Chaotic sparrow search and attention bidirectional time domain convolution network-based sea clutter weak target detection method and system

Abstract

The application discloses a sea clutter weak target detection method and system based on a chaotic sparrow search and attention bidirectional time domain convolution network, first acquires a radar echo one-dimensional time sequence and phase space reconstruction; constructs an attention bidirectional time domain convolution network model which fuses bidirectional dilated convolution, time domain self-attention and adaptive weighted error loss; adopts a chaotic sparrow search algorithm to globally optimize key superparameters of the model; uses optimal superparameters to train the model on pure sea clutter data, inputs a sequence to be detected to perform single-step prediction, and obtains a time domain prediction error sequence; finally, performs frequency domain transformation on the sequence, and judges whether a weak target exists by analyzing whether an energy peak different from background noise exists in a spectrum. The application can fully utilize sea clutter context information, adaptively focus on key features, overcome the defect that a traditional optimization algorithm is prone to local optimization, and significantly improves weak target detection performance under a low signal-to-noise ratio.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing, specifically relating to a method and system for detecting weak targets in sea clutter based on a chaotic sparrow search and attention-based bidirectional temporal convolutional network. Background Technology

[0002] In fields such as ocean surveillance, maritime safety, and sea surface detection, radar is the core sensor for acquiring information about sea surface targets. However, the echoes received by radar during sea surface detection not only contain target signals but also background signals generated by the reflection of the sea surface itself, i.e., sea clutter. Sea clutter exhibits strong nonlinearity and non-stationarity; under complex sea conditions, its amplitude can fluctuate wildly and randomly, forming "sea spikes" that are highly similar to the echo characteristics of real, weak targets (such as small boats and buoys). This phenomenon makes it easy for traditional detection methods to misidentify strong clutter as targets or miss real targets submerged in strong clutter, severely limiting the detection performance and reliability of radar systems.

[0003] With the development of nonlinear science, research has shown that sea clutter is essentially a high-dimensional, complex chaotic time series. Based on this understanding, the academic community has pioneered the idea of ​​target detection based on nonlinear prediction: constructing a model that can accurately predict the evolution of pure sea clutter. When a target signal appears, because it does not conform to the inherent dynamic laws of sea clutter, the model prediction will inevitably fail, resulting in significant prediction errors, thereby achieving target detection.

[0004] However, these prediction-based approaches still face a series of serious challenges in practical applications. First, the expressive power of existing prediction models is still insufficient. Early models such as Radial Basis Function Networks (RBFN) or Long Short-Term Memory Networks (LSTM) have network structures that are insufficient to fully characterize the strong nonlinearity and long-range correlation of sea clutter. Even Temporal Convolutional Networks (TCN), which have improved in long-range dependency modeling in recent years, cannot simultaneously utilize the contextual information of the sequence due to their inherent unidirectional structure, resulting in a need to improve the prediction accuracy of common local abrupt changes and detailed features such as "sea spikes" in sea clutter. Second, the performance of these deep models is highly dependent on the precise configuration of hyperparameters such as learning rate and number of network layers, and the optimization process is extremely difficult. The high-dimensional complexity of sea clutter signals makes the hyperparameter optimization space huge and non-convex. Traditional manual parameter tuning methods are inefficient and subjective. Existing intelligent optimization algorithms, such as Particle Swarm Optimization (PSO) or Standard Sparrow Search (SSA), suffer from poor population diversity and premature convergence when dealing with such complex problems, often only finding local optima and failing to fully explore the performance limits of the prediction model. More critically, existing technologies also have inherent limitations in distinguishing between outliers and real targets. Traditional loss functions tend to treat outliers like "sea peaks" as noise during model training, rather than learning their unique evolutionary patterns. This directly leads to the trained model still being unable to effectively distinguish between prediction errors caused by real targets and those caused by "sea peaks" during testing, resulting in a persistently high false alarm rate.

