Generation enhancement and detection framework (GASD-Net) based on adversarial signal decoupling

Abstract

The invention discloses a generation enhancement and detection framework (GASD-Net) based on adversarial signal decoupling, and belongs to the technical field of radar signal processing. According to the method, a variational adversarial network architecture of a double-hidden-variable space is provided for solving the problems of scarcity of small target samples and serious feature coupling under the sea clutter background. The method comprises the following steps: firstly, constructing a three-state data set comprising pure clutters, a synthetic target and a real target; then a double-flow encoder is trained to decouple radar echoes into independent clutter latent variables and target latent variables; in the training process, sparse consistency loss (Sparse Push) is introduced to suppress projection of clutters in a target space, and activation hinge loss (Hinge Pull) is utilized to forcibly extract deep features of a real target. According to the method, high-fidelity enhancement of radar signals is achieved, the blind source separation capability is achieved, weak target signals can be effectively purified from strong sea clutters, and the detection precision and robustness under the complex sea condition are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of radar signal processing, deep learning, marine remote sensing, and intelligent countermeasures technology. Specifically, this invention relates to a radar echo processing method based on a hybrid architecture of Variational Auto-Encoder (VAE) and Generative Adversarial Network (GAN), and particularly to a technique that utilizes a dual-latent space to decouple the physical features of the signal, thereby achieving data enhancement, blind source separation, and high-precision detection of weak target signals against a sea clutter background. Background Technology

[0002] 1. Current Status and Challenges of Radar Sea Clutter Detection

[0003] Radar systems play an irreplaceable role in modern missions such as safeguarding maritime rights, military coastal defense early warning, maritime search and rescue, and monitoring illegal fishing. However, with the continuous improvement of radar resolution (such as high-resolution X-band radar), the echoes generated by sea surface reflections (i.e., "sea clutter") exhibit extremely complex characteristics: Non-Gaussianity: At low grazing angles and high resolutions, the amplitude distribution of sea clutter deviates significantly from the Rayleigh distribution, exhibiting a significant long tail characteristic (such as K distribution, Weibull distribution, or Pareto distribution), making it difficult to control the false alarm rate.

[0004] Non-stationarity: Sea state changes dynamically with wind speed, wind direction, wave swell and time, and the statistical parameters of clutter may fluctuate drastically in a short period of time.

[0005] Strong coherence and Spikes: Breaking waves produce strong echo spikes that last for several seconds, with amplitude, Doppler spectral width and duration that are very similar to those of low-speed small targets (such as floating mines, periscopes, and small wooden boats).

[0006] Traditional detection methods are primarily based on statistical signal processing theory, such as constant false alarm rate (CFAR) detectors. Cell-averaged CFAR (CA-CFAR) is optimal against a Gaussian background, but its performance deteriorates sharply in multi-target or clutter-edge environments. While ordered statistical CFAR (OS-CFAR) improves robustness, it often necessitates raising the detection threshold to maintain a constant false alarm rate (PFA) when facing long-tailed clutter, resulting in extremely low detection probability (Pd) for weak targets with a signal-to-clutter ratio (SCR) below 0 dB. Furthermore, traditional methods typically process data based on a single dimension (such as amplitude or Doppler), making it difficult to fully utilize the multi-dimensional coupling characteristics of time, frequency, and space.

[0007] 2. Applications and limitations of deep learning in the radar field

[0008] In recent years, deep learning has been introduced into the field of radar signal processing. Convolutional Neural Network (CNN)-based classifiers have shown excellent performance in feature extraction, but they still face two major challenges:

[0009] The Small Sample Size Problem (SSM) presents a severe shortage of real-world small target data: acquiring radar data containing real small targets is costly and rare, while pure sea clutter data is readily available. This significant imbalance between positive and negative samples makes supervised learning models highly susceptible to overfitting and exhibits poor generalization ability.

[0010] Generative models lack physical interpretability: While existing generative adversarial networks (GANs) can generate realistic radar echo samples to augment datasets, most standard GANs only learn the surface probability distribution of the samples and cannot understand the internal physical composition of the signal (i.e., the superposition relationship between the "target" and the "background"). The generated samples are often "mixed states," and it is impossible to independently control the generation of "clean targets" or "clean clutter," making it difficult for the model to directly assist in removing clutter or improving the signal-to-clutter ratio.

