construction: The 1×256 latent vector output by the VAE is considered as the basic feature. To implement the hierarchical guidance mechanism, this embodiment designs a hybrid token sequence as the input to the Transformer. Input sequence Its composition is as follows: (1); In formula (1), This is the sequence vector after embedding the time-frequency graph; A learnable type classification token is used to aggregate global features for target recognition; Tokens are extracted from the parameters, specifically for regressing micro-motion parameters.

[0037] This step introduces Transformer training, focusing on the target recognition and parameter regression capabilities of the parameter extraction module. The optimizer uses either Adam or AdamW and incorporates a cosine annealing learning rate decay strategy. The VAE model weights are frozen, keeping the parameters of the latent feature extraction part unchanged. Only the Transformer module and subsequent classification and regression heads are trained to minimize the task loss. This strategy ensures the stability of low-level feature extraction while leveraging the powerful sequence modeling capabilities of the Transformer to achieve progressive inference from coarse-grained classification to fine-grained parameter extraction through a hierarchical guidance mechanism.

[0038] Temporal slicing and cross-domain bridging: To support the serialization attention computation within the Transformer mentioned above, Figure 5 The document details the underlying data format conversion and bridging process. After VAE processing, the features are no longer treated as a single static 2D image, but rather segmented along the time axis into continuous multi-frame feature slices (token sequences). Simultaneously, positional encoding is injected into each slice to strictly preserve the temporal physical evolution of the radar signal. This mechanism perfectly demonstrates how continuous image frames smoothly transition from the VAE's image domain to the sequence domain required by the Transformer, completely breaking the traditional CNN's dependence on fixed input sizes and giving the model extremely strong architectural compatibility.

[0039] S4: Design a cascaded Transformer structure to obtain high-precision target type recognition results and micro-motion parameter regression results.

[0040] To address the problems of fragmented multi-task branches and rigid feature concatenation in traditional CNNs, this invention creatively introduces the Prompt paradigm of large language models and designs a cascaded Transformer architecture. For example... Figure 6The diagram shows the category context-guided cascaded Transformer architecture under the large model Prompt paradigm. The main Transformer architecture, considering the balance between computational efficiency and model generalization ability, uses a relatively shallow number of network layers, specifically 4 to 6 layers. Each layer includes a multi-head self-attention mechanism and a feedforward neural network. The parameters such as the number of heads and hidden layer dimensions of the multi-head attention mechanism follow the standard Transformer architecture design. The number of heads for the multi-head attention mechanism is set to 8, and the dimension of the feedforward network is set to 1024. The self-attention mechanism is used to capture long-distance temporal dependencies in time-frequency features. Layer-guided output: Extracting the corresponding... The location vector is passed through a fully connected layer and a softmax activation function to output the target class probability distribution. The corresponding... The vector of position. Due to the self-attention mechanism, this vector has been fused with... Category prior information and The feature information is obtained. Through a fully connected regression layer, micro-motion parameters such as precession frequency and precession angle are output. This includes two stages, as detailed below: Phase 1: Prior category extraction; The initial token sequence, after being processed by the VAE, is first input into the first-level Transformer module for global feature interaction. Its output is connected to the classification head to identify and output the category label of the current spatial target, i.e., the easily obtainable privileged information.

[0041] Prompt context splicing and second-level guidance regression: This is the core guidance mechanism of this invention. For example... Figure 6 As shown, the classification result output from the first level is transformed into a contextual cue word vector [CTX] through an embedding mapping layer. This contextual cue word is pre-concatenated into the reused original token sequence, forming a completely new concatenated sequence. This concatenated sequence is then fed into the second-level Transformer module. Here, the network utilizes a "context-guided attention" mechanism, allowing the contextual cue words to deeply participate in the feature extraction and attention calculation of the original sequence tokens, thus performing a second deep attention calculation under the explicit cues of the category context. Finally, the parameter regression head at the end outputs highly complex micro-motion parameters (such as precession frequency and precession angle).