[0005] Therefore, there is an urgent need for a detection method that can comprehensively address the above challenges in order to achieve accurate and stable detection of weak targets against a background of strong sea clutter. Summary of the Invention

[0006] To address the problems of low detection accuracy, high false alarm rate, and difficulty in model optimization of weak sea surface signals under sea clutter backgrounds, the present invention aims to provide a method and system for detecting weak targets under sea clutter based on chaotic sparrow search and attention-based bidirectional temporal convolutional networks, so as to achieve accurate and stable detection of weak targets under strong sea clutter backgrounds.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for detecting weak targets in sea clutter based on a chaotic sparrow search and attention-based bidirectional temporal convolutional network includes the following steps:

[0009] Step 1: Obtain the one-dimensional time series of radar echoes and reconstruct their phase space to map the one-dimensional time series data to a high-dimensional state space.

[0010] Step 2: Construct an attention-based bidirectional temporal convolutional network model that integrates bidirectional dilated convolution, temporal self-attention mechanism, and adaptive weighted error loss function to predict pure sea clutter sequences;

[0011] Step 3: The Chaotic Sparrow Search algorithm is used to globally optimize the key hyperparameters of the attention-based bidirectional temporal convolutional network model constructed in Step 2, in order to find the optimal parameter combination that minimizes the model's prediction error.

[0012] Step 4: Using the optimal hyperparameters obtained in Step 3, train the attention-based bidirectional temporal convolutional network model on pure sea clutter data; input the sequence to be detected into the model for single-step prediction, calculate the temporal prediction error sequence between the predicted value and the true value; and perform frequency domain transformation on the error sequence, and finally determine whether there is a weak target by analyzing whether there are energy peaks in the error spectrum that are different from the background noise.

[0013] Furthermore, in step 1, after obtaining the one-dimensional time series of radar echoes but before phase space reconstruction, a step of normalizing the one-dimensional time series of radar echoes is also included.

[0014] Furthermore, in step 1, the phase space reconstruction uses the CC method to determine the optimal embedding dimension. and delay time and one-dimensional time series Reconstructed into multidimensional phase space vectors , which serves as the input for subsequent models.

[0015] Furthermore, in step 2, the attention-based bidirectional temporal convolutional network model is constructed as follows:

[0016] The bidirectional dilated convolution module consists of a forward temporal convolutional network that processes the original reconstructed sequence and a backward temporal convolutional network that processes the time-reversed sequence. Each temporal convolutional network is composed of multiple layers of dilated causal convolutions stacked together, and the output features of the two networks are spliced ​​and fused.

[0017] The temporal self-attention module connects a scaled dot product attention layer to the output of the bidirectional temporal convolutional network to dynamically allocate attention weights;

[0018] The Adaptive Weighted Error Loss Function module is used to apply an adaptive weighted penalty to the prediction error during the model training phase.

[0019] Furthermore, in the bidirectional dilated convolution module, the one-dimensional dilated causal convolution operation is defined as:

[0020]

[0021] in, For the current time step The convolution output features at that point For the convolution kernel at the index The weight value at the location; Index for the current time step, The index for traversing the elements of the convolution kernel; For convolution kernel, The kernel size is the convolution kernel size. It is the expansion factor.

[0022] Furthermore, in the temporal self-attention module, the calculation formula for the scaled dot product attention layer is:

[0023]

[0024] in, These are the query, key, and value matrices, respectively; softmax is the normalized exponential activation function used to transform the calculated attention score into a probability distribution; Key matrix The feature dimension is used here as a scaling factor to prevent the inner product result from becoming too large and causing gradient vanishing.

[0025] Furthermore, the adaptive weighted error loss function adopts the following loss function:

[0026]

[0027] in, For the true value, For predicted values, These are weighted control coefficients; The total number of samples involved in the loss calculation; For the natural constant An exponential function with base 0.