[0011] Therefore, there is an urgent need for a new network architecture that can understand the physical synthesis process of radar signals and has the ability to deeply decouple the "target component" from the "clutter component" in order to solve the problems of data scarcity and weak signal extraction. Summary of the Invention

[0012] This invention aims to address the problems of scarce training samples, difficulty in extracting weak target features, and lack of physical interpretability in the existing technologies, and proposes a generative enhancement and detection framework (GASD-Net) based on adversarial signal decoupling.

[0013] The core design concept of this invention is a shift from "simple sample generation" to "deep feature decoupling." By constructing a dual-stream latent variable space, the network is forced to learn and decompose radar echo signals into orthogonal or independent "background texture features" and "target physical features." This invention is not only a high-fidelity data augmenter but also a signal purifier with blind source separation capabilities.

[0014] The technical solution adopted in this invention is as follows:

[0015] Firstly, a generation enhancement and detection method based on adversarial signal decoupling is provided, comprising the following steps:

[0016] Step S1: Constructing a three-state radar signal dataset under physical constraints. Unlike traditional binary classification (target present / no target) datasets, this invention constructs a training set containing three physical states:

[0017] 1. Pure Clutter State (PCS): Derived from range cells in radar data where no target is confirmed, representing the background distribution of the actual sea state.

[0018] 2. Synthetic Target State (STS): Based on the pure clutter state, a simulated target signal is injected according to the radar equations and target motion model. This state has a clear "ground truth" and is used to provide a strong supervision signal.

[0019] 3. Real Target State (RTS): This is derived from cells in the measured data that contain cooperative targets (such as experimental vessels). This state contains the real target-wave coupling effect but lacks the pure target ground truth value, and is used for semi-supervised or self-supervised learning.

[0020] Step S2: Construct a Dual Hidden Space Decoupling Network (GASD-Net) to build an end-to-end deep neural network containing three core components:

[0021] 1. Dual-Stream Encoder: Maps the input time-series signal I / Q (In-phase / Quadrature) to two independent latent distribution spaces:

[0022] Clutter latent variable space (Zc): Specifically encodes the texture, speckle noise, and long trailing shape of sea clutter.

[0023] Target latent variable space (Zt): specifically encodes the target's amplitude fluctuations, Doppler frequency shifts, and micro-motion characteristics.

[0024] 2. Disentangled Decoder: Contains a generative network with shared weights, but has two logical paths:

[0025] Path-Reconstruction: The input is [Zc, Zt], and the goal is to reconstruct the original mixed signal.

[0026] Path-Purify: The input is [0, Zt] (forced masking of clutter features), and the goal is to output a clean target signal after denoising.

[0027] 3. Adversarial Discriminator: Used to determine the authenticity of the generated signal and constrain the generated distribution to approximate the real radar echo distribution.

[0028] Step S3: Design a multi-task joint decoupling loss function. To drive the automatic separation of features in the dual latent space, a joint objective function containing the following four types of losses was designed:

[0029] 1. Basic reconstruction loss (L_recon): Ensures that no information is lost.

[0030] 2. Strongly supervised separation loss (L_sup): Using synthetic target state samples, the purification path output is forced to be consistent with the injected analog signal.

[0031] 3. Sparse Consistency Loss: Using pure clutter state samples, any non-zero activation of the target latent variable Zt is penalized, forcing the network to not "phantom" the target in clutter.

[0032] 4. Hinge Pull: Using real target state samples, it penalizes cases where the Zt activation is too low, forcing the network to mine weak target features from real data.

[0033] Step S4: Adversarial iterative training employs an alternating optimization strategy, where the generator (encoder + decoder) and discriminator engage in game-like training until the network converges and possesses stable signal separation capabilities.

[0034] Step S5: Perform target detection and signal purification. During the inference phase, the signal to be tested is input into the network, and detection can be performed in two modes:

[0035] 1. Latent Space Detection Mode: Directly monitors the energy modulus of the Zt space.

[0036] 2. Reconstruction Domain Detection Mode: The target waveform is reconstructed using the purification path output, and its energy or signal-to-noise ratio is calculated for decision-making.

[0037] Secondly, a computer system and a radar signal processing terminal for implementing the above method are provided.

[0038] The beneficial effects of this invention are:

[0039] It achieves physical-level signal decoupling: GASD-Net can separate sea clutter from the target signal, thereby "purifying" the target at an extremely low signal-to-noise ratio, which is something that traditional black-box deep learning models cannot do.

