Radar target detection method based on state space model autoencoder

Abstract

The application discloses a radar target detection method based on a state space model self-encoder and belongs to the technical field of radar signal processing, which is used for sea monitoring radar target detection and comprises the following steps: acquiring multi-distance unit time domain echo data of a sea monitoring radar, extracting pure sea clutter echo samples without target signals to form a training data set, constructing a selective state space model self-encoder network based on the training data set and training the network, inputting to-be-detected radar echo data into the trained selective state space model self-encoder network, outputting a reconstructed signal, calculating a detection statistic and setting an adaptive detection threshold, and screening a target area. Through the construction of the selective state space self-encoder and the training of the time-frequency domain composite loss, high-fidelity modeling of the sea clutter and high-sensitivity detection of the target are realized, and the detection probability under a low signal-to-clutter ratio is significantly improved and the false alarm rate is reduced.

Description

Technical Field

[0001] This invention discloses a radar target detection method based on a state-space model autoencoder, belonging to the field of radar signal processing technology. Background Technology

[0002] The backscattered echo signal from the sea surface received by maritime surveillance radar, known as sea clutter, is typically much stronger than the echo signal from small targets on the sea surface, severely impacting the radar's target detection performance. Under low grazing angle observation conditions, sea clutter exhibits complex statistical characteristics that are non-Gaussian and non-stationary, accompanied by anomalous scattering phenomena such as sea spikes, posing significant challenges to the detection of weak targets on the sea surface. Target detection against a strong clutter background is a core issue in maritime radar signal processing.

[0003] Currently, sea surface target detection mainly employs the following methods: CFAR detection methods based on statistical models, which design constant false alarm rate (CFAR) detectors after establishing a distribution model of clutter; subspace decomposition methods, which use techniques such as singular value decomposition to decompose echoes into different subspaces for target extraction; and time-frequency analysis methods, which transform signals to the time-frequency domain and utilize the characteristic differences between clutter and targets for detection. However, statistical model methods rely on the accuracy of distribution assumptions, and their performance degrades sharply when there is a distribution mismatch; subspace methods require manual setting of decomposition parameters, and their separation effect is limited when clutter and target spectra overlap; time-frequency analysis methods are limited by the inherent trade-offs in time-frequency resolution, and the selection of analysis parameters depends on experience.

[0004] In recent years, deep learning-based target detection methods have received considerable attention. Predictive methods, in particular, utilize temporal neural networks to predict pure clutter components and subtract them from the measured signal to achieve target detection. However, these methods implicitly assume that the network's predicted output does not contain target components, an assumption that is difficult to rigorously guarantee in real-world scenarios. Furthermore, Long Short-Term Memory (LSTM) networks suffer from the vanishing gradient problem when processing long sequences, while the computational complexity of attention-based networks is proportional to the square of the sequence length, which is detrimental to real-time processing of radar signals. Summary of the Invention

[0005] The purpose of this invention is to provide a radar target detection method based on a state-space model autoencoder, so as to solve the problem of target detection difficulties caused by the non-stationarity and non-Gaussianity of sea clutter under low signal-to-clutter ratio in the prior art.

[0006] Radar target detection methods based on state-space model autoencoders include: S1. Acquire time-domain echo data of multi-range cells of the sea radar, normalize the echo data, and reorganize it into the input channel format of the neural network. Use a sliding window method to extract time series samples of a fixed length and extract pure sea clutter echo samples without target signals to form a training dataset. S2. Based on the training dataset, construct a selective state-space model autoencoder network. With the goal of minimizing the reconstruction loss function, train the neural network and iteratively optimize the network parameters until convergence to obtain the trained selective state-space model autoencoder network. The selective state-space model autoencoder network consists of an input mapping layer, an encoder, a bottleneck layer, a decoder, and an output mapping layer. Both the encoder and decoder are composed of several layers of selective state-space model modules stacked together. The selective state-space model module includes discretized state transition equations. S3. Input the radar echo data to be detected into the trained selective state-space model autoencoder network, output the reconstructed signal, calculate the time-step reconstruction error between the original signal and the reconstructed signal as the detection statistic, set an adaptive detection threshold based on the detection statistic, and determine the time period when the detection statistic exceeds the threshold as the target area.