[0028] Furthermore, in step 3, the chaotic sparrow search algorithm performs global optimization of the hyperparameters of the attention-based bidirectional temporal convolutional network model through the following mechanism:

[0029] Chaotic initialization: The initial population is generated using a Logistic chaotic mapping, and its iterative equation is as follows: Among them, control parameters ;

[0030] Population update strategy: Divide the sparrow population into discoverers, joiners and vigilants. Discoverers are given non-linear adaptive weights to balance global exploration and local development. Joiners move toward the optimal position or explore randomly based on their fitness. Vigilants jump out of the local optimum by forced offset when the algorithm stalls.

[0031] Fitness function construction: The root mean square error of the predictions of the attention-based bidirectional temporal convolutional network model on the validation set is used as the fitness function for each individual in the chaotic sparrow search algorithm, and it is defined as follows:

[0032]

[0033] in, Given the size of the validation set, the goal of the algorithm is to find the hyperparameter combination that minimizes the RMSE.

[0034] The key hyperparameters include learning rate, kernel size, hidden layer dimension, and number of attention heads.

[0035] Furthermore, in step 4, the final decision is as follows: if the error spectrum exhibits a broadband noise substrate with extremely low energy levels and a flat distribution, it is determined to be pure sea clutter with no target present; if the error spectrum shows one or more narrowband spikes with high signal-to-noise ratio and concentrated energy on a flat noise substrate, it is determined to be the presence of a weak target.

[0036] A weak target detection system for sea clutter based on a chaotic sparrow search and attention-based bidirectional temporal convolutional network includes:

[0037] The data preprocessing module is used to acquire the one-dimensional time series of radar echoes and reconstruct their phase space, mapping the one-dimensional time series data to a high-dimensional state space.

[0038] The prediction model building module is used to build an attention-based bidirectional temporal convolutional network model that integrates bidirectional dilated convolution, temporal self-attention mechanism, and adaptive weighted error loss function;

[0039] The hyperparameter global optimization module is used to perform global optimization of the key hyperparameters of the attention bidirectional temporal convolutional network model using the chaotic sparrow search algorithm, so as to find the optimal parameter combination that minimizes the model prediction error.

[0040] The model training and error generation module is used to train the attention-based bidirectional temporal convolutional network model on pure sea clutter data using the optimal hyperparameters, and input the sequence to be detected into the model for single-step prediction, and calculate the temporal prediction error sequence between the predicted value and the true value.

[0041] The frequency domain decision module is used to perform frequency domain transformation on the time domain prediction error sequence, and to ultimately determine whether there is a weak target by analyzing whether there are energy spikes in the error spectrum that are different from the background noise.

[0042] Beneficial effects: Compared with the prior art, the present invention has the following beneficial effects:

[0043] (1) The present invention constructs a prediction model that integrates bidirectional temporal convolutional network and self-attention mechanism, which can simultaneously utilize the contextual information of the sea clutter sequence and adaptively focus on key dynamic features through attention mechanism, thus significantly improving the prediction accuracy of strong nonlinear and non-stationary sea clutter.

[0044] (2) The present invention designs an adaptive weighted error loss function, which assigns higher weights to difficult samples such as “sea peaks” during the training phase, forcing the model to accurately learn their evolutionary rules, fundamentally enhancing the model’s ability to distinguish sea peaks from real targets and reducing the false alarm rate.

[0045] (3) The present invention proposes a chaotic sparrow search algorithm to globally optimize the key hyperparameters of the model, and uses chaotic mapping to initialize the population, which improves the ergodicity; it introduces nonlinear adaptive weights and a watchdog forced offset mechanism, which effectively overcomes the defect of traditional optimization algorithms that are prone to getting trapped in local optima, and has a fast convergence speed and high optimization accuracy.

[0046] (4) This invention combines time-domain prediction error and frequency-domain transformation decision, and achieves target detection by analyzing whether there are narrow-band spikes with concentrated energy in the error spectrum. It can still maintain excellent detection performance under extremely low signal-to-noise ratio conditions.