[0040] It solves the problem of training with small samples: by providing strong supervision through "synthetic targets" and combining them with "real targets" for domain adaptation, the model can learn the ideal features of the target and adapt to the nonlinear coupling of the real environment, which greatly reduces the dependence on large-scale measured labeled data.

[0041] The robustness of detection is improved: the introduced Sparse Push and Hinge Pull mechanisms essentially construct a "rejection threshold" for background clutter and a "sensitivity threshold" for the target, effectively reducing the false alarm rate (PFA) while improving the detection probability (Pd) of weak targets. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the overall network architecture of the GASD-Net of the present invention, showing the connection relationship between the dual-stream encoder, dual-path decoder and discriminator.

[0043] Figure 2 This is a schematic diagram illustrating the principle of constructing a three-state dataset and decoupling features from the dual latent space in this invention, which intuitively demonstrates the difference in response between Zc and Zt when processing different types of signals.

[0044] Figure 3 This is a flowchart illustrating the calculation process of the multi-task joint decoupling loss function of the present invention, which details the different loss calculation paths for pure clutter, synthetic targets, and real targets.

[0045] Figure 4 This is a flowchart illustrating the training process and detection logic of the method of the present invention.

[0046] Figure 5 The graph (ROC curve) shows the performance comparison of the method of this invention on the measured IPIX radar dataset, illustrating the performance differences with CA-CFAR, OS-CFAR and conventional CNN detectors. Detailed Implementation

[0047] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The following embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0048] Example 1: Radar Environment Configuration and Construction of Three-State Dataset

[0049] This embodiment first describes the data acquisition and preprocessing process, which is the foundation for training a robust model.

[0050] 1.1 Data Sources and Parameters: The IPIX radar measured sea clutter database (or similar X-band marine surveillance radar data) provided by McMaster University, Canada, was selected. Typical data parameters are as follows:

[0051] Radar system: shore-based coherent radar

[0052] Operating frequency: 9.39GHz (X-band)

[0053] Polarization: HH (horizontal transmit, horizontal receive) or VV (vertical transmit, vertical receive)

[0054] Pulse Repetition Frequency (PRF): 1000Hz~2000Hz

[0055] Distance resolution: 30 meters (pulse width 200 ns)

[0056] Sampling data format: I (in-phase) / Q (orthogonal) complex sequence

[0057] 1.2 Data Preprocessing: Perform the following standardization operations on the raw I / Q data:

[0058] 1. Modulus Normalization: Calculate the mean μ_global and standard deviation σ_global of the time series for all distance units. Perform Z-Score normalization for each sampling point x(t): X′(t)=(x(t)-μ_global) / σ_global

[0059] 2. Sequence Slicing: A sliding window is used to slice the long weighted sequence into sample segments of fixed length. In this embodiment, the sequence length N = 512 points (corresponding to an observation time of approximately 0.5 seconds).

[0060] 1.3 Construction Logic of Tri-State Dataset The core of this invention lies in using samples with three different properties for joint training:

[0061] State A: Pure Clutter (PC)

[0062] Selection: Select all distance cell data except for the distance cell where the target is located and its adjacent protection cells.

[0063] Label: Label = 0, auxiliary label Mask_target = all zero sequence.

[0064] Function: To learn the background distribution P(Clutter) of sea clutter and suppress the response in the target latent variable space through "Sparse Push" loss.

[0065] State B: Synthetic Target (ST)

[0066] Selection: Randomly select pure cluttered wave state samples as background.

[0067] Injection: Simulated target echo s(t) is superimposed on the background.

[0068] Simulation target model: in:

[0069] A(t): Amplitude envelope. To simulate weak targets, the signal-to-noise ratio (SCR) is set to be randomly and uniformly distributed between -5dB and +5dB.

[0070] f_d: Doppler frequency. The normalized frequency is randomly distributed in the range [0.01, 0.3].

[0071] k: Chirp Rate. Simulates the target's accelerated motion; k is random between [-1e-4, 1e-4].

[0072] The initial phase is uniformly distributed in the range [0, 2π].

[0073] Label: Label = 1, auxiliary label Mask_target = s(t) (pure analog signal).