[0007] S1 includes the in-phase and quadrature components of the echo data output by the radar receiver; The input channel formats for recombining into neural networks include using the in-phase and quadrature components as two separate input channels, or combining the in-phase and quadrature components into a complex signal as a single input channel.

[0008] Constructing the input feature sequence : ; In the formula, For input feature index, For the first Each input feature.

[0009] S2 includes a selective state-space model module that includes discretized state transition equations: ; ; In the formula, For the first The hidden state vector at each time step This is the current output; This is the discretized state transition matrix; For depend on The generated input projection matrix, For depend on The generated output projection matrix.

[0010] The selective state-space model module adopts the Mamba architecture, and layer normalization and residual links are set between the various Mamba modules; The encoder and decoder consist of several layers of stacked Mamba modules; The input mapping layer, bottleneck layer, and output mapping layer are linear transformation layers.

[0011] The Mamba module receives input features and divides them into two branches. The first branch performs one-dimensional convolution and gated activation sequentially, while the second branch first performs selective projection to generate... and Then, the selective projection results, , and The discretized state transition equation is used to process the data, resulting in the second branch processing result. The second branch processing result is then multiplied element-wise with the first branch result to obtain the output feature. A skip connection is set between the encoder and the decoder to concatenate or add the output features of each layer of the encoder with the input features of the corresponding layer of the decoder element by element. The output mapping layer maps the decoder output to 1 dimension through linear transformation to obtain the reconstructed signal.

[0012] The reconstruction loss function is a composite loss function, including temporal reconstruction loss. and frequency domain consistency loss : ; ; In the formula, The original signal, To reconstruct the signal, The length of the time series; Through the and The error between the spectra is obtained by performing Fourier transforms on each spectrum.

[0013] S3 includes inputting the radar echo data to be detected into a trained selective state-space model autoencoder network, and outputting... ,calculate and The time-step reconstruction error between them is used as the detection statistic. : ; Set up a sliding window, for Applying a moving average filter yields the averaged detection statistic. ;calculate mean and standard deviation And set an adaptive detection threshold. : ; In the formula, This is the threshold coefficient; Will The time period is determined to include the target area.

[0014] For echo segments identified as pure clutter, continuously collect and add them to the training dataset, and perform incremental training at a learning rate lower than the initial learning rate.

[0015] Compared with existing technologies, this invention has the following advantages: This invention only requires pure sea clutter echo data to complete network training, without the need for labeled samples or paired data containing targets, significantly reducing the difficulty of acquiring training data; this invention transforms the target detection problem into an out-of-distribution sample detection problem, without relying on the assumption in predictive methods that the predicted output does not contain target components, making the method more applicable; there is a deep technical fit between the selective state-space model and the characteristics of sea clutter signals. Its long-range memory capability is suitable for modeling the long-range temporal correlation of sea clutter, and the state transition matrix structure is naturally suitable for encoding the quasi-periodic Doppler characteristics of clutter. The out-of-distribution generalization capability of the autoencoder makes the target naturally stand out in the reconstruction error; the computational complexity of the selective state-space model is linear to the sequence length, the inference process adopts a recursive approach, resulting in low memory consumption and the potential for real-time deployment on embedded processing platforms; the reconstruction error directly constitutes the detection statistics, eliminating the need for independent detector design. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the network structure of the selective state-space model autoencoder of the present invention; Figure 3 This is a schematic diagram of the internal structure of the selective state-space model module of the present invention; Figure 4 This is a schematic diagram illustrating the principle of reconstruction error detection in this invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0018] Radar target detection methods based on state-space model autoencoders include: S1. Acquire time-domain echo data of multi-range cells of the sea radar, normalize the echo data, and reorganize it into the input channel format of the neural network. Use a sliding window method to extract time series samples of a fixed length and extract pure sea clutter echo samples without target signals to form a training dataset. S2. Based on the training dataset, construct a selective state-space model autoencoder network. With the goal of minimizing the reconstruction loss function, train the neural network and iteratively optimize the network parameters until convergence to obtain the trained selective state-space model autoencoder network. The selective state-space model autoencoder network consists of an input mapping layer, an encoder, a bottleneck layer, a decoder, and an output mapping layer. Both the encoder and decoder are composed of several layers of selective state-space model modules stacked together. The selective state-space model module includes discretized state transition equations. S3. Input the radar echo data to be detected into the trained selective state-space model autoencoder network, output the reconstructed signal, calculate the time-step reconstruction error between the original signal and the reconstructed signal as the detection statistic, set an adaptive detection threshold based on the detection statistic, and determine the time period when the detection statistic exceeds the threshold as the target area.