[0047] (5) The present invention has been verified on the measured IPIX radar database. The results show that compared with the existing methods, the prediction error of the present invention is significantly reduced, and it can still clearly distinguish weak targets in sea states with low signal-to-noise ratio, and has outstanding robustness and practicality. Attached Figure Description

[0048] Figure 1 This is an overall flowchart of the detection method of the present invention;

[0049] Figure 2 A comparison of fitness convergence curves for different optimization algorithms in the hyperparameter optimization process;

[0050] Figure 3 The target detection results are shown in the figure for #40hh, where (a) is the actual sea clutter prediction figure, (b) is the time domain prediction error and (c) is the error signal spectrum;

[0051] Figure 4 The prediction error analysis diagram for the pure clutter gate (gate 1) under #30hh is shown, where (a) is the time-domain prediction error and (b) is the error signal spectrum.

[0052] Figure 5 The graph shows the prediction error analysis of the target gate (gate 7) under #30hh, where (a) is the time-domain prediction error and (b) is the error signal spectrum. Detailed Implementation

[0053] The invention will now be further explained with reference to the accompanying drawings.

[0054] like Figure 1 As shown, this invention presents a method for detecting weak targets in sea clutter based on chaotic sparrow search and attention-based bidirectional temporal convolutional networks. This method aims to address the difficulty in detecting weak targets in strong sea clutter environments due to the nonlinear characteristics of the signal and the presence of "sea spikes," as well as the problem that traditional prediction models easily get trapped in local optima during hyperparameter optimization. This method reconstructs the phase space of the original sea clutter data and inputs it into an Attention-BiTCN network with hyperparameters optimized using the CSSA algorithm for single-step prediction. By accurately fitting the chaotic trajectory of pure sea clutter, sequences containing weak targets generate significant prediction errors. Finally, the detection of weak targets is achieved by analyzing the abrupt changes in the frequency domain of this error. Specifically, the method includes the following steps:

[0055] Step 1: Obtain the one-dimensional time series of radar echoes, normalize them, and reconstruct their phase space to map the one-dimensional time series data to a high-dimensional state space.

[0056] In this process, the phase space reconstruction uses the CC method to determine the optimal embedding dimension. and delay time and one-dimensional time series Reconstructed into multidimensional phase space vectors , which serves as the input for subsequent models.

[0057] Step 2: Construct an Attention-BiTCN model that integrates bidirectional dilated convolution, temporal self-attention mechanism, and adaptive weighted error loss function to predict pure sea clutter sequences;

[0058] The Attention-BiTCN model is specifically structured as follows:

[0059] The bidirectional dilated convolution module consists of a forward temporal convolutional network that processes the original reconstructed sequence and a backward temporal convolutional network that processes the time-reversed sequence. Each temporal convolutional network is composed of multiple layers of dilated causal convolutions stacked together, and the output features of the two networks are spliced ​​and fused.

[0060] The one-dimensional dilated causal convolution operation is defined as follows:

[0061]

[0062] in, For the current time step The convolution output features at that point For the convolution kernel at the index The weight value at the location; Index for the current time step, The index for traversing the elements of the convolution kernel; For convolution kernel, The kernel size is the convolution kernel size. The expansion factor is used. This structure can efficiently capture long-range temporal dependencies with an exponentially growing receptive field and concatenate and fuse the output features of the two networks.

[0063] The temporal self-attention module connects a scaled dot-product attention layer to the output of the bidirectional TCN, and its calculation formula is as follows:

[0064]

[0065] in, These are the query, key, and value matrices, respectively; softmax is the normalized exponential activation function used to transform the calculated attention score into a probability distribution; Key matrix The feature dimension is used here as a scaling factor to prevent the inner product from becoming too large and causing gradient vanishing. This module dynamically allocates attention weights by calculating the similarity between features at different time steps, enabling the model to adaptively focus on key time segments that play a decisive role in prediction.