[0074] Function: Provides Ground Truth, enabling strongly supervised networks to learn "denoising" and "reconstruction" capabilities.

[0075] State C: Real Target (RT)

[0076] Selection: Select range cells that contain targets as confirmed by radar logs (such as range gates 7-9 in IPIX data).

[0077] Label: Label = 2, auxiliary label Mask_target = Unknown.

[0078] Function: To force the network to adapt to the nonlinear coupling between real targets and sea clutter (such as occlusion effect and multipath effect), and to prevent "domain shift" caused by training only on synthetic data.

[0079] Example 2: Detailed Explanation of GASD-Net Network Architecture

[0080] like Figure 1 As shown, GASD-Net adopts the VAE-GAN architecture based on one-dimensional convolution.

[0081] 2.1 Dual-Stream Encoder: Encoder E maps the input sequence x (dimension 2×512, I / Q channels) to latent features.

[0082] Shared feature extraction layer:

[0083] Layer 1: Conv1d(in=2, out=32, kernel=5, stride=2), BatchNorm, LeakyReLU(0.2)

[0084] Layer 2: Conv1d (in=32, out=64, kernel=5, stride=2), BatchNorm, LeakyReLU (0.2)

[0085] Layer 3: Conv1d (in=64, out=128, kernel=5, stride=2), BatchNorm, LeakyReLU (0.2)

[0086] Flatten: Flattens the feature map into a one-dimensional vector.

[0087] Clutter Branch:

[0088] FC_mu_c: Fully connected layer, outputs clutter mean μ_c (dimension Dim_z = 32).

[0089] FC_var_c: Fully connected layer, output clutter log-variance log(σ) 2 _c).

[0090] Target Branch:

[0091] FC_mu_t: Fully connected layer, outputting the target mean μ_t (dimension Dim_z = 32).

[0092] FC_var_t: Fully connected layer, outputs the target log-variance log(σ) 2 _t).

[0093] Reparameterization:

[0094] z_c=μ_c+exp(0.5·log(σ 2 _c))·ε, where ε~N(0,I)

[0095] z_t=μ_t+exp(0.5·log(σ 2 _t))·ε, where ε~N(0, I)

[0096] 2.2 Disentangled Decoder: Decoder G aims to recover time-domain signals from latent variables.

[0097] Input strategy: The decoder itself has a simple structure, but it accepts two different combinations of inputs (achieved by controlling the gating).

[0098] Network structure:

[0099] FC_input maps the input z (dimension 64 or 32+32) back to the feature map size.

[0100] Reshape: Transforms the shape into (128, 64).

[0101] Layer1: ConvTranspose1d (in=128, out=64, kernel=5, stride=2), BatchNorm, ReLU

[0102] Layer 2: ConvTranspose1d (in=64, out=32, kernel=5, stride=2), BatchNorm, ReLU

[0103] Layer 3: ConvTranspose1d(in=32,out=2,kernel=5,stride=2),Tanh(output range normalized to [-1,1])

[0104] Output path:

[0105] 1. x_recon = G([z_c, z_t]): Reconstruct the mixed signal.

[0106] 2.x_target = G([0,z_t]): Set z_c to zero and use only z_t to generate the signal to achieve purification.

[0107] 2.3 Discriminator: Discriminator D is used to distinguish between real radar echoes and generated echoes.

[0108] Structure: A 4-layer 1D CNN is used, with the end connected to a scalar probability output by a sigmoid activation function.

[0109] Input: real sample x or reconstructed sample x_recon.

[0110] Example 3: Design of Joint Loss Function for Multiple Tasks

[0111] This embodiment details the mathematical principles of driving feature decoupling. The total loss function is defined as:

[0112] L_total=λ1L_recon+λ2L_sup+λ3L_sparse+λ4L_hinge+λ5L-KL+λ6L_adv

[0113] 3.1 The basic reconstruction loss (L_recon) measures the decoder's ability to reconstruct the original input and is applicable to all samples. L_recon = / / x - x_recon / / 2 2

[0114] 3.2 The strongly supervised separation loss (L_sup) is calculated only for the synthesized target state (Label = 1). Since we know the injected pure signal s(t) during synthesis, we require the output x_target of the purification path to approximate s(t). L_sup = / / s(t) - x_target / / 2 2

[0115] 3.3 Sparse Push Loss is calculated only for pure clutter states (Label = 0). Physical truth: Pure clutter does not contain the target. Therefore, the network's response in the target latent variable space Z_t should be completely suppressed. L_sparse = / / z_t / / 1 Note: The L1 norm is used to induce sparsity.

