Sea surface weak target detection method based on improved PatchTST
By improving the phase space reconstruction and dual-path differential input of the PatchTST model, and combining it with the PSO algorithm to optimize hyperparameters, the problem of low detection rate and efficiency of weak targets in the context of sea clutter is solved, and efficient detection under different sea areas and sea conditions is achieved.
Patent Information
- Application Number
- CN202511681059.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-17
- Publication Date
- 2026-02-27
AI Technical Summary
Existing technologies still need improvement in terms of detection rate and computational efficiency for weak targets against sea clutter backgrounds, and are difficult to effectively identify echo signals from stealth or unmanned platforms on the sea surface.
An improved PatchTST model is adopted, and an IPatchTST network is constructed through phase space reconstruction and dual-path differential input. The hyperparameters are optimized by combining the PSO algorithm, and weak target signal detection is achieved by utilizing the peaks of prediction error in the time and frequency domains.
It significantly improves detection rate and computational efficiency under different sea areas and sea conditions, showing significant advantages over traditional methods. In particular, it can accurately detect weak targets through frequency domain analysis under severe sea conditions.
Smart Images
Figure CN121578261A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of radar signal processing technology, and more specifically to a method for detecting weak targets on the sea surface based on an improved PatchTST. Background Technology
[0002] With the continuous development of technology, the understanding and monitoring of the ocean can be achieved through radar. Radar is the "eye" of the maritime battlefield and maritime detection scenarios. To further enhance the maritime detection capabilities of radar, a deeper understanding, more refined perception, and full utilization of the characteristics of sea clutter are crucial. Sea clutter refers to the clutter signals reflected from the sea surface in a radar system. Sea clutter has strong energy and exhibits non-Gaussian, non-stationary, and nonlinear characteristics. These clutter signals interfere with the detection of objects on the sea surface (such as ships, small boats, or buoys), reducing the performance of the radar system. Small maritime targets such as stealth and unmanned platforms have weak echo energy, exhibiting very low signal-to-noise ratios in the spatial, temporal, and frequency domains, making them difficult to detect reliably. Therefore, researching methods for detecting weak signals in the context of sea clutter is of great significance.
[0003] Recent studies have confirmed that sea clutter signals are not completely random, but rather typical chaotic sequences. This chaotic characteristic can be used to effectively detect whether target signals are present in sea clutter radar echoes. The method involves treating sea clutter as a nonlinear dynamic system and constructing a predictive model to simulate its time series generation process. If a target signal is present in this chaotic background, its specificity will invalidate the previously established model and lead to a large prediction error, indicating that the target signal is present in the echo.
[0004] Currently, researchers both domestically and internationally are using various methods to predict chaotic time series by establishing predictive models. Early researchers used methods such as SVM (Support Vector Machines) and BP neural networks to predict chaotic time series and detect targets, and these methods have been widely applied. Liang employed a nonlinear prediction method based on correlation dimension and a repository to detect target signals, while Shi et al. used a multi-domain, multi-dimensional feature fusion method to construct a high-dimensional feature space. In recent years, with the rapid development of neural networks, more improved neural network models have been applied to data prediction problems. Recurrent neural networks (RNNs) and long short-term memory networks (LSTMs), as improvements to traditional neural networks, have been applied to chaotic time series prediction problems. In 2022, Yan developed a novel detection technique that utilizes the prediction error of LSTM and performs frequency domain transformation analysis. By calculating the Doppler spectrum of the error signal, weak targets can be identified in the frequency domain, significantly improving the detection performance of the LSTM model under low signal-to-noise ratio conditions. De et al. constructed a hybrid monitoring model combining Convolutional Neural Networks (CNNs) and LSTMs, which achieves target discrimination by capturing the amplitude fluctuations of echo signals over time. Abdelkader et al. used a hybrid structure combining RNNs and LSTMs for time series prediction, effectively improving the accuracy of the prediction results. In 2025, Zhou Xing et al. used temporal convolutional networks (TCNs) for weak signal detection. However, these methods still need improvement in terms of detection rate and computational efficiency for weak targets on the sea surface. Summary of the Invention
[0005] The purpose of this invention is to provide a method for detecting weak targets on the sea surface based on an improved PatchTST, so as to solve the problems existing in the prior art mentioned in the background section.