[0042] The second stage is Prompt context concatenation and second-level guided regression, which is the core guidance mechanism of this invention. The classification result output from the first stage is transformed into context cue word vectors ([CTX]) through an embedding mapping layer. These context cue words are pre-concatenated into the reused original token sequence to form a new concatenated sequence. This concatenated sequence is then fed into the second-level Transformer module. At this point, the network utilizes the "context-guided attention" mechanism, allowing the context cue words to deeply participate in the feature extraction and attention calculation of the original sequence tokens, thereby performing a second deep attention calculation under the explicit cues of the category context. Finally, the parameter regression head at the end outputs highly complex micro-motion parameters, such as precession frequency and precession angle.

[0043] After completing the pre-training phase one, this invention enters the downstream task joint training phase. To efficiently achieve implicit guidance of target type on micro-motion parameters, this invention designs a serialization processing architecture based on multi-task-specific tokens. The cascaded guidance training mechanism based on multi-task token sequences is as follows: Figure 7 As shown, it includes: S401: Noise-resistant feature extraction based on frozen VAE. Only noisy time-frequency maps are input during the training phase. The signal passes through the frozen unified encoder. Processing and outputting potential noise reduction features Since the encoder weights are fixed, this step ensures that the denoising capabilities gained during the pre-training phase are not catastrophically forgotten in subsequent training.

[0044] S402: Multi-task Joint Token Sequence Construction (Core Innovation). This invention abandons the cumbersome design of traditional models that require multiple forward propagations or explicit feature splicing. As shown in the middle of Figure 4, the denoising features are divided into sequences along the time axis. 1, 2... The sequence is then prepended with two learnable, proprietary tokens: a target type token [Type_Token] and a micro-parameter token [Param_Token]. These three parts together constitute a joint input sequence. S403: Implicit guidance and feature interaction based on self-attention mechanism. The constructed joint sequence is input into a trainable Transformer encoder. During the self-attention computation of the deep network, [Param_Token] will spontaneously interact with [Type_Token] and feature slices. This process involves intensive information exchange. At the underlying physical mechanism level, this is equivalent to using the macroscopic semantic features of the target type to guide and constrain the local feature extraction of micro-motion parameters, thus achieving efficient implicit feature fusion. S404: Dual-head parallel output and loss calculation. After processing by the Transformer, the updated [Type_Token] and [Param_Token] are extracted from the top layer of the network and fed into the "Classification Task Head" and "Parameter Regression Head" respectively. Combining the true type labels and true parameter labels of the input data, the network calculates the classification cross-entropy loss in parallel. and regression mean square error loss And summed up as total loss . S405: Gradient Truncation and Joint Optimization. As shown at the bottom and left of Figure 4, the joint optimizer (Adam) calculates the gradient based on the total loss and performs error backpropagation. This process simultaneously updates the Transformer encoder, the two task heads, and the embedding weights of the initial token. Specifically, when the error gradient propagates to the output of the unified encoder, the system performs a gradient truncation operation, thereby achieving perfect decoupling between front-end denoising and back-end recognition parameter extraction.

[0045] Explanation of the advantages of this architecture in the actual inference phase: In the actual inference process after model deployment, the input test signal, after preprocessing, also constructs a sequence containing [Type_Token] and [Param_Token]. The data only needs to go through a single forward propagation of the Transformer to obtain classification and regression results simultaneously at the top level. This multi-Token parallel architecture not only achieves deep parameter guidance during training but also avoids the computational delay caused by serial cascading during inference, greatly satisfying the real-time requirements of radar non-cooperative target recognition systems.

[0046] Level 1 processing and privileged information extraction: such as Figure 5 As shown, the initial token sequence after VAE processing is first input into the first-level Transformer Block for global feature interaction. Its output is first connected to the classification head to identify and output the category label of the current spatial target, i.e., the easily obtainable privileged information. Prompt context concatenation and second-level guided regression: This is the core guidance mechanism of this invention. The classification result output from the first level is transformed into a context prompt vector, and it is concatenated again to the beginning or end of the original token sequence. The concatenated new sequence is fed into the second-level Transformer Block. At this time, the network performs a second deep attention calculation under the explicit prompt of the category context. Finally, the highly complex micro-motion parameters are output by the parameter regression head at the end.