[0116] 3.4 The activation hinge loss is calculated only for the real target state (Label = 2). Physical truth: The real target sample must contain target features. To prevent the network from mistaking a real, weak target for clutter (i.e., z_t collapses to 0), a hinge loss is introduced to force the energy of z_t to exceed a specific threshold η. L_hinge = max(0, η - / / z_t / / 1). Parameter setting: The threshold η is empirically set to 1.0–2.0.

[0117] 3.5KL divergence loss (L_KL) is the standard loss of the VAE, constraining the latent variable distribution to approximate a standard normal distribution N(0, I), ensuring the continuity of the latent space. L_KL = -0.5·∑(1+log(σ) 2 )-μ 2 -σ 2 )

[0118] 3.6 Adversarial Loss (L_adv) utilizes feedback provided by the discriminator D to improve the realism of the generated data.

[0119] L_adv = -E[log(D(x_recon))]

[0120] Recommended weight configuration: λ1 = 10 (to ensure waveform accuracy), λ2 = 50 (strong supervision core), λ3 = 20 (strong suppression of false alarms), λ4 = 10 (to recall weak targets), λ5 = 0.1, λ6 = 1.0.

[0121] Example 4: Training and Testing Process

[0122] 4.1 Training Strategy

[0123] 1. Initialization: Randomly initialize network parameters.

[0124] 2. Iterative loop:

[0125] A batch (e.g., 64 samples, including 32 pure clutter, 16 synthetic targets, and 16 real targets) is sampled from the three-state dataset.

[0126] Update the discriminator: Train D using only L_adv so that it can distinguish between the real x and the generated X_recon.

[0127] Update the generator (E+G): With D fixed, calculate L_total and update the parameters of E and G through backpropagation. Note that, depending on the sample label, L_sup, L_sparse, and L_hinge are calculated and accumulated only on the corresponding sample.

[0128] 3. Termination: Train for approximately 100–200 epochs until the loss curve converges.

[0129] 4.2 Detection and Purification Process After the network training is completed, it is deployed on the radar signal processing terminal (such as NVIDIA Jetson or FPGA accelerator card).

[0130] 1. Signal input: Real-time acquisition of radar echo sequence x_input.

[0131] 2. Feature extraction: z_t is obtained through encoder E.

[0132] 3. Mode 1 (Fast Detection): Calculate the energy feature E_feat = / / μ_t / / 2. If E_feat > γ (detection threshold), it is identified as a target; otherwise, it is identified as clutter. Advantages: Minimal computational cost, requiring only half the computation of a single forward propagation.

[0133] 4. Mode 2 (High-precision purification): Obtain x_out = G([0, z_t]) through the decoder purification path. Calculate the signal-to-noise ratio (SCR) of the output signal: SCR_out = 10log 10 ( / / x_out / / 2 / / / x_input-x_out / / 2 If SCR_out exceeds the preset value, it is identified as a target. Advantages: The target waveform after clutter removal can be visually observed, facilitating subsequent feature identification.

[0134] Example 5: Experimental Verification and Effect Analysis

[0135] Validation was performed using IPIX radar datasets from 1993 (e.g., datasets #17, #54, #280).

[0136] 5.1 Experimental Setup

[0137] Comparison methods: CA-CFAR, OS-CFAR, traditional CNN classifier, standard VAE.

[0138] Evaluation metrics: Receiver operating characteristic curve (ROC), area under the curve (AUC).

[0139] 5.2 Performance

[0140] Clutter suppression capability: In pure sea clutter scenarios, GASD-Net's L_sparse loss makes z_t extremely close to zero, resulting in a very low false alarm rate (PFA), which is superior to OS-CFAR's performance at clutter edges.

[0141] Weak target detection: On a test set with an average SCR of 3dB, when PFA = 10 -3 hour:

[0142] CA-CFAR detection probability (Pd) ≈ 0.45

[0143] The detection probability (Pd) of a standard CNN is approximately 0.72.

[0144] The detection probability (Pd) of GASD-Net is approximately 0.89.

[0145] Generalization ability: Due to the introduction of Real Target State (RTS) for Hinge Pull training, the performance degradation of the model at unseen sea state levels (such as generalizing from sea state 3 to sea state 4) is much smaller than that of the model trained only on synthetic data.

