Faint target detection method on sea surface based on improved bidirectional long short-term memory network

Abstract

The present invention discloses a method for detecting weak targets on the sea surface based on an improved bidirectional long short-term memory network, which relates to the technical field of radar signal processing. The method comprises the following steps: obtaining a signal to be detected, the signal to be detected comprising a sea clutter signal and a target echo signal, dividing the signal to be detected into training data and test data, and reconstructing a chaotic phase space of the signal to be detected; using a bidirectional long short-term memory network Bi-LSTM algorithm to initially model the signal to obtain an LSTM model; using the reconstructed phase space signal as input to the BiLSTM network, determining the length of the training data by embedding dimension and delay time, optimizing the parameters of the BiLSTM model by using an improved Newton-Raphson optimization algorithm, improving the model prediction accuracy and running speed, reducing the target detection threshold, performing single-step prediction in combination with the BiLSTM model, and using the prediction error to detect weak target signals under strong chaotic background noise.

Description

Technical Field

[0001] The present invention relates to the technical field of radar signal processing, and in particular to a method for detecting weak targets on the sea surface based on an improved bidirectional long short-term memory network. Background Art

[0002] Sea clutter refers to the clutter signals reflected from the sea surface in radar systems. These signals interfere with the detection of objects on the sea surface (such as ships, boats, or buoys), reducing the performance of the radar system. Therefore, it is of great significance to study the detection methods of weak signals in the background of sea clutter.

[0003] To address the problem of detecting weak signals on the sea surface in the presence of sea clutter, scholars at home and abroad have proposed numerous research methods. Early researchers focused on statistical theory, modeling sea clutter as distributional models such as the Rayleigh distribution, the log-normal distribution, and the composite K distribution. However, these methods suffered from issues with accuracy and generalization. At the end of the 20th century, Haykin et al. proposed that sea clutter modeling is a prediction problem and used chaotic systems to model sea clutter. With the development of statistical learning theory, support vector machines (SVM) theory emerged, providing new insights into small target detection. In 2010, Xing et al. proposed a weak signal detection method based on the LS-SVM model, optimizing the SVM model and improving prediction accuracy. Although SVM and its improved algorithms can handle certain nonlinear problems through kernel functions, they still suffer from insufficient accuracy when predicting complex dynamic nonlinear relationships, especially chaotic time series. Therefore, a method for weak target detection on the sea surface based on an improved bidirectional long short-term memory network is needed to address these issues. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for detecting faint targets on the sea surface based on an improved bidirectional long short-term memory network to solve the problems existing in the prior art raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] The method for detecting faint targets on the sea surface based on the improved bidirectional long short-term memory network includes the following steps:

[0007] S1: Acquire the signal to be detected, which includes sea clutter signal and target echo signal, divide the signal to be detected into training data and test data, and reconstruct the chaotic phase space of the signal to be detected;

[0008] S2: Use the Bi-LSTM algorithm to initially model the signal and obtain an LSTM model.

[0009] S3: Optimize the initial model using the improved Newton-Raphson optimization algorithm NRBO;

[0010] S4: The signal to be detected enters the NRBO-BiLSTM model, which outputs a predicted signal. The predicted signal is denormalized, the error between the predicted signal and the actual signal is calculated, and a relative amplitude difference curve is drawn. When the error curve reaches a peak, it can be determined that a weak signal exists.

[0011] Preferably, in said S1, the chaotic phase space reconstruction uses CC method to solve the embedding dimension m and delay time τ of the sea clutter time series. Reconstructed space sequence The correlation dimension of the chaotic characteristic strange attractor is solved using the correlation integral, and the delay time τ and the embedding dimension m are solved according to the critical radius when the correlation index is saturated.

[0012] Preferably, in S2, the specific calculation formula of the LSTM model is as follows:

[0013]

[0014]

[0015]

[0016]

[0017]

[0018]

[0019] The symbol is the input gate , the Forget Gate , output gate , the hidden layer state at the previous moment , the current hidden layer state , input at the current moment , temporary unit status , the current unit state , the unit state at the previous moment , Sigmoid activation function , hyperbolic tangent activation function tanh, the bias vector of the forget gate, input gate, candidate memory unit and output gate 、 、 、 and the weight matrix 、 、 、 The forget gate is used to determine the information to be forgotten and discarded. The memory gate uses the sigmoid function and tanh function to determine the units that need to be updated and the updated unit information. The output gate uses the sigmoid function and tanh function to determine the output value based on the unit state. Compared with the traditional unidirectional LSTM, the bidirectional long short-term memory network can capture the context information of the sequence data by introducing both forward and reverse LSTM modules.

