Method for constructing GLSTM model and application of GLSTM model in sea clutter suppression

Abstract

This invention discloses a method for constructing a GLSTM model and its application in sea clutter suppression. The GLSTM model is a combination of a generator and an LSTM network. A DCGAN is trained and the generator is extracted. The LSTM network is then combined with the trained generator to obtain the GLSTM model. The generator parameters are fixed, and the GLSTM sea clutter prediction model is trained using a training dataset. During sea clutter suppression, the real and imaginary parts of the sequence are separated and preprocessed for normalization. This invention utilizes the chaotic characteristics of sea clutter and the generator's ability to capture sea clutter distribution features, employing a GLSTM model to predict and suppress sea clutter, providing a new method for improving target detection and recognition performance in sea clutter backgrounds.

Description

Technical Field

[0001] This invention relates to a method for suppressing sea clutter in marine environments, and more particularly, to a sea clutter suppression method based on a GLSTM model. This invention combines a generator with an LSTM (Long Short-Term Memory) network to form a GLSTM model, referred to as a "generative long short-term memory network model". Background Technology

[0002] Radar operates in a marine environment, and the backscattered signals received from the sea surface are commonly referred to as sea clutter. Sea clutter often severely limits the radar's ability to detect targets on the sea surface. In June 2019, the *Journal of Radar*, Volume 8, Issue 3, published an article entitled "Design and Key Technology Research of Airborne Multifunctional Maritime Surveillance Radar System," authored by Jiang Qian, Wu Hao, and Wang Yanning. The airborne multifunctional maritime surveillance radar consists of two parts: airborne radar equipment and display / control / information processing software (hereinafter referred to as...). Figure 1 The airborne radar is installed on the aircraft platform and performs reconnaissance missions as an airborne mission device with the support of the data transmission subsystem, etc. The display and control / processing software is installed in the mission control station to monitor and control the status of the airborne radar, receive reconnaissance data sent in real time through the data transmission analysis chart, form a preliminary situation display, and send the processed reconnaissance information to the backend for intelligence processing.

[0003] Radars operating in marine environments encounter various types of clutter interference. These clutter types are diverse and complex, including ground clutter reflected from land, reefs, and islands; meteorological clutter reflected from rain, snow, and other weather phenomena; and sea clutter reflected from the sea surface. Among these, sea clutter has a high power level and often severely affects radar detection of sea surface targets such as ships and navigation buoys. The physical mechanism of sea clutter generation is very complex, with many influencing factors, such as radar parameters (including band, resolution, polarization, beamwidth, bandwidth, pulse repetition frequency, etc.), geographic geometric parameters (including platform height, azimuth, etc.), platform type (including shore-based, airborne, etc.), marine environmental parameters (including temperature, wave height, wave direction, wind direction, wind speed, etc.), and sea area (including nearshore, shallow sea, deep sea, etc.). These factors are interconnected, producing complex sea clutter characteristics. Compared to ground clutter, the spatiotemporal variation characteristics of sea clutter are also more complex, posing significant difficulties and challenges to target detection against sea clutter backgrounds. Under high resolution, low skidding angle, and high sea state conditions, sea clutter exhibits non-Gaussian, nonlinear, and non-stationary statistical characteristics, manifesting as interference signals resembling targets (such as sea spikes). Radar struggles to distinguish target signals from interference signals, leading to false alarms and missed alarms. In strong sea clutter backgrounds, for slow-moving targets with small radar cross-sections, the signal-to-clutter ratio of the received radar echo signal is low, hindering target detection. Therefore, sea clutter suppression is of significant research importance.

[0004] Sea clutter suppression is an important research direction in the radar field. Scholars both domestically and internationally have conducted extensive research on sea clutter characteristics and proposed many sea clutter suppression methods. These methods can be categorized into cancellation methods, subspace decomposition methods, wavelet transform methods, and neural network methods. Early sea clutter suppression principles were relatively simple, mainly focusing on modeling the amplitude distribution of sea clutter. Model parameters were estimated using measured sea clutter data, and then the original echo was canceled with the sea clutter obtained from the model to achieve suppression. Subspace decomposition methods separate sea clutter from the target signal based on the aggregation characteristics of clutter in subspace, and are divided into eigenvalue decomposition (EVD) and singular value decomposition (SVD) methods. Short-time Fourier transform (STFT), wavelet transform, and Wigner-Ville distribution (WVD) are time-frequency analysis methods that can describe the change of the signal spectrum over time. Wavelet transform can perform "zoom" analysis on frequency domain (or time domain) signals to describe the time-frequency variation characteristics of local signal ranges. Based on the different characteristics of sea clutter and target signals in the wavelet domain, threshold processing is used for separation to achieve sea clutter suppression. Haykin S. et al. found that sea clutter signals exhibit many characteristics of chaotic signals. Measured sea clutter data showed that sea clutter has a limited correlation dimension, a positive Lyapunov exponent, and short-term predictability. Based on Takens' theorem, Haykin established a sea clutter prediction model based on a Radial Basis Function Neural Network (RBFNN). This model also showed good sea clutter prediction performance on unlearned sea clutter datasets. Since then, researchers have begun to study a new class of sea clutter suppression methods, utilizing the chaotic short-term predictability of sea clutter to remove predicted sea clutter from the original echo to suppress it. In recent years, domestic and international scholars have applied deep learning networks to the radar field, achieving better results than traditional algorithms in clutter suppression, anti-jamming, target detection, and target recognition and tracking. Deep learning networks can also be used to further explore the deep features and fine information of sea clutter, achieve more accurate clutter sample selection, and improve the sea clutter suppression effect.