[0006] To achieve the above objectives, the present invention provides the following technical solution: The sea surface weak target detection method based on the improved PatchTST includes the following steps: S1: Acquire the signal to be detected, which includes sea clutter signal and target echo signal. Normalize the signal to be detected, then reconstruct the phase space, and then divide the reconstructed signal into training set and test set. S2: Use a dual-path differential module to improve the PatchTST model and obtain the improved IPatchTST model; S3: Input the training set signal into the IPatchTST model and train the IPatchTST model; S4: Optimize the hyperparameters of the IPatchTST model using the PSO algorithm and OneCycle strategy to obtain the PSO-IPatchTST detection framework; S5: Input the test set signal into the PSO-IPatchTST detection framework to obtain the prediction result. Calculate the error between the prediction result and the true value point by point. Based on the time domain and frequency domain peak values of the prediction error, determine whether a small target exists.
[0007] Preferably, the specific steps of phase space reconstruction in S1 are as follows: For sea clutter time series The CC method is used to calculate the embedding dimension m and the delay time τ, and then the phase space sequence is reconstructed. Where n=1,2, Based on this, the correlation dimension of the singular attractor is calculated by correlation integral, and the appropriate embedding dimension and time delay window parameters are determined according to the saturation characteristics of the correlation index.
[0008] Preferably, the improvement steps of the PatchTST model in S2 are as follows: Adding a dual-path differential module to PatchTST specifically involves performing a first-order difference on the input univariate radar signal, and then combining the differential signal with the original signal as a new input; the original time series... its difference sequence ,in This indicates performing a first-order difference; zeros are padded at the ends of the difference sequences to ensure they are of the same length, and the two signals are aligned and combined into a single multivariable input as shown in Equation 1:
[0009] Dual-channel differential input enhances high-frequency abrupt change information by introducing differential features, with relatively low added uncertainty and computational cost. The differential signal serves as an auxiliary input; due to the channel independence of PatchTST, the two signals are not mixed at the data level. However, since both inputs share the same Transformer encoder, the attention weights can learn the features of both signals through backpropagation.
[0010] Preferably, the PSO-IPatchTST detection framework in S4 is as follows: The PSO algorithm is used to optimize the partition step size, sequence segment length, and feature dimension hyperparameter of the IPatchTST model, while the OneCycle strategy is used to optimize the learning rate hyperparameter.
[0011] Preferably, the specific process for improving the PatchTST model is as follows: For each univariate time input First, the patch is divided into several overlapping or non-overlapping blocks, with a patch length of P and a stride of S, generating a patch sequence. Where N is the number of patch sequences, and its calculation formula is: The patch sequence is then fed into the Transformer encoder. In the Transformer encoder, a learnable weight matrix is used. Each patch sequence is projected into a D-dimensional vector space and a learnable positional encoding is added. To preserve the timing sequence, new input is obtained. As shown in Formula 3:
[0012] For each attention head, a query, key, and value matrix is generated using a learnable weight matrix: , , ,in , , The output of each attention head for the Query matrix, Key matrix, and Value matrix. The result is obtained through calculation using Formula 4:
[0013] The multi-head attention block also includes a BatchNorm layer and an FFN network. The output signal is finally fed into a flatten layer with a linear head to obtain the prediction result. ; The model uses MSE as the loss function. The loss of each channel was collected and averaged over several time series. This represents the mean square error, calculated as shown in Formula 5: .
[0015] The preferred steps for optimizing the IPatchTST model are as follows: S41: The learning rate is gradually increased from a low value to a preset peak value in a linear manner; S42: Linearly reduce the learning rate to a level close to the initial value, allowing the model to be fine-tuned near the found optimization region; S43: Further decay the learning rate to an extremely low value to ensure stable convergence during the training process.
[0016] Preferably, the maximum learning rate is set to 0.0025.
[0017] Compared with the prior art, the beneficial effects of the present invention are: The method of this invention first reconstructs the phase space of the original sea clutter time series data, constructs the IPatchTST network using dual-path differential input, establishes a single-step prediction model for sea clutter amplitude, optimizes the hyperparameters of the model using the PSO algorithm, and utilizes the spikes in the prediction error in the time and frequency domains to detect weak target signals. Experiments on the IPIX and SDRDSP datasets show that the method has good detection performance under different sea areas and sea states. Compared with traditional methods, this method has significant improvements in both detection rate and computational efficiency. Attached Figure Description
[0018] Figure 1 This is a flowchart of the multimodal time-frequency graph fusion algorithm of the present invention.
[0019] Figure 2 This is a diagram of the overall architecture of the PatchTST invention.
[0020] Figure 3 This is the backbone network diagram of the Transformer of this invention.