[0066] Adaptive Weighted Error (AWE) Loss Function: During the model training phase, the following loss function is used instead of the traditional mean squared error:

[0067]

[0068] in, For the true value, For predicted values, These are weighted control coefficients; The total number of samples involved in the loss calculation; For the natural constant It is an exponential function with a base of 0. This loss function imposes a greater penalty on difficult samples with large prediction errors (such as sea spikes), forcing the model to make a more accurate fit.

[0069] Step 3: The Chaotic Sparrow Search (CSSA) algorithm is used to globally optimize the key hyperparameters of the attention-based bidirectional temporal convolutional network model constructed in Step 2, in order to find the optimal parameter combination that minimizes the model's prediction error.

[0070] The Chaotic Sparrow Search Algorithm (CSSA) improves upon the Standard Sparrow Search Algorithm (SSA) through the following mechanism to achieve global optimization of the hyperparameters of the Attention-BiTCN model (including learning rate, kernel size, hidden layer dimension, number of attention heads, etc.):

[0071] Chaotic initialization: using Logistic chaotic mapping (where control parameters) This method generates an initial population. It replaces the random generation of the standard SSA, ensuring a more uniform distribution of initial hyperparameter combinations in the search space and improving ergodicity.

[0072] Population update strategy: The sparrow population is divided into discoverers, joiners, and vigilants. Discoverers are given non-linear adaptive weights to balance global exploration and local development; joiners either move towards the optimal position or explore randomly based on their fitness; vigilants escape local optima by forced offset when the algorithm stagnates.

[0073] Fitness function construction: The root mean square error (RMSE) of the Attention-BiTCN model's predictions on the validation set is used as the fitness function for each individual (i.e., each set of hyperparameters) in the CSSA algorithm, defined as follows:

[0074]

[0075] in, Let be the sample size of the validation set. The goal of the algorithm is to find the hyperparameter combination that minimizes the RMSE.

[0076] Step 4: Using the optimal hyperparameters obtained in Step 3, train the attention-based bidirectional temporal convolutional network model on pure sea clutter data; input the sequence to be detected into the model for single-step prediction, calculate the temporal prediction error sequence between the predicted value and the true value; and perform frequency domain transformation on the error sequence, and finally determine whether there is a weak target by analyzing whether there are energy peaks in the error spectrum that are different from the background noise.

[0077] Specifically, the final judgment is as follows: if the error spectrum exhibits a broadband noise substrate with extremely low energy levels and a flat distribution, it is judged to be pure sea clutter and no target exists; if the error spectrum shows one or more narrowband spikes with high signal-to-noise ratio and concentrated energy on a flat noise substrate, it is judged to be the presence of a weak target.

[0078] The present invention will be further described below with reference to the embodiments.

[0079] Example 1

[0080] 2000 consecutive points of radar data were selected, normalized, and phase space reconstruction was performed using the CC method.

[0081] Initialize the Chaotic Sparrow Search (CSSA) algorithm, setting the population size to 30 and the maximum number of iterations to 50. Define the hyperparameter search space for the Attention-BiTCN model to be optimized, including the learning rate, kernel size, number of attention heads, and batch size. Use a Logistic chaotic mapping to initialize the population positions to ensure the traversal of the initial hyperparameter combinations in the search space. Use the root mean square error (RMSE) of the Attention-BiTCN model on the validation set as the fitness function of the CSSA algorithm, and perform iterative optimization until the maximum number of iterations is reached.

[0082] Output the optimal hyperparameter combination that minimizes the fitness function value. Use this optimal combination to construct the final Attention-BiTCN prediction model and train it on the training set.

[0083] Table 1 IPIX Radar Data Description

[0084]

[0085] The test set data to be detected is input into the trained optimal Attention-BiTCN model for single-step prediction. The absolute error between the predicted data and the original radar echo data is calculated to generate a time-domain prediction error sequence. A Fast Fourier Transform (FFT) is performed on this error sequence to obtain the error spectrum. By analyzing whether there are narrow-band spikes with concentrated energy that are distinct from the background noise floor in the error spectrum, the presence of a weak target is ultimately determined.