[0005] Current research on predicting sea clutter time signals based on neural network models mainly focuses on the amplitude prediction of sea clutter. The LSTM (Long Short-Term Memory) network model has achieved good results in sea clutter prediction. However, the nonlinear mapping capability of the LSTM model is limited, which restricts the prediction accuracy of sea clutter. Therefore, the sea clutter suppression method based on the GLSTM model proposed in this invention combines the LSTM network and the generator, which improves the nonlinear mapping capability of the network, helps to improve the prediction results of sea clutter, and thus improves the suppression effect of sea clutter. Summary of the Invention

[0006] One objective of this invention is to propose a method for constructing a GLSTM model, which combines an LSTM network and a DCGAN generator. The DCGAN generator is extracted by training a DCGAN model.

[0007] The second objective of this invention is to train a GLSTM model using radar echo data to obtain a GLSTM sea clutter prediction model. During training, the chaotic characteristics of sea clutter and the nonlinear mapping characteristics of the GLSTM model are utilized; based on the GLSTM model, a definite chaotic generation pattern is grasped from the seemingly random sea clutter time series to achieve sea clutter prediction.

[0008] The third objective of this invention is the application of the GLSTM sea clutter prediction model in an airborne multi-functional maritime surveillance radar system. Sea clutter suppression is achieved by canceling the predicted sea clutter results with the original echo. Compared to the original LSTM model, the GLSTM model improves both the model's nonlinear mapping capability and the sea clutter prediction accuracy, thereby enhancing the sea clutter suppression effect.

[0009] like Figure 2 The present invention illustrates a sea clutter suppression method based on a GLSTM model, which is embedded in... Figure 1 The data terminal unit is used to receive radar echo data transmitted by the airborne data terminal, denoted as... In this invention, a portion of the radar echo data is used. Used to construct the training dataset, denoted as Another part of the radar echo data Used to construct the validation dataset, denoted as .

[0010] like Figure 2 As shown, the present invention provides a sea clutter suppression method that combines a phase processing mechanism with an LSTM network model. The method specifically includes three parts: generator modeling, GLSTM sea clutter prediction modeling, and real-time sea clutter processing based on the GLSTM sea clutter prediction model (GLSTM model application). Part 1: Training the DCGAN model to extract the DCGAN generator.

[0011] Generative Adversarial Networks (GANs) consist of a generator and a discriminator. The generator captures the latent distribution of real data and generates new samples, while the discriminator is a binary classifier that determines whether the input is real data or a generated sample. The generator and discriminator learn from each other through a game, enabling the generator to generate samples that are closer to real data. (Refer to pages 427-229 of *Deep Learning*, published by Posts & Telecom Press, July 2017, by Ian Goodfellow). A deep convolutional layer is added to the GAN network to form the DCGAN (Deep Convolutional Generative Adversarial Network) model. In this invention, the DCGAN model is as follows: Figure 6 As shown, the specific steps are as follows: In step 1 of the generator, the measured radar echo data is processed by range pulse compression.

[0012] In this invention, an airborne multi-functional maritime surveillance radar system acquires measured radar echo data. The radar echo data The two-dimensional data formed after range pulse compression is denoted as radar echo-range pulse compressed data. It can be represented in matrix form as ;

[0013] Indicates the number of pulse points in the azimuth direction.

[0014] This indicates the number of sampling points in the distance direction.

[0015] This represents the first distance sampling point at the first azimuth time.

[0016] This represents the second distance sampling point at the first azimuth time.

[0017] This indicates the first position time at which the data was collected. A distance-oriented sampling point.

[0018] This represents the first distance sampling point at the second position time.

[0019] This represents the second distance sampling point at the second location time.

[0020] This indicates the second position time at which the data was collected. A distance-oriented sampling point.

[0021] Indicates the number of collected data. The first distance sampling point at each azimuth time.

[0022] Indicates the number of collected data. The second distance sampling point at each azimuth time.

[0023] Indicates the number of collected data. The first position at time _____ A distance-oriented sampling point.

[0024] Due to the obtained measured radar echo data Typically, the data is complex; therefore, range-wise normalization of range-directed compressed echoes involves dividing the complex sea clutter time series into real and imaginary parts and normalizing them separately. The real part of the sequence is then normalized. The processing method is as follows ;

[0025] Indicates the identifier of the distance sampling point, satisfying .

[0026] Indicates the first One-dimensional sea clutter pulse sequence within a distance cell .

[0027] This indicates the first position time at which the data was collected. A distance-oriented sampling point.

[0028] This indicates the second position time at which the data was collected. A distance-oriented sampling point.

[0029] Indicates the number of collected data. The first position at time _____ A distance-oriented sampling point.

[0030] Indicates to The real part is taken from all sampling points.

[0031] express The minimum value after taking the real part of all sampling points.

[0032] express The maximum value after taking the real part of all sampling points.

[0033] In this invention, the imaginary part sequence is normalized. The processing method is as follows ;

[0034] Indicates to The imaginary part is taken for all sampling points.

[0035] express The minimum value after taking the imaginary part of all sampling points.