[0021] Figure 4 This is the OneCycle strategy diagram of the present invention.
[0022] Figure 5 This is a graph showing the IPIX prediction error of the present invention.
[0023] Wherein: (a) is the prediction error map of the Data17 dataset; (b) is the prediction error map of the Data26 dataset; (c) is the prediction error map of the Data54 dataset.
[0024] Figure 6 This is a graph showing the prediction error of the SDRDSP in this invention.
[0025] Figure 7 This is the SDRDSP prediction error spectrum diagram of the present invention.
[0026] Where: (a) is the clutter prediction error spectrum; (b) is the target wave prediction error spectrum. Detailed Implementation
[0027] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.
[0028] Please see Figure 1-7 The present invention provides the following technical solutions: To address the insufficient detection capability of traditional methods in the context of sea clutter, this invention proposes a weak target detection method based on an improved PatchTST. This method first reconstructs the phase space of the original sea clutter time series data, constructs an IPatchTST (improved patch time series transformer) network using dual-path differential input, establishes a single-step prediction model for sea clutter amplitude, and optimizes the model's hyperparameters using the Particle Swarm Optimization (PSO) algorithm. The detection of weak target signals is achieved by utilizing the peaks in the prediction error in the time and frequency domains. Experiments on the IPIX and SDRDSP datasets demonstrate that the proposed method exhibits good detection performance under different sea areas and sea states. Compared with traditional methods, this invention shows improvements in both detection rate and computational efficiency.
[0029] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.
[0030] like Figure 1 As shown, this method consists of three parts: data processing, model building, and parameter optimization.
[0031] 1. Data Processing: I and Q components are obtained from radar data, and their L2 norm is used to obtain a one-dimensional amplitude time series. The input sequence is then normalized. Due to the chaotic characteristics of sea clutter, the signal is reconstructed in phase space with m=33 and τ=1, resulting in an m-dimensional phase space vector as shown in Equation 2. The reconstructed sequence is then divided into training and test sets; the reconstructed signal is divided into training and test sets in a 7:3 ratio.
[0032]
[0033] 2. Model Building In the field of time series forecasting, especially in long-series forecasting tasks, traditional Transformer models face core challenges such as high computational complexity, difficulty in effectively capturing local semantics, and overfitting due to multivariate mixtures. PatchTST, as an innovative variant of the Transformer architecture, effectively addresses these challenges by introducing two key design features.
[0034] First, a sequence segmentation mechanism is employed. Time series prediction aims to understand the correlations between data at different time steps. However, a single time step does not have semantic meaning like words in a sentence; therefore, extracting local semantic information is crucial for analyzing the relationships between them. Dividing the time series into segments with local semantic information not only significantly reduces the computational complexity of attention, enabling the model to handle longer historical sequences, but also enhances the model's ability to capture local temporal patterns.
[0035] Secondly, by adopting a channel independence strategy, the multivariate time series is decomposed into multiple univariate sequences and input into a Transformer encoder with shared weights. This not only significantly reduces the number of model parameters and alleviates the risk of overfitting, but also implicitly learns the common time features among different variables through the weight sharing mechanism.
[0036] Furthermore, PatchTST inherits the core advantages of Transformer, especially its powerful global dependency modeling capabilities. Through the self-attention mechanism, the model can directly establish the dependency relationship between any two time points in the sequence, thereby effectively capturing complex patterns such as long-range periodicity and trends in time series.
[0037] The overall structure of PatchTST is as follows: Figure 2 As shown, the Transformer backbone network is as follows: Figure 3 As shown.
[0038] The model processing flow is as follows: For each univariate time input First, the patch is divided into several overlapping or non-overlapping blocks, with a patch length of P and a stride of S, generating a patch sequence. Where N is the number of patch sequences, and its calculation formula is: The patch sequence is then fed into the Transformer encoder.
[0039] In the Transformer encoder, a learnable weight matrix is used. Each patch sequence is projected into a D-dimensional vector space and a learnable positional encoding is added. To preserve the timing sequence, new input is obtained. As shown in Formula 3:
[0040] For each attention head, a query, key, and value matrix is generated using a learnable weight matrix: , , ,in , , The output of each attention head for the Query matrix, Key matrix, and Value matrix. It is calculated using Formula 4.
[0041]
[0042] like Figure 2 As shown, the multi-head attention block also includes a BatchNorm layer and an FFN (feed forward network) network. The output signal is finally fed into a flatten layer with a linear head to obtain the prediction result. .