[0019] S1 includes the in-phase and quadrature components of the echo data output by the radar receiver; The input channel formats for recombining into neural networks include using the in-phase and quadrature components as two separate input channels, or combining the in-phase and quadrature components into a complex signal as a single input channel.

[0020] Constructing the input feature sequence : ; In the formula, For input feature index, For the first Each input feature.

[0021] S2 includes a selective state-space model module that includes discretized state transition equations: ; ; In the formula, For the first The hidden state vector at each time step This is the current output; This is the discretized state transition matrix; For depend on The generated input projection matrix, For depend on The generated output projection matrix.

[0022] The selective state-space model module adopts the Mamba architecture, and layer normalization and residual links are set between the various Mamba modules; The encoder and decoder consist of several layers of stacked Mamba modules; The input mapping layer, bottleneck layer, and output mapping layer are linear transformation layers.

[0023] The Mamba module receives input features and divides them into two branches. The first branch performs one-dimensional convolution and gated activation sequentially, while the second branch first performs selective projection to generate... and Then, the selective projection results, , and The discretized state transition equation is used to process the data, resulting in the second branch processing result. The second branch processing result is then multiplied element-wise with the first branch result to obtain the output feature. A skip connection is set between the encoder and the decoder to concatenate or add the output features of each layer of the encoder with the input features of the corresponding layer of the decoder element by element. The output mapping layer maps the decoder output to 1 dimension through linear transformation to obtain the reconstructed signal.

[0024] The reconstruction loss function is a composite loss function, including temporal reconstruction loss. and frequency domain consistency loss : ; ; In the formula, The original signal, To reconstruct the signal, The length of the time series; Through the and The error between the spectra is obtained by performing Fourier transforms on each spectrum.

[0025] S3 includes inputting the radar echo data to be detected into a trained selective state-space model autoencoder network, and outputting... ,calculate and The time-step reconstruction error between them is used as the detection statistic. : ; Set up a sliding window, for Applying a moving average filter yields the averaged detection statistic. ;calculate mean and standard deviation And set an adaptive detection threshold. : ; In the formula, This is the threshold coefficient; Will The time period is determined to include the target area.

[0026] For echo segments identified as pure clutter, continuously collect and add them to the training dataset, and perform incremental training at a learning rate lower than the initial learning rate.

[0027] This invention also proposes a sea surface target detection device based on a selective state-space model autoencoder, which is implemented based on the sea surface target detection method described above. It includes a data preprocessing module for normalizing and windowing radar echo data; a selective state-space model autoencoder module, comprising an encoder, a low-dimensional bottleneck layer, and a decoder composed of multiple stacked selective state-space model modules, wherein the input and output projection parameters of the selective state-space model module are dynamically generated depending on the current input; a training module for training the autoencoder module using pure sea clutter data as training samples with the goal of minimizing reconstruction loss; a target detection module for calculating the reconstruction error between the original signal and the reconstructed signal as a detection statistic and making target decisions based on an adaptive threshold; and an online adaptive update module for continuously collecting echo segments determined to be pure clutter during operation and periodically incrementally training the autoencoder module with a fine-tuned learning rate, enabling the model to adapt to changes in the statistical characteristics of sea clutter caused by sea state changes.

[0028] The present invention also proposes a computer-readable storage medium in which the method of the present invention is implemented when a computer program is executed by a processor.

[0029] The core technical concept of this invention is to construct an autoencoder network using a selective state-space model, and to efficiently model the temporal evolution of sea clutter signals using a selective state-space mechanism. There is a deep technical fit between the selection mechanism of the selective state-space model and the characteristics of sea clutter signals: First, sea clutter signals exhibit significant long-range temporal correlation, with their time-domain echo amplitude being periodically modulated by wave motion, and the correlation spanning hundreds to thousands of pulse periods. The selective state-space model naturally possesses long-range memory capabilities through its hidden state recursion mechanism, and its computational complexity is only linear. Second, the Doppler spectrum of sea clutter has a finite broadening characteristic centered on the Bragg frequency. The state transition matrix of the state-space model can be interpreted as a set of damped oscillators, naturally suitable for encoding such quasi-periodic signals. Third, sea surface target signals exhibit abrupt amplitude changes relative to the clutter background. After the autoencoder is trained only on pure clutter data, it cannot effectively reconstruct the target component, thus naturally highlighting the target in the reconstruction error, achieving target determination in the sense of out-of-distribution sample detection.