[0036] express The maximum value after taking the imaginary part of all sampling points. Generator step 2: Obtaining the one-dimensional real part-sea clutter pulse sequence;

[0037] In this invention, radar echo-range pulse compression data is used. The real part sequence is obtained by performing range-direction extraction to obtain a one-dimensional real part-sea clutter pulse sequence. . Generator step 3: Obtaining the one-dimensional imaginary part – sea clutter pulse sequence;

[0038] In this invention, radar echo-range pulse compression data is used. The imaginary part sequence is obtained by performing range-direction sequencing to obtain a one-dimensional imaginary part-sea clutter pulse sequence. . In generator step 4, the delay time of the sea clutter time series is estimated using the autocorrelation function method;

[0039] In this invention, an autocorrelation function is used to estimate the time delay of a sea clutter time series. ,and ;

[0040] Indicates the delay time.

[0041] n represents the identifier of the azimuth sampling point, satisfying... .

[0042] Represents the nth position time at which the data was collected. A distance-oriented sampling point.

[0043] Indicates the number of collected data. The first position at time _____ A distance-oriented sampling point.

[0044] express The conjugate of complex numbers.

[0045] In this invention, the delay time It is an autocorrelation function Decrease to the initial value The value of n corresponding to the time.

[0046] In this invention, the following is adopted: For one-dimensional real part - sea clutter pulse sequence Perform delay time calculation to obtain the estimated delay time. .

[0047] In this invention, the following is adopted: For one-dimensional imaginary part - sea clutter pulse sequence Perform delay time calculation to obtain the estimated delay time. . In generator step 5, the embedding dimension of the sea clutter time series is estimated using the Cao algorithm; In this invention, the Cao algorithm for estimating the embedding dimension of sea clutter time series is expressed as follows: 1) Define the time delay vector:

[0048] ;

[0049] Indicates the delay time.

[0050] This indicates the embedding dimension that needs to be estimated.

[0051] Indicates the first Based on the embedding dimension Construct a time delay vector.

[0052] Indicates the number of collected data. The first position at time _____ A distance-oriented sampling point.

[0053] Indicates the number of collected data. The first position at time _____ A distance-oriented sampling point.

[0054] Indicates the number of collected data. The first position at time _____ A distance-oriented sampling point.

[0055] 2) Determine the embedding dimension function ,and ;

[0056] Indicates the first Based on the embedding dimension Construct a time delay vector.

[0057] Indicates the first Based on the next embedding dimension Construct a time delay vector.

[0058] Indicates the first Based on the embedding dimension Construct a time delay vector.

[0059] Indicates the first Based on the next embedding dimension Construct a time delay vector.

[0060] In the formula For Euclidian distance calculation, the distance is defined as follows: ;

[0061] Indicates the first Based on the embedding dimension The constructed time delay vector. Based on Then there is , Indicates the number of collected data. The first position at time _____ Each distance-oriented sampling point, Indicates the number of collected data. The first position at time _____ Each distance-oriented sampling point, Indicates the number of collected data. The first position at time _____ A distance-oriented sampling point.

[0062] Indicates the first Based on the embedding dimension The constructed time delay vector. Based on Then there is , Indicates the number of collected data. The first position at time _____ Each distance-oriented sampling point, Indicates the number of collected data. The first position at time _____ Each distance-oriented sampling point, Indicates the number of collected data. The first position at time _____ A distance-oriented sampling point.

[0063] Represents the time delay vector The One element, The integer is a number and its value range satisfies .

[0064] Represents the time delay vector The One element, The integer is a number and its value range satisfies .

[0065] 3) Definition The mean is ;

[0066] The value depends on the time delay. and embedding dimension .

[0067] 4) Definitions used for research from Change to The function is ;

[0068] In this invention, Increase to a certain value Then continue to increase The function value is maintained. The surrounding area hardly changed at this point. It is the estimated optimal embedding dimension. Generator step 6: Based on the estimated delay time and the optimal embedding dimension, construct the DCGAN training dataset using a one-dimensional real part-sea clutter pulse sequence.

[0069] In this invention, the DCGAN training dataset The construction method is as follows: (A) using estimated delay time and optimal embedding dimension One-dimensional real part - sea clutter pulse sequence Segmented into shorter sequences (B) Short sequences The length of the sea clutter time series input into the DCGAN model satisfies If the sea clutter time series length of a single range cell is At most, it can construct A length of Short sequences. During the construction process, each short sequence... The next sampled point value is used as its corresponding label data, therefore the maximum number of training sets that a single distance unit can construct is [number missing]. . Generator step 7: Based on the estimated delay time and the optimal embedding dimension, construct the DCGAN training dataset using the one-dimensional imaginary part – sea clutter pulse sequence.

[0070] In this invention, the DCGAN training dataset The construction method is as follows: (A) using estimated delay time and optimal embedding dimension One-dimensional imaginary part – sea clutter pulse sequence Segmented into shorter sequences (B) Short sequences The length of the sea clutter time series input into the DCGAN model satisfies If the sea clutter time series length of a single range cell is At most, it can construct A length of Short sequences. During the construction process, each short sequence... The next sampled point value is used as its corresponding label data, therefore the maximum number of training sets that a single distance unit can construct is [number missing]. . Step 8 of the generator process: Extract the generator;

[0071] In this invention, a sequence is used. After training the DCGAN model to achieve real part balance, extract the trained real part into the generator. .

[0072] In this invention, a sequence is used. After training the DCGAN model to achieve imaginary part balance, extract the trained imaginary part into the generator. . Generator step 9: Generate the GLSTM training model;

[0073] In this invention, an LSTM network is used in conjunction with a real part generator. Thus, the real part-GLSTM training model is obtained.