[0086] The experimental data used in this embodiment came from the IPIX radar sea clutter database of McMaster University, Canada. This radar operates in the X-band with a carrier frequency of 9.39 GHz and a pulse repetition frequency (PRF) of 1 kHz. The cooperative target in the experiment was a polystyrene foam ball approximately 1 m in diameter wrapped in wire mesh, a typical weak surface target. The experiment selected measured data covering different wind speeds and wave heights under HH and VV polarization modes; some data descriptions are shown in Table 1.

[0087] Figure 2 The graph shows a comparison of the fitness convergence curves of different optimization algorithms. To verify the superiority of the proposed CSSA algorithm in hyperparameter optimization, it was compared with Particle Swarm Optimization (PSO) and Standard Sparrow Search (SSA) algorithms. The fitness convergence curves of the three algorithms are shown below. Figure 2As shown in the figure, the PSO algorithm gets stuck in a local optimum around the 38th iteration. The standard SSA algorithm stalls after the 32nd iteration. In contrast, the CSSA algorithm proposed in this invention converges quickly to the globally lowest RMSE value around the 16th iteration. Experimental results show that the CSSA algorithm significantly outperforms both the PSO and SSA algorithms in terms of convergence speed and optimization accuracy.

[0088] Table 2 shows a performance comparison of different prediction models. To verify the performance of the Attention-BiTCN model proposed in this invention in sea clutter prediction, it was compared with traditional radial basis function neural networks (RBF), long short-term memory networks (LSTM), one-way temporal convolutional networks (TCN), and bidirectional TCN (BiTCN) on the #280 group of pure clutter data.

[0089] Table 2 Performance Comparison of Different Prediction Models

[0090]

[0091] As can be seen from the data in Table 2, the traditional RBF and LSTM have large errors when dealing with strong non-stationary sea clutter; while the CSSA-Attention-BiTCN model proposed in this invention achieves the best prediction performance, with an RMSE of only 0.013568, which is about 49.5% more accurate than the second best BiTCN model.

[0092] Figure 3 , Figure 4 and Figure 5 The images show the detection results of this invention under different signal-to-noise ratio sea states. First, the effectiveness of the method was verified using data from the #40hh group with medium to high signal-to-noise ratios, such as... Figure 3 As shown, in the sample interval where the target exists, the temporal prediction error exhibits a significant and anomalous "bulge" ( Figure 3 (a)). After converting the prediction error sequence to the frequency domain, the background clutter energy is uniformly suppressed to an extremely low baseline level, while at the Doppler frequency band corresponding to the target, it exhibits an extremely sharp high-energy peak (a). Figure 3 (b)) This achieved precise targeting.

[0093] To further evaluate the detection limit of this method under extremely harsh sea conditions, experiments were conducted using data from the #30hh group, which has an extremely low signal-to-noise ratio. Analysis was performed on a pure clutter gate without a target, and the time and frequency domain characteristics of its prediction error are as follows: Figure 4 As shown. Its time-domain prediction error amplitude is extremely small ( Figure 4 In (a), the error spectrum energy distribution is flat, forming a very low background noise base. Figure 4(b) demonstrates that the model of this invention has excellent suppression capability for pure sea clutter. Subsequently, analysis was performed on the 7th range gate containing weak targets, and the results are as follows... Figure 5 As shown. Because the target signal does not conform to the inherent evolution law of sea clutter, the model prediction fails at this point, resulting in a prediction error with significantly enhanced energy in the time domain ( Figure 5 (a)). When the time-domain prediction error sequence is transformed by FFT, as shown in (a) Figure 5 As shown in (b), the energy of the target signal is effectively coherently accumulated in the corresponding Doppler frequency band, forming a sharp peak with an extremely high signal-to-noise ratio, which stands out clearly from the flat clutter error spectrum.