[0043] The model uses MSE as the loss function. The loss of each channel was collected and averaged over several time series. This represents the mean squared error. The calculation method is shown in Formula 5:
[0044] Regarding the model structure, the differences in structural parameters are shown in Table 1: Table 1 Differences in Model Structure
[0045] Parameters such as the number of encoder layers affect model complexity and computation speed, while parameters such as sequence segment length, partition step size, and model feature dimension affect segment length and prediction accuracy. Therefore, the PSO optimization algorithm is used to optimize these parameters.
[0046] The improvements to PatchTST in this invention mainly focus on the input structure and model structure. PatchTST, due to its segmented characteristics and channel independence, is suitable for processing multivariate sequences, while radar signal input is a univariate time series. Therefore, improvements to the model structure are necessary. The improvement involves adding a dual-path differential module to PatchTST. Specifically, this involves performing a first-order difference on the input univariate radar signal, and then combining the differential signal with the original signal as a new input. (Original time series) its difference sequence ,in This means performing a first-order difference, padding the ends of the difference sequences with zeros to ensure they are of the same length, aligning the two signals and combining them into a single multivariable input.
[0047] Dual-channel differential input enhances high-frequency abrupt changes by introducing differential features, with relatively low added uncertainty and computational cost. The differential signal serves as an auxiliary input; due to the channel independence of PatchTST, the two signals are not mixed at the data level. However, since both inputs share the same Transformer encoder, the attention weights can learn the features of both signals through backpropagation. Furthermore, certain modifications were made to the model's structural parameters.
[0048] The PatchTST model was improved by adjusting the differential input and model structure, resulting in the IPatchTST model, which comprises a three-level structure: differential input, patch, and Transformer backbone network. The processed signal sequence was input into the model for prediction, and the prediction results were inversely normalized. The relative error between the predicted and actual values was calculated point-by-point to obtain the relative error curve. An FFT was performed on the prediction error to obtain the frequency domain result, and the time and frequency domain results were combined for the final target determination. Furthermore, the RMSE (Real-Time Sequence) was used to measure the model's fit to sea clutter, as shown in Equation 6.
[0049] 3. Parameter optimization This invention combines the PSO algorithm and the OneCycle strategy to optimize hyperparameters. The sequence segment length, partition step size, and feature dimension of the model determine the granularity of time series decomposition, information overlap, and expressive power. These three parameters interact and jointly determine model performance. Manual parameter tuning rarely yields the optimal solution; therefore, the PSO algorithm is used for spatial optimization of the three parameters, with RMSE as the optimization metric. The particle count is set to 30, the maximum number of iterations to 30, and the spatial dimension to 3. Furthermore, the learning rate is dynamically adjusted using the OneCycle strategy, such as... Figure 4 As shown, the OneCycle strategy optimizes the training performance of deep learning models by designing a dynamic learning rate change process. This strategy divides the training process into three key stages: first, the learning rate is linearly increased from a low value to a preset peak, which helps the model quickly explore the parameter space and escape local optima; then, the learning rate is linearly decreased to a level close to the initial value, allowing the model to fine-tune near the found optimization region; finally, the learning rate is further decayed to an extremely low value to ensure stable convergence of the training process. This invention sets the maximum learning rate to 0.0025. The OneCycle strategy is used to find the optimal learning rate in each PSO optimization process.
[0050] This invention employs two representative radar detection datasets: IPIX data acquired in 1993 and recently released SDRDSP data. The IPIX dataset was acquired by the IPIX radar system located atop a cliff in the Dartmouth area on the east coast of Canada. This system operates at a frequency of 9.3 GHz, a pulse repetition frequency of 2000 Hz, a range sampling interval of 15 meters, and supports four polarization modes: HH, VV, HV, and VH. The observed target was a 1-meter diameter polystyrene foam sphere encased in a wire mesh. The data was obtained in sea states 3 and 4, corresponding to wave heights of 0.7 to 2.1 meters. The IPIX dataset consists of 14 distinct sub-files, each containing detailed information on target range gates. The range gate with the most prominent target characteristics is identified as the primary target range gate, while the secondary target range gate refers to adjacent gates where the target is still discernible. All other range gates besides the primary and secondary target range gates are classified as clutter gates. This invention refers to the echo from the primary target range gate as the target-containing echo, and the echo from the clutter gate as pure clutter.