[0074] In this invention, an LSTM network is used in conjunction with an imaginary part generator. Thus, the imaginary part-GLSTM training model is obtained.

[0075] In this invention, the real part-GLSTM training model and the imaginary part-GLSTM training model together constitute the GLSTM training model.

[0076] In this invention, the LSTM network model is referenced from pages 248-250 of "Deep Learning" by Ian Goodfellow, published by People's Posts and Telecommunications Press. LSTM (convolutional long short-term memory network) adds long-term memory units to a gated RNN, controlling the flow of temporal information through forget gates, input gates, and output gates, and updating the network weights through backpropagation to learn long-term dependencies. RNNs using tanh and ReLU activation functions have excessively short and unstable learning steps, while LSTM neural networks can stably learn multi-step temporal relationships.

[0077] In the generator construction process of this invention, a DCGAN model is trained using a training dataset. The DCGAN model consists of a generator and a discriminator (e.g., ...). Figure 7 As shown, the generator and discriminator learn the sea clutter distribution characteristics through adversarial training. The generator's input is a random vector of a specified dimension. The dimension of this vector is consistent with the dimension of the hidden feature vector output by the LSTM, and then the generator generates sample sea clutter data. Compared with measured sea clutter data The input data is fed into the discriminator, which outputs the probability of classifying the input data as real sample data. The training objective of the discriminator is to distinguish between real and fake sea clutter data as much as possible, which means determining the probability of discrimination for real sea clutter data. The discrimination probability of the sample sea clutter data generated by the generator should be as close to 1 as possible. It should be as close to 0 as possible. The training objective of the generator is to generate data that more closely resembles the real sea clutter distribution to confuse the discriminator, that is, to make the discriminator's discrimination probability for the generated sample sea clutter data as close as possible to 0. The goal is to get as close to 1 as possible. The generator and discriminator are trained adversarially, each becoming stronger until they reach an equilibrium state. At this point, the discriminator can no longer distinguish the data generated by the generator from the real sea clutter data. In other words, the generator has learned the distribution characteristics of the real sea clutter, and the distribution of the data generated by the generator is very close to the distribution of the real sea clutter data. The second part, in constructing the GLSTM sea clutter prediction model, includes the following steps: Step 1 involves performing range pulse compression processing on the measured radar echo data.

[0078] In this invention, an airborne multi-functional maritime surveillance radar system acquires measured radar echo data. The radar echo data The two-dimensional data formed after range pulse compression is denoted as radar echo-range pulse compressed data. .

[0079] Due to the obtained measured radar echo data Typically, the data is complex; therefore, range-wise normalization of range-directed compressed echoes involves dividing the complex sea clutter time series into real and imaginary parts and normalizing them separately. The real part of the sequence is then normalized. Imaginary part sequence normalization . In step two, a range cell with a large sea clutter intensity that is not affected by the target signal is selected as the selected one-dimensional real part-sea clutter pulse sequence.

[0080] In this invention, the radar echo-range pulse compressed data is used... Select a range cell with high sea clutter intensity that is unaffected by target signal interference. If the selected cell is the [number]th [cell]... The real part echo data of each range cell is denoted as the selected one-dimensional real part-sea clutter pulse sequence. The The sequence form is represented as ;

[0081] This represents the first position time when the real part is acquired. A distance-oriented sampling point.

[0082] This represents the second position time when the real part is acquired. A distance-oriented sampling point.

[0083] Indicates the number of times the real part has been collected. The first position at time _____ A distance-oriented sampling point. In step three, a range cell with high sea clutter intensity and no interference from the target signal is selected as the selected one-dimensional imaginary part – sea clutter pulse sequence.

[0084] In this invention, the radar echo-range pulse compressed data is used... Select a range cell with high sea clutter intensity that is unaffected by target signal interference. If the selected cell is the [number]th [cell]... The imaginary part echo data of each range cell is denoted as the selected one-dimensional imaginary part-sea clutter pulse sequence. The The sequence form is represented as ;

[0085] This represents the first position time when the imaginary part is acquired. A distance-oriented sampling point.

[0086] This represents the second position time when the imaginary part is acquired. A distance-oriented sampling point.

[0087] Indicates the number of imaginary parts collected. The first position at time _____ A distance-oriented sampling point. Step four involves constructing a real part-training dataset using a one-dimensional real part-sea clutter pulse sequence based on the estimated delay time and optimal embedding dimension.

[0088] In this invention, the time delay of the sea clutter time series is estimated based on the autocorrelation function method. (i.e., generator step 4) and estimating the embedding dimension of the sea clutter time series based on the Cao algorithm. (i.e., generator step 5) for The data is processed to obtain the real part-training dataset. . Step 5 involves constructing an imaginary part training dataset using a one-dimensional imaginary part-sea clutter pulse sequence based on the estimated delay time and optimal embedding dimension.

[0089] In this invention, the time delay of the sea clutter time series is estimated based on the autocorrelation function method. (i.e., generator step 4) and estimating the embedding dimension of the sea clutter time series based on the Cao algorithm. (i.e., generator step 5) for The data is processed to obtain the training dataset used to construct the imaginary part. . Step six: Generate the GLSTM sea clutter prediction model;

[0090] In this invention, the real part of the training dataset is used. The real-part GLSTM training model is used to obtain the real-part GLSTM sea clutter prediction model.

[0091] In this invention, the imaginary part is used to train the dataset. The imaginary part-GLSTM sea clutter prediction model is obtained by training the model in the imaginary part-GLSTM training model.