[0094] Example 2

[0095] Based on the above method, this invention also provides a weak target detection system for sea clutter based on a chaotic sparrow search and attention-based bidirectional temporal convolutional network. The system includes:

[0096] The data preprocessing module is used to acquire the one-dimensional time series of radar echoes and reconstruct their phase space, mapping the one-dimensional time series data to a high-dimensional state space.

[0097] The prediction model building module is used to build an attention-based bidirectional temporal convolutional network model that integrates bidirectional dilated convolution, temporal self-attention mechanism, and adaptive weighted error loss function;

[0098] The hyperparameter global optimization module is used to perform global optimization of the key hyperparameters of the attention bidirectional temporal convolutional network model using the chaotic sparrow search algorithm, so as to find the optimal parameter combination that minimizes the model prediction error.

[0099] The model training and error generation module is used to train the attention-based bidirectional temporal convolutional network model on pure sea clutter data using the optimal hyperparameters, and input the sequence to be detected into the model for single-step prediction, and calculate the temporal prediction error sequence between the predicted value and the true value.

[0100] The frequency domain decision module is used to perform frequency domain transformation on the time domain prediction error sequence, and to ultimately determine whether there is a weak target by analyzing whether there are energy spikes in the error spectrum that are different from the background noise.

[0101] Each module of this system performs its corresponding function according to the steps in the above method embodiments. Specific implementation methods can be found in the descriptions in the method embodiments, and will not be repeated here. This system can be integrated into a radar signal processing terminal to achieve real-time detection of weak targets against a strong sea clutter background.

[0102] In summary, this invention enhances the expression of chaotic dynamic features of sea clutter through phase space reconstruction, constructs an Attention-BiTCN model that integrates bidirectional temporal convolution, self-attention mechanism and adaptive weighted error loss to make high-precision predictions of pure sea clutter, and uses CSSA to globally optimize key hyperparameters. Finally, it achieves the detection of weak targets on the sea surface by the abnormal abrupt change of prediction error in the frequency domain, and maintains good stability and robustness in complex sea conditions and low signal-to-noise ratio scenarios.

[0103] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting weak targets in sea clutter based on a bidirectional temporal convolutional network of chaotic sparrow search and attention, characterized in that, Includes the following steps: Step 1: Obtain the one-dimensional time series of radar echoes and reconstruct their phase space to map the one-dimensional time series data to a high-dimensional state space. Step 2: Construct an attention-based bidirectional temporal convolutional network model that integrates bidirectional dilated convolution, temporal self-attention mechanism, and adaptive weighted error loss function to predict pure sea clutter sequences; Step 3: The Chaotic Sparrow Search algorithm is used to globally optimize the key hyperparameters of the attention-based bidirectional temporal convolutional network model constructed in Step 2, in order to find the optimal parameter combination that minimizes the model's prediction error. Step 4: Using the optimal hyperparameters obtained in Step 3, train the attention-based bidirectional temporal convolutional network model on pure sea clutter data; input the sequence to be detected into the model for single-step prediction, calculate the temporal prediction error sequence between the predicted value and the true value; and perform frequency domain transformation on the error sequence, and finally determine whether there is a weak target by analyzing whether there are energy peaks in the error spectrum that are different from the background noise.

2. The method according to claim 1, characterized in that, In step 1, after obtaining the one-dimensional time series of radar echoes but before phase space reconstruction, the step also includes a normalization process for the one-dimensional time series of radar echoes.

3. The method according to claim 1, characterized in that, In step 1, the phase space reconstruction uses the CC method to determine the optimal embedding dimension. and delay time and one-dimensional time series Reconstructed into multidimensional phase space vectors , which serves as the input for subsequent models.