[0051] The SDRDSP dataset was provided by the "Radar Oceanographic Detection Data Sharing Program" led by the Naval Aviation University. It was acquired using the Tian'ao SPPR50P radar, with polarization modes including HH and VV. The transmission sequence included three pulse types: single pulse T1, linear frequency modulated pulses T2 and T3, and a system repetition frequency of 2000 Hz. Two steel buoys were deployed as targets in the experimental scenario, located at 2.97 nautical miles and 3.19 nautical miles respectively. Each buoy had a diameter of 2.4 meters, covering sea states 3-5. The signal length was 131,000, and the total observation time was 65.5 seconds.
[0052] The HH polarization data of IPIX and SDRSDP used in this invention are shown in Table 2.
[0053] Table 2 Dataset Information
[0054] Sea clutter is the backscattered echo of radar signals originating from the local sea surface, influenced by various factors such as waves and wind. Recent studies have confirmed that sea clutter signals are not completely random, but rather typical chaotic sequences. Treating sea clutter as a nonlinear dynamic system, a predictive model is established to simulate its process. If the sea clutter contains a target signal, the previously established model will become invalid, resulting in significant prediction errors. The presence of a target signal within the measured signal can be determined by observing the error spikes.
[0055] First, three datasets (Data17, Data26, and Data54) from the IPIX dataset were selected for the experiment. The original radar IQ orthogonal signals were preprocessed, with 7500 points selected from each dataset. The training and test sets were divided in a 7:3 ratio, and the data were input into PSO-IPatchTST to obtain the prediction results. The difference between the predicted results and the true values was calculated as the prediction error. The prediction errors for the three datasets are as follows: Figure 5 As shown in the figure, blue represents the prediction error of pure clutter signals, and red represents the prediction error of signals containing target echoes. It can be seen that the prediction errors of signals containing target echoes in Data17 and Data26 have obvious peaks in the early stages, while the prediction errors of pure clutter signals fluctuate relatively little. The prediction error of the Data54 dataset containing target signals has a significant peak between 1500 and 2000 points, with a maximum peak value of 0.648, while the maximum prediction error of pure clutter is 0.320. The results demonstrate that the method of this invention can detect targets well in sea states below level 4.
[0056] To further validate the method's detection performance under adverse conditions, data from the SDRDSP dataset 20221113000008 was used. This dataset contains sea state 5, the highest sea state among currently available datasets. T1 pulses were used, with a total of 950 range gates, where the target echo is located at range gates 442-454. Range gate 1 was selected as pure clutter, and range gate 447 contained the target echo signal. The prediction error was obtained as follows: Figure 6 As shown in the figure, the prediction error of the signal containing the target fluctuates relatively large, but there is no obvious peak, making it impossible to distinguish intuitively, and the time-domain discrimination method fails. This is because of the rough sea conditions and high waves. On the one hand, the high waves cover the target interference detection effect; on the other hand, the impact of the waves generates sea spikes, affecting radar detection. This invention uses the frequency domain transformation of the prediction error for discrimination, treating the prediction error as a signal and studying its frequency domain distribution. If there is a peak, it is considered that there is a high-power component in the prediction error, i.e., a weak signal exists. The frequency domain image obtained by applying FFT to the prediction error is shown below. Figure 7 As shown in the figure, the prediction error containing the target signal has a spike at 600-700 points, while the clutter prediction error does not have this spike. Therefore, the presence of weak target signals in the sea clutter background can be determined by the frequency domain spike. Experiments have shown that combining the time and frequency domains can accurately detect weak signals.
[0057] In single-step prediction, the PSO-IPatchTST detection framework proposed in this invention achieves good results. To better verify the effectiveness and accuracy of the proposed method, three other datasets with HH polarization from the IPIX and SDRDSP datasets were selected for comparison with currently used GF-RNN, BiLSTM, and TCN methods. In the experimental settings, the training epochs for all methods were uniformly set to 100, and the batch size was set to the maximum value allowed by the GPU memory capacity. The evaluation metrics used were root mean square error (RMSE) and computation time, where computation time was defined as the total time required to complete 100 training epochs. The RMSE and computation time results for each method are shown in Table 3.