[0092] In this invention, the real part-GLSTM sea clutter prediction model and the imaginary part-GLSTM sea clutter prediction model together constitute the GLSTM sea clutter prediction model.

[0093] In this invention, a trained DCGAN generator is obtained from a balanced DCGAN, and an LSTM network is combined with the DCGAN generator to obtain a GLSTM training model. The LSTM network in the GLSTM training model learns the temporal variation characteristics of the input one-dimensional sea clutter time series. Initially, the feature vector and memory unit state are randomly initialized. As time progresses, the memory unit stores the current input state, continuously updates the latest input information, and passes it on, ultimately outputting a hidden feature vector. The DCGAN generator decodes and outputs the predicted sea clutter value based on this feature vector. The third part, applying the GLSTM sea clutter prediction model to complete sea clutter suppression, includes the following steps: Step A: The measured radar echo data is processed by range pulse compression.

[0094] In this invention, an airborne multi-functional maritime surveillance radar system acquires measured radar echo data. The radar echo data The two-dimensional data formed after range pulse compression is denoted as radar echo-range pulse compressed data. .from Real part radar echo-range pulse compressed data were extracted from the data. And imaginary radar echo-range pulse compression data . Step B: Normalize each distance cell individually;

[0095] In this invention, the time delay of the sea clutter time series is estimated based on the autocorrelation function method. (i.e., generator step 4) and estimating the embedding dimension of the sea clutter time series based on the Cao algorithm. (i.e., generator step 5) for The data is processed to obtain a dataset for constructing the real part and verifying the data. .

[0096] In this invention, the time delay of the sea clutter time series is estimated based on the autocorrelation function method. (i.e., generator step 4) and estimating the embedding dimension of the sea clutter time series based on the Cao algorithm. (i.e., generator step 5) for The data is processed to obtain the dataset used for constructing the imaginary part and for verification. .

[0097] In this invention, the training dataset ( , ) and validation dataset ( , The difference lies in the fact that the radar echo data used in the aforementioned verification dataset does not include the selected one-dimensional sea clutter pulse sequence. , ). Step C: Run the GLSTM sea clutter prediction model;

[0098] In this invention, the real part-verification dataset is... After processing, the real part of the GLSTM sea clutter prediction model is input into the real part of the sea clutter prediction model to obtain the predicted real part of the sea clutter. .

[0099] In this invention, the imaginary part-verification dataset is... After processing, the imaginary part of the GLSTM sea clutter prediction model is used to obtain the predicted values ​​of the sea clutter. .

[0100] In this invention, and Combined, the predicted sea clutter value is obtained. . Step D: The predicted sea clutter value is canceled out with the original range compression echo to suppress sea clutter;

[0101] In this invention, the predicted sea clutter value Compressed echo at original distance Cancellation is performed, and during the cancellation process, a difference operation of complex numbers is executed. If the target signal exists in the original range compression echo, then after the sea clutter signal is cancelled and suppressed, the signal retained in the result is the target signal. Compared with existing technologies, the advantages of the sea clutter suppression method based on the GLSTM model in this invention are as follows:

[0102] The GLSTM model in the sea clutter suppression method of this invention is a combination of a generator and an LSTM network; DCGAN is trained and the generator is extracted, and the LSTM network is combined with the trained generator to obtain the GLSTM model, with the generator parameters fixed.

[0103] In this invention, the generator provides the GLSTM model with more effective information describing sea clutter by learning the distribution characteristics of sea clutter during the training process, thus solving the problem of limited sea clutter prediction accuracy.

[0104] Existing sea clutter suppression methods based on LSTM networks limit the prediction accuracy of sea clutter due to the limited nonlinear mapping performance of LSTM networks. The GLSTM model proposed in this invention combines LSTM networks and generators, improving the nonlinear mapping capability of the network model, enhancing the prediction accuracy of sea clutter, and further improving the sea clutter suppression effect. Attached Figure Description

[0105] Figure 1 This is a schematic diagram of the components of an airborne multi-functional maritime surveillance radar system.

[0106] Figure 2 This is a block diagram illustrating the principle of the sea clutter suppression method based on the GLSTM network model of this invention.

[0107] Figure 3 It is the autocorrelation function method for estimating the echo time delay parameter.

[0108] Figure 4 The autocorrelation function method is used to estimate the echo time delay parameter.

[0109] Figure 5 The Cao algorithm estimates the 1-embedding dimension parameters of the echo.

[0110] Figure 6 The Cao algorithm estimates the 2-embedding dimension parameters of the echo.

[0111] Figure 7 This is a schematic diagram of DCGAN training.

[0112] Figure 8 These are the inputs and outputs of the GLSTM sea clutter prediction model.

[0113] Figure 9 The images show the sea clutter suppression results for echo 1. Sub-images (a), (b), (c), and (d) represent the range-compressed echo 1, the range-compressed echo 1 imaging results, and the sea clutter suppression results, respectively.

[0114] Figure 10 This is a comparison of cross-sections before and after echo 1 sea clutter suppression (cross-section at the 4000th pulse of each distance unit).

[0115] Figure 11 The images show the sea clutter suppression results for echo 2. Sub-images (a), (b), (c), and (d) represent the range-compressed echo 2, the range-compressed echo 2 imaging results, and the sea clutter suppression results, respectively.