[0030] The selective projection of this invention is to... A learnable linear layer (fully connected layer) maps the model to a high-dimensional feature space, resulting in projected feature vectors. Selective projection generates dynamic parameters, which serve as inputs for state transitions. Therefore, the input projection matrix in the state-space model... and output projection matrix It's not fixed, but depends on the input at the current time step. Dynamically generated.

[0031] The following description, in conjunction with the accompanying drawings, provides further details. A flowchart of the technical process of the present invention is shown below. Figure 1 As shown, the process includes a training phase and a detection phase. The training phase involves data acquisition and preprocessing to construct a pure sea clutter training set, which is then input into a selective state-space model autoencoder model, ultimately outputting the trained model parameters. The selective state-space model autoencoder model sequentially includes an input mapping layer, an encoder, a bottleneck layer, a decoder, and an output mapping layer. In the detection phase, the radar echo to be detected is input into the trained network model, a reconstruction error calculation threshold is set, and then comparison and decision are performed to ultimately complete target detection.

[0032] The selective state-space model autoencoder network architecture of this invention is as follows: Figure 2 As shown, this includes inputting the radar echo sequence into a mapping layer, and then into an encoder, where the encoder is... The encoder consists of several selective state-space model modules, which are added element-wise to the corresponding selective state-space model modules of the decoder through skip connections. The encoder processing result is input into the bottleneck layer, which is mapped to a low-dimensional feature space and then input into the decoder. The decoder processing result is input into the output mapping layer to obtain the reconstructed signal.

[0033] The internal structure of the selective state-space model module of this invention is as follows: Figure 3 As shown, the input sequence is fed into the selective spatial model module, which is divided into two branches. The first branch performs one-dimensional convolution and gated activation sequentially, while the second branch first performs selective projection to generate... and Then, the selective projection results, , and The discretized state transition equation is used to process the data, resulting in the second branch processing result. The second branch processing result is then multiplied element-wise with the first branch result to obtain the output feature.

[0034] The principle of reconstruction error detection in this invention is as follows: Figure 4 As shown, the radar echo sequence to be detected is input into a trained selective state-space model autoencoder, which outputs a reconstructed signal sequence. The reconstruction error is calculated point by point between the reconstructed signal sequence and the original signal sequence to obtain a reconstruction error sequence. The reconstruction error sequence is analyzed for statistical characteristics to obtain an adaptive detection threshold. The reconstruction error sequence is then tested to determine whether the error is greater than the adaptive threshold. If it is, the region is identified as containing the target, and the detection result is output. If not, the region is identified as still broadcasting, and sea clutter suppression is performed.

[0035] The following description, in conjunction with an embodiment, provides further details. In this embodiment, the radar operates at a frequency of 9.39 GHz and employs a low grazing angle observation system. The data includes measured echoes under various sea state conditions; each data set contains a time-domain echo sequence of 34 range cells, where the locations of the target range cells are marked.

[0036] Multiple sets of measured data from the dataset were selected, with each range cell containing in-phase and quadrature data from thousands of pulses. Echo amplitude values ​​were calculated for each range cell. Zero-mean unit variance standardization was applied to the echo amplitude sequences: the mean of each sequence was subtracted and divided by the standard deviation to eliminate amplitude differences between different range cells caused by propagation attenuation and other factors. The normalized data were then truncated into time series samples using a sliding window of 1024 pulses with a 50% overlap rate. Based on the dataset annotation information, range cell data without targets were selected to form the training dataset; range cells containing targets were used only for testing. All pure clutter samples were randomly divided into training and validation sets in an 8:2 ratio.