Claims

1. A generation enhancement and detection method based on adversarial signal decoupling, characterized in that, Includes the following steps: S1. Construct a three-state radar signal dataset: Based on the radar measured echo data, construct a three-state training sample set that includes pure clutter state, synthetic target state and real target state; S2. Construct a dual-latent space decoupling network (GASD-Net): Build a deep neural network containing a two-stream encoder, a decoupling decoder, and an adversarial discriminator; where the encoder maps the input signal to independent clutter latent variable spaces (Z_c) and target latent variable spaces (Z_t); S3. Define a multi-task joint decoupling loss function: Design a joint objective function that includes reconstruction loss, strong supervision separation loss, sparse consistency loss and activation hinge loss to drive feature decoupling; S4. Adversarial Iterative Training: An alternating optimization strategy is used to train the generator network and the discriminator network, enabling the model to separate the pure target component from the mixed signal. S5. Target Detection and Signal Purification: The radar signal to be tested is input into the trained encoder to extract latent variables. Target detection is achieved by monitoring the activation energy of the target latent variable space (Z_t) or by reconstructing the pure target signal through the decoder.

2. The method of claim 1, wherein, The construction of the three-state radar signal dataset in step S1 specifically includes: Pure clutter state: Echo sequence of non-target range cells selected from measured data; True target state: The range cell echo sequence containing the true target is confirmed from the measured data; Synthetic target state: A simulated target signal is superimposed on a pure clutter state; the simulated target signal is generated by randomizing the amplitude, Doppler frequency, initial phase and frequency modulation slope (Chirp) to cover different signal-to-noise ratios and motion states, and to simulate the characteristics of weak targets.

3. The method of claim 1, wherein, The specific network structure of GASD-Net in step S2 is as follows: The dual-stream encoder takes a temporal I / Q sequence as input, extracts features through multiple one-dimensional convolutional layers (Convld) and nonlinear activation layers, and then branches into two fully connected layers, which output clutter latent variable distribution parameters (μ_c, σ_c) and target latent variable distribution parameters (μ_t, σ_t), respectively. The latent variable vectors z_c and z_t are obtained by sampling through reparameterization techniques. Decoupled decoder: contains two generation paths; Reconstruction path: The input is the concatenated vector [z_c, z_t], which aims to reconstruct the original mixed input signal; Purification path: The input is the concatenated vector [0, z_t], which sets the clutter latent variable to zero, aiming to generate a denoised and pure target signal; Adversarial discriminator: Used to distinguish whether the input signal is a real radar echo or a reconstructed signal generated by the decoder.

4. The method of claim 1, wherein, The multi-task joint decoupling loss function L_total in step S3 is defined as follows: L_total=λ1L_recon+λ2L_sup+λ3L_sparse+λ4L_hinge+λ5L_KL+λ6L_adv Where: L_recon is the overall reconstruction loss, which calculates the mean square error between the input signal and the output of the reconstruction path; L_sup is the strongly supervised separation loss, which is calculated only for the synthetic target state sample, constraining the output of the purification path to approximate the injected pure analog signal; L_sparse is the sparse consistency loss, which is calculated only for pure clutter state samples; and L_hinge is the activation hinge loss, which is calculated only for the real target state sample.

5. The method of claim 4, wherein, The calculation logic for the sparse consistency loss L_sparse and the activation hinge loss L_hinge is as follows: Sparse consistency loss (Sparse Push): When the input sample label is pure clutter, the L1 norm of the target latent variable z_t is forced to approach zero, expressed by the formula: L_sparse= / / z_t / / 1 Activation Hinge Loss: When the input sample label is a real target, the L1 norm of the latent variable z_t of the target is forced to exceed the preset threshold η. The formula is expressed as: L_hinge=max(O,η- / / z_t / / 1) to achieve forced extraction of weak real target features and feature suppression of non-target signals.

6. The method of claim 1, wherein, Step S5 specifically includes: calculating the energy characteristic E = μt in the target latent variable space of the input signal; if E is greater than the detection threshold γ, the target is determined to exist, otherwise it is determined to be background clutter; or the reconstructed signal is directly output from the purification path of the decoder, and its signal-to-clutter ratio (SCR) is calculated for decision.

Images