[0116] Figure 12 This is a comparison of cross-sections before and after echo 2 sea clutter suppression (cross-section at the 2204th pulse of each distance unit). Detailed Implementation

[0117] This embodiment is implemented based on the technical solution of the present invention, and provides detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiment. The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. This embodiment uses measured radar sea detection data, selecting two types of range-compressed echo data, with data numbers 20210106155330_01_staring and 20210106172511_02_staring, respectively. Echo 1 does not contain target signals, while echo 2 contains target signals within range cells 940-953, 1482-1492, and 2140-2810. The relevant parameters (radar echo data) for these two types of data are shown in the table below:

[0118] The specific operation process of the sea clutter suppression method using the GLSTM sea clutter prediction model of this invention is as follows:

[0119] 1. Generate a one-dimensional time series from the collected range-direction compressed echo signal by range-by-range unit, and divide it into real part echo sequence and imaginary part echo sequence, and then perform normalization preprocessing on each.

[0120] 2. Select a range cell with strong sea clutter. Here, the one-dimensional sea clutter time series selected for echo 1 is the first range cell, and the one-dimensional sea clutter time series selected for echo 2 is the 933rd range cell.

[0121] 3. The time delay parameter of the selected sea clutter time series is estimated using the autocorrelation function method. For the time delay of echo 1 The estimated value is 5; for the time delay of echo 2 The estimate is 3. The corresponding normalized autocorrelation function curve is as follows: Figure 3 and Figure 4 As shown.

[0122] 4. The Cao algorithm is applied to estimate the embedding dimension parameters of the selected sea clutter time series. For echo 1, exist After a value greater than 9, it no longer follows The value increases and changes, therefore the optimal embedding dimension parameter estimate is 10; for echo 2, exist After 13, it no longer follows The value increases and changes, therefore the optimal embedding dimension parameter estimate is 14. Correspondingly... Curves Figure 5 and Figure 6 As shown.

[0123] 5. Construct the dataset using the estimated time delay parameter and embedding dimension parameter. For echo 1, the selected sea clutter time series length is 6940, so 6890 short sequences of length 50 can be constructed as the training dataset. Using the same construction method, the corresponding validation datasets are also constructed for the remaining range units of the echo. For echo 2, the selected sea clutter time series length is 4390, so 4348 short sequences of length 42 can be constructed as the training dataset. The corresponding validation datasets are also constructed for the remaining range units of the echo.

[0124] 6. Figure 7 The diagram illustrates DCGAN training. The network hyperparameters are set with a batch size of 64, a learning rate of 0.0002, 25 training iterations, and a cross-entropy loss function. DCGAN is trained using the training dataset until it reaches equilibrium. At this point, the discriminator cannot distinguish the data generated by the generator from real sea clutter data. In other words, the DCGAN generator has learned the distribution characteristics of real sea clutter, and the distribution of the data generated by the generator closely resembles the distribution of real sea clutter data.

[0125] 7. Train the GLSTM network model using the constructed training dataset. Set the number of LSTM layers to 2, the number of hidden layer nodes to 32, the batch size to 64, the learning rate to 0.0002, and the mean squared error loss function to be used. Figure 8 This represents the input and output of a single training iteration of the network model. The network input is a short sequence, and the network output is the predicted value of that short sequence at the next time step. The mean squared error is calculated between the network's output prediction and the corresponding input label data, and the network is optimized in the direction of reducing the mean squared error until the network training ends. At this point, the network has learned the variation pattern of sea clutter and has the ability to predict sea clutter. It is worth noting that the real and imaginary part sequences are respectively input into the trained real and imaginary part sea clutter prediction models, and the prediction results are then recombined to form the complex sea clutter prediction result.

[0126] 8. Validate the trained GLSTM sea clutter prediction network model using a validation dataset. Input the validation data and output the predicted sea clutter. Cancel the sea clutter prediction results with the original echo. Figure 9 In the middle (a), the range compression echo is 1. Figure 9 (b) shows the range-compressed echo 1 imaging result. Figure 9 (c) shows the results of implementing sea clutter suppression. Figure 9 Imaging results after suppressing clutter in the middle (d) sea. Figure 10 Comparison of cross sections before and after sea clutter suppression; Figure 11 In the middle (a), the range compression echo 2 is shown. Figure 11(b) shows the range compression echo 2 imaging results. Figure 11 (c) shows the results of implementing sea clutter suppression. Figure 11 Imaging results after suppressing clutter in the middle (d) sea. Figure 12 The images show a comparison of cross-sections before and after sea clutter suppression. Calculations show that after suppression, the sea clutter energy in echo 1 decreased by 11.8136 dB, while the signal-to-clutter ratio in echo 2 increased by 5.5092 dB. Experimental results demonstrate that the sea clutter suppression method based on the GLSTM model has practical applicability.