[0020] The formula of bidirectional long short-term memory network is as follows:

[0021]

[0022]

[0023]

[0024] Where, 、 Represent the output of the forward LSTM and the output of the reverse LSTM, respectively. 、 Respectively represent the calculation function of the forward LSTM and the calculation function of the reverse LSTM, represents the input time vector, represents the forward hidden state at the previous moment, Represents the reverse hidden state at the next moment, Represents the final output of the bidirectional LSTM, using The symbol represents vector concatenation.

[0025] Preferably, in S3, the specific steps of the NRBO optimization algorithm are as follows:

[0026] S31: Set initial parameters: determine the population size, set the maximum number of iterations, and define the decision factor;

[0027] S32: Initialize the population: randomly generate the position vector of the initial population as the set of initial solutions;

[0028] S33: Evaluate the fitness function, calculate the fitness value of each individual in the population, record the best individual and this best individual, and select RMSE as the fitness function value;

[0029] S34: Apply the NRSR rule to approximate the solution and gradually determine the list of hyperparameters with the best effect: Through the iterative optimization process, NRSR can accurately adjust the hyperparameters to reduce the model prediction error and improve the overall performance;

[0030] S35: Trap avoidance operation: During the NRSR iteration process, the TAO method is used to detect and avoid falling into the local optimal solution. When the local optimal solution is detected, TAO explores a better solution by combining the positions of the current vector and the optimal vector, thereby effectively jumping out of the dilemma of the local optimal solution.

[0031] Preferably, in S4, the NRBO-BiLSTM model can predict the sea clutter chaotic system with a lower error, but will not predict the weak target signal superimposed on the clutter. For the sea clutter system, the prediction error of the model is small, while for the target echo, the model will produce a large prediction error. With the help of the error peak, the signal containing the target echo and the pure sea clutter signal can be intuitively distinguished. When an obvious error peak appears in the difference curve, it can be determined that the target signal exists.

[0032] Preferably, in S1, the distribution ratio of training data to test data is 7:3.

[0033] Preferably, the delay time τ =11, and the embedding dimension m=3.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] The present invention uses the reconstructed phase space signal as the input of the BiLSTM network, determines the length of the training data by embedding dimension and delay time, and optimizes the parameters of the BiLSTM model using an improved Newton-Raphson optimization algorithm. This improves the model's prediction accuracy and running speed, reduces the target detection threshold, and combines the BiLSTM model for single-step prediction. The prediction error is used to detect weak target signals from strong chaotic background noise. Experimental results show that compared with traditional LSTM and improved LSTM models, this method improves detection effect by more than 22%. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a detection flow chart of the present invention.

[0037] Figure 2 This is the single-step prediction diagram of the Lorenz system of the present invention.

[0038] Figure 3 This is the prediction error diagram of the Lorenz system of the present invention.

[0039] Figure 4 This is the sea clutter prediction error diagram of the present invention.

[0040] Figure 5 This is the RMSE diagram of sea clutter prediction of the present invention. DETAILED DESCRIPTION

[0041] In order to make the technical means, creative features, objectives and effects achieved by the present invention easier to understand, the present invention is further described below in conjunction with specific implementation methods.

[0042] See also Figure 1-5 , the present invention provides the following technical solutions:

[0043] The method for detecting faint targets on the sea surface based on the improved bidirectional long short-term memory network includes the following steps:

[0044] S1: 7500 consecutive points of radar data were selected and divided into training and test sets with a ratio of 7:3. The embedding dimension and delay time were determined by the CC method. The embedding dimension m was 3 and the delay function τ was 11, and the chaotic phase space was reconstructed.

[0045] S2: Normalize the data and input it into the BILSTM model for single-step prediction.

[0046] S3: Initialize NRBO parameters, determine the population size to be 30, set the maximum number of iterations to 50, and use the RMSE of the model prediction results as the fitness function.

[0047] S4: Explore new solutions according to the NRSR rule, using trap avoidance operations to avoid falling into local optimality. For the nth hyperparameter group in the tth iteration, apply the NRSR rule to explore individual positions. The new position is expressed as:

[0048]

[0049] Where, new locations explored for application of NRSR rules; and are the number of iterations and individual serial numbers respectively; and Represent random numbers between (0,1), and the TAO is introduced by setting the optimal position and the vector position obtained by NRSR Combining to generate solutions with enhanced quality , by comparing the random number rand and DF value in the middle of (0,1), a new value is generated. The DF value is usually 0.6, and the new value generated is:

[0050]

[0051]

[0052] Where, and They are and A random number between and is a random number, and the Mean function represents the mean.