[0058] Table 3 Comparison of Predictive Performance
[0059] As can be seen from the table, Bi-LSTM has the highest computational time, which is due to its two-layer LSTM structure. LSTM is a special type of RNN, whose core feature is its recurrent connections. This means that when processing time series data, the computation at time step t heavily depends on time step t-1. Therefore, RNNs and LSTMs are not suitable for efficient parallel computing. The PSO-IPatchTST method is based on the Transformer architecture, abandons the recurrent structure, and fully utilizes the self-attention mechanism, making it highly parallelizable. Therefore, compared to other methods, the PSO-IPatchTST of this invention significantly increases computational speed and reduces computation time by approximately 58%, making it more suitable for engineering needs. In terms of prediction accuracy, PSO-IPatchTST has an average RMSE of 0.1292, compared to 0.6815 for the other three methods, indicating smaller errors and higher accuracy.
[0060] In summary, this invention solves the problem of difficulty in detecting weak signals against a sea clutter background, and has good detection results for different sea conditions.
[0061] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for detecting weak targets on the sea surface based on an improved PatchTST, characterized in that, Includes the following steps: S1: Acquire the signal to be detected, which includes sea clutter signal and target echo signal. Normalize the signal to be detected, then reconstruct the phase space, and then divide the reconstructed signal into training set and test set. S2: Use a dual-path differential module to improve the PatchTST model and obtain the improved IPatchTST model; S3: Input the training set signal into the IPatchTST model and train the IPatchTST model; S4: Optimize the hyperparameters of the IPatchTST model using the PSO algorithm and OneCycle strategy to obtain the PSO-IPatchTST detection framework; S5: Input the test set signal into the PSO-IPatchTST detection framework to obtain the prediction result. Calculate the error between the prediction result and the true value point by point. Based on the time domain and frequency domain peak values of the prediction error, determine whether a small target exists.
2. The method for detecting weak sea surface targets based on the improved PatchTST according to claim 1, characterized in that, The specific steps for phase space reconstruction in S1 are as follows: For sea clutter time series The CC method is used to calculate the embedding dimension m and the delay time τ, and then the phase space sequence is reconstructed. Where n=1,2, Based on this, the correlation dimension of the singular attractor is calculated by correlation integral, and the appropriate embedding dimension and time delay window parameters are determined according to the saturation characteristics of the correlation index.
3. The method for detecting weak targets on the sea surface based on the improved PatchTST according to claim 1, characterized in that, The improvement steps of the PatchTST model in S2 are as follows: Adding a dual-path differential module to PatchTST specifically involves performing a first-order difference on the input univariate radar signal, and then combining the differential signal with the original signal as a new input; the original time series... its difference sequence ,in This indicates performing a first-order difference; zeros are padded at the ends of the difference sequences to ensure they are of the same length, and the two signals are aligned and combined into a single multivariable input as shown in Equation 1: 。 4. The method for detecting weak sea surface targets based on the improved PatchTST according to claim 1, characterized in that, The PSO-IPatchTST detection framework in S4 is as follows: The PSO algorithm is used to optimize the partition step size, sequence segment length, and feature dimension hyperparameter of the IPatchTST model, while the OneCycle strategy is used to optimize the learning rate hyperparameter.
5. The method for detecting weak sea surface targets based on the improved PatchTST according to claim 3, characterized in that, The specific process for improving the PatchTST model is as follows: For each univariate time input First, the patch is divided into several overlapping or non-overlapping blocks, with a patch length of P and a stride of S, generating a patch sequence. Where N is the number of patch sequences, and its calculation formula is: The patch sequence is then fed into the Transformer encoder. In the Transformer encoder, a learnable weight matrix is used. Each patch sequence is projected into a D-dimensional vector space and a learnable positional encoding is added. To preserve the timing sequence, new input is obtained. As shown in Formula 3: For each attention head, a query, key, and value matrix is generated using a learnable weight matrix: , , ,in , , The output of each attention head for the Query matrix, Key matrix, and Value matrix. The result is obtained through calculation using Formula 4: The multi-head attention block also includes a BatchNorm layer and an FFN network. The output signal is finally fed into a flatten layer with a linear head to obtain the prediction result. ; The model uses MSE as the loss function. The loss of each channel was collected and averaged over several time series. This represents the mean square error, calculated as shown in Formula 5: 。 6. The method for detecting weak sea surface targets based on the improved PatchTST according to claim 4, characterized in that, The specific steps to optimize the IPatchTST model are as follows: S41: The learning rate is gradually increased from a low value to a preset peak value in a linear manner; S42: Linearly reduce the learning rate to a level close to the initial value, allowing the model to be fine-tuned near the found optimization region; S43: Further decay the learning rate to an extremely low value to ensure stable convergence during the training process.
7. The method for detecting weak sea surface targets based on the improved PatchTST according to claim 6, characterized in that, The maximum learning rate is set to 0.0025.