Claims

1. A method for constructing a GLSTM model, characterized in that... It includes the following steps: In step 1 of the generator, the measured radar echo data is processed by range pulse compression. radar echo data The two-dimensional data formed after range pulse compression is denoted as radar echo-range pulse compressed data. Represented in matrix form as ; Indicates the number of pulse points in the azimuth direction; Indicates the number of sampling points in the distance direction; This represents the first distance sampling point at the first time of the first location acquisition. This represents the second distance sampling point at the first acquired location time. This indicates the first position time at which the data was collected. One distance-oriented sampling point; This represents the first distance sampling point at the second position time. This represents the second distance sampling point at the second time of the acquisition. This indicates the second position time at which the data was collected. One distance-oriented sampling point; Indicates the number of collected data. The first distance sampling point at each azimuth time; Indicates the number of collected data. The second distance sampling point at each azimuth time; Indicates the number of collected data. The first position at time _____ One distance-oriented sampling point; Due to the obtained measured radar echo data Since it is complex data, the range-wise normalization process for compressed echoes involves dividing the complex sea clutter time series into real and imaginary parts and normalizing them separately. Real part sequence normalization The processing method is as follows ; Indicates the identifier of the distance sampling point, satisfying ; Indicates the first One-dimensional sea clutter pulse sequence within a distance cell ; This indicates the first position time at which the data was collected. One distance-oriented sampling point; This indicates the second position time at which the data was collected. One distance-oriented sampling point; Indicates the number of collected data. The first position at time _____ One distance-oriented sampling point; Indicates to Take the real part of all sampling points in the middle; express The minimum value after taking the real part of all sampling points in the middle; express The maximum value after taking the real part of all sampling points in the middle; Imaginary part sequence normalization The processing method is as follows ; Indicates to Take the imaginary part of all sampling points in the middle; express The minimum value after taking the imaginary part of all sampling points in the middle; express The maximum value after taking the imaginary part of all sampling points; Generator step 2: Obtaining the one-dimensional real part-sea clutter pulse sequence; Radar echo-range pulse compression data The real part sequence is obtained by performing range-direction extraction to obtain a one-dimensional real part-sea clutter pulse sequence. ; Generator step 3: Obtaining the one-dimensional imaginary part – sea clutter pulse sequence; Radar echo-range pulse compression data The imaginary part sequence is obtained by performing range-direction sequencing to obtain a one-dimensional imaginary part-sea clutter pulse sequence. ; In generator step 4, the delay time of the sea clutter time series is estimated using the autocorrelation function method; The autocorrelation function used to estimate the time delay of sea clutter time series is: ,and ; Indicates the delay time; Indicates the identifier of the azimuth sampling point, satisfying ; Indicates the number of collected data. The first position at time _____ One distance-oriented sampling point; Indicates the number of collected data. The first position at time _____ One distance-oriented sampling point; express The conjugate of complex numbers; use For one-dimensional real part - sea clutter pulse sequence Perform delay time calculation to obtain the estimated delay time. ; use For one-dimensional imaginary part - sea clutter pulse sequence Perform delay time calculation to obtain the estimated delay time. ; In generator step 5, the embedding dimension of the sea clutter time series is estimated using the Cao algorithm; The Cao algorithm for estimating the embedding dimension of sea clutter time series is expressed as follows: 1) Define the time delay vector: ； Indicates the delay time; This indicates the embedding dimension that needs to be estimated. Indicates the first Based on the embedding dimension Constructed time delay vector; Indicates the number of collected data. The first position at time _____ One distance-oriented sampling point; Indicates the number of collected data. The first position at time _____ One distance-oriented sampling point; Indicates the number of collected data. The first position at time _____ One distance-oriented sampling point; 2) Determine the embedding dimension The function is ,and ； Indicates the first Based on the embedding dimension Constructed time delay vector; Indicates the first Based on the next embedding dimension Constructed time delay vector; Indicates the first Based on the embedding dimension Constructed time delay vector; Indicates the first Based on the next embedding dimension Constructed time delay vector; In the formula For Euclidian distance calculation, the distance is defined as follows: ; Indicates the first Based on the embedding dimension The constructed time delay vector is based on Then there is , Indicates the number of collected data. The first position at time _____ Each distance-oriented sampling point, Indicates the number of collected data. The first position at time _____ Each distance-oriented sampling point, Indicates the number of collected data. The first position at time _____ One distance-oriented sampling point; Indicates the first Based on the embedding dimension The constructed time delay vector is based on Then there is , Indicates the number of collected data. The first position at time _____ Each distance-oriented sampling point, Indicates the number of collected data. The first position at time _____ Each distance-oriented sampling point, Indicates the number of collected data. The first position at time _____ One distance-oriented sampling point; Represents the time delay vector The One element, The integer is a number and its value range satisfies ; Represents the time delay vector The One element, The integer is a number and its value range satisfies ; 3) Definition The mean is: ； The value depends on the time delay. and embedding dimension ; 4) Definitions used for research from Change to The function is ; exist Increase to a certain value Then continue to increase The function value is maintained. The surrounding area remains unchanged at this time. It is the estimated optimal embedding dimension; Generator step 6: Based on the estimated delay time and the optimal embedding dimension, construct the DCGAN training dataset using a one-dimensional real part-sea clutter pulse sequence. DCGAN training dataset The construction method is (A) using estimated delay time. and optimal embedding dimension One-dimensional real part - sea clutter pulse sequence Segmented into short sequences (B) Short sequences The length of the sea clutter time series input into the DCGAN model satisfies If the sea clutter time series length of a single range cell is Construct A length of Short sequences; during the construction process, each short sequence The next sampled point value is used as its corresponding label data, therefore the number of training sets that can be constructed by a single distance unit is . ; Generator step 7: Based on the estimated delay time and the optimal embedding dimension, construct the DCGAN training dataset using the one-dimensional imaginary part – sea clutter pulse sequence. DCGAN training dataset The construction method is (A) using estimated delay time. and optimal embedding dimension One-dimensional imaginary part – sea clutter pulse sequence Segmented into short sequences (B) Short sequences The length of the sea clutter time series input into the DCGAN model satisfies If the sea clutter time series length of a single range cell is Construct A length of Short sequences; during the construction process, each short sequence The next sampled point value is used as its corresponding label data, therefore the number of training sets that can be constructed by a single distance unit is... ; Step 8 of the generator process: Extract the generator; Use sequence After training the DCGAN model to achieve real part balance, extract the trained real part into the generator. ; Use sequence After training the DCGAN model to achieve imaginary part balance, extract the trained imaginary part into the generator. ; Generator step 9: Generate the GLSTM training model; Based on LSTM networks and combined with real part generators , thus obtaining the real part-GLSTM trained model; Based on LSTM networks and combined with virtual part generators Thus, the imaginary part-GLSTM training model is obtained; The real part-GLSTM training model and the imaginary part-GLSTM training model together constitute the GLSTM training model.