4. The method according to claim 1, characterized in that, In step 2, the attention-based bidirectional temporal convolutional network model is constructed as follows: The bidirectional dilated convolution module consists of a forward temporal convolutional network that processes the original reconstructed sequence and a backward temporal convolutional network that processes the time-reversed sequence. Each temporal convolutional network is composed of multiple layers of dilated causal convolutions stacked together, and the output features of the two networks are spliced ​​and fused. The temporal self-attention module connects a scaled dot product attention layer to the output of the bidirectional temporal convolutional network to dynamically allocate attention weights; The Adaptive Weighted Error Loss Function module is used to apply an adaptive weighted penalty to the prediction error during the model training phase.

5. The method according to claim 4, characterized in that, In the bidirectional dilated convolution module, the one-dimensional dilated causal convolution operation is defined as: in, For the current time step The convolution output features at that point For the convolution kernel at the index The weight value at the location; For convolution kernel, The kernel size is the convolution kernel size. It is the expansion factor.

6. The method according to claim 4, characterized in that, In the temporal self-attention module, the calculation formula for the scaled dot product attention layer is: in, These are the query, key, and value matrices, respectively; softmax is the normalized exponential activation function used to transform the calculated attention score into a probability distribution; Key matrix The feature dimension is used here as a scaling factor to prevent the inner product result from becoming too large and causing gradient vanishing.

7. The method according to claim 4, characterized in that, The adaptive weighted error loss function adopts the following loss function: in, For the true value, For predicted values, These are weighted control coefficients; The total number of samples involved in the loss calculation; For the natural constant An exponential function with base 0.

8. The method according to claim 1, characterized in that, In step 3, the chaotic sparrow search algorithm performs global optimization of the hyperparameters of the attention-based bidirectional temporal convolutional network model through the following mechanism: Chaotic initialization: The initial population is generated using a Logistic chaotic mapping, and its iterative equation is as follows: Among them, control parameters ; Population update strategy: Divide the sparrow population into discoverers, joiners and vigilants. Discoverers are given non-linear adaptive weights to balance global exploration and local development. Joiners move toward the optimal position or explore randomly based on their fitness. Vigilants jump out of the local optimum by forced offset when the algorithm stalls. Fitness function construction: The root mean square error of the predictions of the attention-based bidirectional temporal convolutional network model on the validation set is used as the fitness function for each individual in the chaotic sparrow search algorithm, and it is defined as follows: in, Given the size of the validation set, the goal of the algorithm is to find the hyperparameter combination that minimizes the RMSE. The key hyperparameters include learning rate, kernel size, hidden layer dimension, and number of attention heads.

9. The method according to claim 1, characterized in that, In step 4, the final decision is as follows: if the error spectrum shows a broadband noise substrate with extremely low energy level and flat distribution, it is determined to be pure sea clutter and no target exists; if the error spectrum shows one or more narrowband peaks with high signal-to-noise ratio and concentrated energy on a flat noise substrate, it is determined to be a weak target.

10. A weak target detection system for sea clutter based on a bidirectional temporal convolutional network of chaotic sparrow search and attention, characterized in that, include: The data preprocessing module is used to acquire the one-dimensional time series of radar echoes and reconstruct their phase space, mapping the one-dimensional time series data to a high-dimensional state space. The prediction model building module is used to build an attention-based bidirectional temporal convolutional network model that integrates bidirectional dilated convolution, temporal self-attention mechanism, and adaptive weighted error loss function; The hyperparameter global optimization module is used to perform global optimization of the key hyperparameters of the attention bidirectional temporal convolutional network model using the chaotic sparrow search algorithm, so as to find the optimal parameter combination that minimizes the model prediction error. The model training and error generation module is used to train the attention-based bidirectional temporal convolutional network model on pure sea clutter data using the optimal hyperparameters, and input the sequence to be detected into the model for single-step prediction, and calculate the temporal prediction error sequence between the predicted value and the true value. The frequency domain decision module is used to perform frequency domain transformation on the time domain prediction error sequence, and to ultimately determine whether there is a weak target by analyzing whether there are energy spikes in the error spectrum that are different from the background noise.

Images