[0047] The loss function for this implementation is designed as follows: The total loss function in this embodiment It is composed of a weighted average of VAE reconstruction loss, KL divergence loss, consistency loss, classification loss, and regression loss. The formula is as follows: (2); In formula (2), This is the basic loss of VAE; The mean square error is used to measure the difference between the time-frequency plot reconstructed by the decoder and the original clean time-frequency plot as the reconstruction loss. KL divergence loss is used to calculate the KL divergence between the latent vector distribution and the standard normal distribution, ensuring the continuity and smoothness of the latent space. The standard cross-entropy loss function is used for classification loss to calculate the target class prediction error. For regression loss, Smooth L1 Loss or mean squared error loss is used to calculate the regression error of the micro-motion parameters.

[0048] To address the characteristic of simulation datasets containing multiple noise level labels, a consistency loss algorithm is designed to achieve denoising. The core idea is that the feature distribution of the same target in the latent space should remain consistent under different signal-to-noise ratio conditions.

[0049] The specific implementation logic is as follows: For any two samples in the training batch... and If they belong to the same target category and have the same physical parameters, but have different noise level labels, then their latent feature vectors output by the encoder are calculated. and The distance between them. This embodiment uses cosine similarity or mean square error as the distance metric, forcing the network to learn noise-invariant features. The specific calculation formula is defined as: (3); In formula (3), Ω represents a set of sample pairs with the same target attributes but different noise levels, and μ represents the mean vector of the encoder output.

[0050] Through the above implementation methods, this invention utilizes VAE to achieve efficient dimensionality reduction and denoising of high-dimensional time-frequency graphs, and solves the problem of the difficulty in directly extracting micro-motion parameters through a hierarchical guided Transformer architecture, thereby improving the accuracy and robustness of non-cooperative target recognition.

[0051] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention and within the spirit and principles of the present invention, shall still fall within the protection scope of the present invention.

Claims

1. A method for non-cooperative target recognition and micro-motion parameter extraction based on consistent VAE and Transformer, characterized in that, include: Step 1: Obtain the time-frequency diagram based on the radar echo signal; Step 2: Perform phased training on the consistent variational autoencoder and the Transformer encoder; Step 3: Input the time-frequency graph into the consistency variational autoencoder. The encoder maps the time-frequency graph to the latent space and uses the consistency loss constraint to make the latent feature vectors of the same target at different noise levels tend to be consistent, thus obtaining the denoised low-dimensional latent feature vectors. Step 4: Divide the denoised low-dimensional latent feature vector into a feature slice sequence along the time axis, and inject position encoding into each feature slice to obtain a token sequence with temporal information; Step 5: Concatenate the token sequence with the learnable target type-specific token and the micro-motion parameter-specific token to construct a joint input sequence; Step 6: Input the joint input sequence into the Transformer encoder, and use the self-attention mechanism to enable the micro-motion parameter-specific token to interact with the target type-specific token and feature slices to obtain the parameter feature representation of the fused category prior information; Step 7: Obtain the target type recognition result based on the output corresponding to the target type-specific token; Step 8: Obtain the regression results of the micro-motion parameters based on the parameter characteristics.

2. The method for non-cooperative target recognition and micro-motion parameter extraction based on consistent VAE and Transformer according to claim 1, characterized in that, Step 1 includes: The radar echo signal is transformed into a time-frequency domain representation by performing time-frequency transformation, thereby obtaining a time-frequency map.