2. The method for constructing a GLSTM model according to claim 1, characterized in that... The process of constructing a GLSTM sea clutter prediction model includes the following steps: Step 1 involves performing range pulse compression processing on the measured radar echo data. radar echo data The two-dimensional data formed after range pulse compression is denoted as radar echo-range pulse compressed data. ; Due to the obtained measured radar echo data Since it is complex data, the range-wise normalization of the compressed echo involves dividing the complex sea clutter time series into real and imaginary parts and normalizing them separately. The real part of the sequence is normalized. Imaginary part sequence normalization ; In step two, a range cell that is free from sea clutter and unaffected by the target signal is selected as the selected one-dimensional real part-sea clutter pulse sequence. From radar echo-range pulse compression data Select a range cell that is free from sea clutter and interference from the target signal. If the selected cell is the [number]th [cell]... The real part echo data of each range cell is denoted as the selected one-dimensional real part-sea clutter pulse sequence. The The sequence form is represented as ; This represents the first position time when the real part is acquired. One distance-oriented sampling point; This represents the second position time when the real part is acquired. One distance-oriented sampling point; The number of times the real part has been collected is the first. The first position at time _____ One distance-oriented sampling point; In step three, a range cell that is free from sea clutter and unaffected by the target signal is selected as the selected one-dimensional imaginary part – sea clutter pulse sequence. From radar echo-range pulse compression data Select a range cell that is free from sea clutter and interference from the target signal. If the selected cell is the [number]th [cell]... The imaginary part echo data of each range cell is denoted as the selected one-dimensional imaginary part-sea clutter pulse sequence. The The sequence form is represented as ; This represents the first position time when the imaginary part is acquired. One distance-oriented sampling point; This represents the second position time when the imaginary part is acquired. One distance-oriented sampling point; Indicates the number of imaginary parts collected. The first position at time _____ One distance-oriented sampling point; Step four involves constructing a real part-training dataset using a one-dimensional real part-sea clutter pulse sequence based on the estimated delay time and optimal embedding dimension. Delay time of sea clutter time series estimated using autocorrelation function method And the embedding dimension of the sea clutter time series estimated based on the Cao algorithm. right The data is processed to obtain the real part-training dataset. ; Step 5 involves constructing an imaginary part training dataset using a one-dimensional imaginary part-sea clutter pulse sequence based on the estimated delay time and optimal embedding dimension. Delay time of sea clutter time series estimated using autocorrelation function method And the embedding dimension of the sea clutter time series estimated based on the Cao algorithm. right The data is processed to obtain the training dataset used to construct the imaginary part. ; Step six: Generate the GLSTM sea clutter prediction model; Real part - training dataset The real part-GLSTM sea clutter prediction model is obtained by training the real part-GLSTM training model. Imaginary part - training dataset The imaginary part-GLSTM sea clutter prediction model is obtained by training the imaginary part-GLSTM training model. The real-part GLSTM sea clutter prediction model and the imaginary-part GLSTM sea clutter prediction model together constitute the GLSTM sea clutter prediction model.

3. The method for constructing a GLSTM model according to claim 2, characterized in that... The GLSTM sea clutter prediction model, implemented on an airborne multi-functional maritime surveillance radar system, employs the following steps to achieve sea clutter suppression: Step A: The measured radar echo data is processed by range pulse compression. radar echo data The two-dimensional data formed after range pulse compression is denoted as radar echo-range pulse compressed data. ;from Real part radar echo-range pulse compressed data were extracted from the data. And imaginary radar echo-range pulse compression data ; Step B: Normalize each distance cell individually; Delay time of sea clutter time series estimated using autocorrelation function method And the embedding dimension of the sea clutter time series estimated based on the Cao algorithm. right The data is processed to obtain a dataset for constructing the real part and verifying the data. ; Delay time of sea clutter time series estimated using autocorrelation function method And the embedding dimension of the sea clutter time series estimated based on the Cao algorithm. right The data is processed to obtain the dataset used for constructing the imaginary part and for verification. ; Step C: Run the GLSTM sea clutter prediction model; Real part - verification dataset After processing, the real part of the GLSTM sea clutter prediction model is input into the real part of the sea clutter prediction model to obtain the predicted real part of the sea clutter. ; Imaginary part - verification dataset After processing, the imaginary part of the GLSTM sea clutter prediction model is used to obtain the predicted values ​​of the sea clutter. ; Will and Combined, the predicted sea clutter value is obtained. ; Step D: The predicted sea clutter value is canceled out with the original range compression echo to suppress sea clutter; Sea clutter prediction value Compressed echo at original distance Cancellation is performed, and during the cancellation process, a difference operation of complex numbers is executed. If the target signal exists in the original range compression echo, then after the sea clutter signal is cancelled and suppressed, the signal retained in the result is the target signal.

Images