[0037] A selective state-space model autoencoder network is constructed. The input mapping layer receives the normalized echo sequence and maps it from 1D amplitude values ​​to a 128D feature space through a linear transformation. The encoder consists of three layers of selective state-space model modules stacked sequentially. The core of each module is the selective state-space transformation, with input projection parameters... and output projection parameters It is not a fixed parameter, but rather determined by the input at the current time step. Dynamically generated through learnable linear transformations. This input-dependent parameter generation mechanism endows the network with selective information filtering capabilities: it efficiently encodes regular clutter components into the hidden state, while suppressing the writing of anomalous components that deviate from the clutter distribution.

[0038] The selective state-space model module is implemented using the Mamba architecture. Each Mamba module contains a one-dimensional causal convolutional layer and a gated activation mechanism. The one-dimensional causal convolutional kernel size is set to 4 to extract local temporal features. The gated activation mechanism achieves non-linear modulation of features through element-wise multiplication. Layer normalization and residual connections are set between adjacent modules to stabilize deep network training. The expansion factor is set to 2.

[0039] The bottleneck layer is a linear transform layer that compresses the encoder output from 128 dimensions to 32 dimensions. This information bottleneck forces the network to retain only the most essential regularities of the clutter signal within a limited representational capacity. Since the training data contains only pure clutter, the low-dimensional representation space learned by the bottleneck layer only characterizes the temporal statistical regularities of the clutter.

[0040] The decoder structure is symmetrical to the encoder, consisting of three stacked Mamba modules. An element-wise additive skip connection is established between the encoder and decoder, directly transferring features from each layer of the encoder to the corresponding layer of the decoder, thus improving the fidelity of clutter detail reconstruction.

[0041] The network training uses pure sea clutter samples as both network input and training labels, employing an autoencoder training paradigm. Frequency domain consistency loss is calculated by performing Fourier transforms on both the original and reconstructed signals and then calculating the mean square error between their spectra. The introduction of frequency domain loss ensures that the reconstructed signal maintains spectral consistency with the original clutter. The weighting coefficients of the composite loss function... The initial learning rate was set to 0.1. Training employed an adaptive moment estimation optimizer with an initial learning rate of 0.001, coupled with a cosine annealing decay strategy. The batch size was set to 64. During training, the reconstruction loss on the validation set was monitored, and a convergence condition was set to trigger an early stopping mechanism when the validation loss no longer decreased after 20 consecutive rounds, saving the optimal model. After training, the network had fully learned the temporal distribution characteristics of pure sea clutter signals and was able to perform high-fidelity reconstruction of clutter components.

[0042] The trained network is applied to the radar echo data to be detected. The echo sequence, after undergoing the same normalization and windowing preprocessing, is input into the network to obtain the reconstructed signal. The reconstruction error at each time step is calculated as the detection statistic. Since the network is trained only on pure clutter data, the learned low-dimensional representation space only covers the clutter distribution. For pure clutter time periods, the input signal is consistent with the distribution learned by the network, resulting in high reconstruction accuracy and a low detection statistic. For time periods containing targets, the target signal is a sample outside the network's learned distribution, and the network cannot effectively reconstruct it, leading to a significant increase in the detection statistic. This detection principle is essentially out-of-distribution sample detection. A moving average filter of 10 pulses is applied to the detection statistic sequence to obtain a smoothed detection statistic, reducing false alarms caused by clutter fluctuations. An adaptive detection threshold is set based on the statistical characteristics of the detection statistic in pure clutter regions. When the smoothed detection statistic exceeds the threshold T, the presence of a target signal in that time period is determined. In actual operation, echo segments identified as pure clutter can be continuously collected, and the network can be incrementally trained periodically with a small learning rate to adapt the model to changes in clutter statistical characteristics caused by sea state changes, thus achieving online adaptive updates.

[0043] The method of this invention is not limited to a specific radar operating frequency band or system. When applied under different radar systems, the sliding window length needs to be adjusted according to the pulse repetition frequency and clutter correlation time, and the hyperparameters such as the number of network layers and bottleneck dimension need to be adjusted appropriately according to the clutter characteristic complexity. The overall process of network training and detection decision remains unchanged.

[0044] The selective state-space model module described in this invention is not limited to the specific implementation of the Mamba architecture. In other embodiments, a state-space model architecture with an input-dependent parameter generation mechanism is acceptable as long as the following condition is met: the input projection parameters and output projection parameters are dynamically generated depending on the current input signal, thereby endowing the network with selective coding capability for clutter signals and enabling it to have a higher reconstruction error for target components that deviate from the clutter distribution. All the above equivalent substitutions are within the protection scope of this invention.