3. The method for non-cooperative target recognition and micro-motion parameter extraction based on consistent VAE and Transformer according to claim 1, characterized in that, Step 3 specifically includes: The clean signal sample and the noisy signal sample are input into the unified encoder with shared weights to obtain the first latent feature vector and the second latent feature vector, respectively. The consistency loss is obtained based on the mean squared error between the first and second latent feature vectors. The first latent feature vector and the second latent feature vector are respectively input into a unified decoder with shared weights to obtain the first reconstructed signal and the second reconstructed signal; The reconstruction loss is obtained based on the differences between the first and second reconstructed signals and the clean signal samples. Based on the consistency loss and reconstruction loss, obtain the total loss and update the weight parameters of the unified encoder and unified decoder according to the total loss; The trained unified encoder is used as the encoder of the consistent variational autoencoder. The time-frequency image to be tested is input into the encoder to obtain the denoised low-dimensional latent feature vector.

4. The method for non-cooperative target recognition and micro-motion parameter extraction based on consistent VAE and Transformer according to claim 3, characterized in that, Obtaining the first latent feature vector and the second latent feature vector includes: Noisy signal samples and clean signal samples are mapped to the internal vector space respectively, and feature recombination and nonlinear transformation are performed in sequence. The features after nonlinear transformation are mapped to mean vector and variance vector respectively, and the first latent feature vector and the second latent feature vector are obtained through reparameterized sampling.

5. The method for non-cooperative target recognition and micro-motion parameter extraction based on consistent VAE and Transformer according to claim 1, characterized in that, Step 4 specifically includes: The low-dimensional latent feature vector is transformed from an image domain representation into a series of feature slices arranged along the time axis. Location encoding information is added to each feature slice to preserve the temporal physical evolution of radar signals and obtain a token sequence with temporal information.

6. The method for non-cooperative target recognition and micro-motion parameter extraction based on consistent VAE and Transformer according to claim 1, characterized in that, Step 5 specifically includes: The target type-specific token and the micro-motion parameter-specific token are placed at the beginning of the token sequence, so that the target type-specific token and the micro-motion parameter-specific token are preferentially associated with the feature slice in the subsequent attention calculation to obtain the joint input sequence.

7. The method for non-cooperative target recognition and micro-motion parameter extraction based on consistent VAE and Transformer according to claim 1, characterized in that, Step 6 specifically includes: The joint input sequence is input into a multi-layer Transformer encoder, and the attention weight between the micro-motion parameter-specific token and the target type-specific token is calculated through a multi-head self-attention mechanism, so that the category semantic information carried by the target type-specific token is transferred to the micro-motion parameter-specific token. Based on the attention weights, obtain the parametric feature representation of the prior information of the fusion category.

8. The method for non-cooperative target recognition and micro-motion parameter extraction based on consistent VAE and Transformer according to claim 1, characterized in that, Step 7 specifically includes: Extract the vector corresponding to the position of the target type-specific token in the output sequence of the Transformer encoder; The vector corresponding to the target type's unique token is input into the classification task head, and the target type recognition result is obtained through a fully connected layer and a Softmax activation function.

9. The method for non-cooperative target recognition and micro-motion parameter extraction based on consistent VAE and Transformer according to claim 1, characterized in that, Step 8 specifically includes: Extract the vector corresponding to the position of the micro-motion parameter-specific token in the output sequence of the Transformer encoder as the parameter feature representation; The parameter feature representation is input into the parameter regression head, and the regression results of the micro-motion parameters of precession frequency and precession angle are obtained through a fully connected regression layer.

10. The method for non-cooperative target recognition and micro-motion parameter extraction based on consistent VAE and Transformer according to claim 1, characterized in that, Step 2, the phased training, specifically includes: In the first stage, the parameters of the Transformer encoder are frozen, an initial learning rate is set, and the weights of the encoder and decoder in the consistent variational autoencoder are updated using the optimizer until convergence, based on the reconstruction loss. KL divergence loss and consistency loss Total loss of VAE construction ; In the second stage, the weights of the consistency variational autoencoder are frozen, and the noisy time-frequency map is input into the encoder to obtain denoising potential features, based on the classification loss. and regression loss Total loss of the construction task The Adam optimizer is used to update the embedding weights of the Transformer encoder, classification task head, parameter regression head, target type-specific token and micro-motion parameter-specific token, and gradient truncation is performed when backpropagating to the encoder output. VAE total loss The expression is: （1）； Total mission losses The expression is: （2）。

Images