[0045] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A radar target detection method based on a state-space model autoencoder, characterized in that, include: S1. Acquire time-domain echo data of multi-range cells of the sea radar, normalize the echo data, and reorganize it into the input channel format of the neural network. Use a sliding window method to extract time series samples of a fixed length and extract pure sea clutter echo samples without target signals to form a training dataset. S2. Based on the training dataset, construct a selective state-space model autoencoder network. With the goal of minimizing the reconstruction loss function, train the neural network and iteratively optimize the network parameters until convergence to obtain the trained selective state-space model autoencoder network. The selective state-space model autoencoder network consists of an input mapping layer, an encoder, a bottleneck layer, a decoder, and an output mapping layer. Both the encoder and decoder are composed of several layers of selective state-space model modules stacked together. The selective state-space model module includes discretized state transition equations. S3. Input the radar echo data to be detected into the trained selective state-space model autoencoder network, output the reconstructed signal, calculate the time-step reconstruction error between the original signal and the reconstructed signal as the detection statistic, set an adaptive detection threshold based on the detection statistic, and determine the time period when the detection statistic exceeds the threshold as the target area.

2. The radar target detection method based on a state-space model autoencoder according to claim 1, characterized in that, S1 includes the in-phase and quadrature components of the echo data output by the radar receiver; The input channel formats for recombining into neural networks include using the in-phase and quadrature components as two separate input channels, or combining the in-phase and quadrature components into a complex signal as a single input channel.

3. The radar target detection method based on a state-space model autoencoder according to claim 2, characterized in that, Constructing the input feature sequence : ； In the formula, For input feature index, For the first Each input feature.

4. The radar target detection method based on a state-space model autoencoder according to claim 3, characterized in that, S2 includes a selective state-space model module that includes discretized state transition equations: ； ； In the formula, For the first The hidden state vector at each time step This is the current output; This is the discretized state transition matrix; For depend on The generated input projection matrix, For depend on The generated output projection matrix.

5. The radar target detection method based on a state-space model autoencoder according to claim 4, characterized in that, The selective state-space model module adopts the Mamba architecture, and layer normalization and residual links are set between the various Mamba modules; The encoder and decoder consist of several layers of stacked Mamba modules; The input mapping layer, bottleneck layer, and output mapping layer are linear transformation layers.

6. The radar target detection method based on a state-space model autoencoder according to claim 5, characterized in that, The Mamba module receives input features and divides them into two branches. The first branch performs one-dimensional convolution and gated activation sequentially, while the second branch first performs selective projection to generate... and Then, the selective projection results, , and The discretized state transition equation is used to process the data, resulting in the second branch processing result. The second branch processing result is then multiplied element-wise with the first branch result to obtain the output feature. A skip connection is set between the encoder and the decoder to concatenate or add the output features of each layer of the encoder with the input features of the corresponding layer of the decoder element by element. The output mapping layer maps the decoder output to 1 dimension through linear transformation to obtain the reconstructed signal.

7. The radar target detection method based on a state-space model autoencoder according to claim 6, characterized in that, The reconstruction loss function is a composite loss function, including temporal reconstruction loss. and frequency domain consistency loss : ； ； In the formula, The original signal, To reconstruct the signal, The length of the time series. These are the weighting coefficients of the composite loss function; Through the and The error between the spectra is obtained by performing Fourier transforms on each spectrum.

8. The radar target detection method based on a state-space model autoencoder according to claim 7, characterized in that, S3 includes inputting the radar echo data to be detected into a trained selective state-space model autoencoder network, and outputting... ,calculate and The time-step reconstruction error between them is used as the detection statistic. : ； Set up a sliding window, for Applying a moving average filter yields the averaged detection statistic. ;calculate mean and standard deviation And set an adaptive detection threshold. : ； In the formula, This is the threshold coefficient; Will The time period is determined to include the target area.

9. The radar target detection method based on a state-space model autoencoder according to claim 8, characterized in that, For echo segments identified as pure clutter, continuously collect and add them to the training dataset, and perform incremental training at a learning rate lower than the initial learning rate.

Images