[0053] S5: Repeat the above two steps until the maximum number of iterations is reached. Set the BILSTM model according to the obtained hyperparameters and train the data to obtain predicted data. Denormalize the predicted data and calculate the relative difference between the predicted result and the amplitude of the original signal to obtain the prediction error. Use the prediction error to determine whether a weak target signal exists.

[0054] S6: Calculate the RMSE value of each range gate, determine the minimum RMSE value of the target gate and the maximum RMSE value of the clutter gate. The interval between them is the detection threshold selectable interval. The larger the selectable interval, the better the weak detection effect. Due to the small RMSE value, the ratio of the minimum RMSE value of the target gate to the maximum RMSE value of the clutter gate is defined as the measurement indicator, which is called the "RMSE maximum value ratio". The formula is as follows:

[0055]

[0056] Where, Indicates the RMSE maximum value ratio, Indicates the minimum RMSE value of the target gate, Indicates the maximum RMSE of the clutter gate; the maximum RMSE ratio can quantify the discrimination between the target and the clutter. The larger the value, the higher the discrimination between the weak signal and the sea clutter, and the better the fault tolerance and detection effect of the model.

[0057] The present invention uses the IPIX radar's Dartmouth 1993 dataset. This data was collected from a clifftop near Dartmouth, Nova Scotia, on Canada's east coast using the IPIX radar. The radar transmit frequency was 9.3 GHz, the pulse repetition frequency was 2000 Hz, the sampling interval was 15 m, and the polarization modes were HH, VV, HV, and VH. The target to be measured was a small ball wrapped in aluminum wire with a diameter of 1 m. Data sets 17, 26, and 54 of the IPIX dataset in the HH polarization mode were used. This data covers different sea conditions and signal-to-noise ratios and is representative.

[0058] Figure 2 This is the single-step prediction result of the Lorenz system. The Lorenz system is a classic three-dimensional chaotic system, and its mathematical description is as follows:

[0059]

[0060] Where: system parameters σ = 16, r = 45.92, b = 4, the initial values ​​of the system are x = 1, y = 0, z = 0.1, and the fourth-order Runge-Kutta method with a step size of 0.01 is used to solve the equation.

[0061] Figure 3 This is the prediction error diagram of the Lorenz system. The neural network can fit the chaotic system and realize short-term prediction, so the overall prediction error is very small. However, the weak signal is an external signal superimposed on the chaotic system and cannot be detected by the neural network, so a large error will occur here. Figure 4 It can be seen that there are two obvious error peaks near 500 points, from which it can be judged that there is a weak transient signal in this area.

[0062] Figure 4 This is the IPIX dataset prediction error map, from Figure 4 It can be seen that under the conditions of groups 17, 26, and 54, there is a significant difference in the prediction error between the no-target range gate and the main target range gate. In the no-target range gate, there is no spike in the prediction error, while in the main target range gate, a clear spike can be clearly seen, indicating the presence of a weak target signal.

[0063] Figure 5 The RMSE graph is predicted for the IPIX dataset. Considering that RMSE reflects the degree of model fit to the chaotic system, the problem in the radar system is essentially a binary classification problem, that is, distinguishing between targets and non-targets, so more attention should be paid to the difference between the target gate and the clutter gate. Figure 5 In the figure, the echo containing the target signal is marked with red dots, and the pure clutter signal is marked with blue dots. The dotted area is the optional threshold space. It can be seen that the RMSE value of the primary target range gate is the largest, significantly higher than the clutter gate, while the difference between the secondary target range gate and the clutter gate is relatively small. Therefore, the key factor affecting target detection is the difference between the minimum value of the target range gate and the maximum value of the clutter range gate.

[0064] Table 1 compares the characteristics of several LSTM methods:

[0065] Table 1 Comparison of various detection methods

[0066] #17 ExRatio Preprocessing denoising Number of iterations LSTM 1.60 no 10000 WOA-LSTM 2.32 yes 300 The present invention 2.85 no 250

[0067] In Table 1, the WOA-LSTM method undergoes wavelet denoising before prediction, and its computational cost is higher than that of other methods. However, the method proposed in the present invention has the largest RMSE maximum value ratio and uses the least number of iterations, and has high discrimination effect and operational efficiency for target signals. The traditional LSTM method cannot directly detect the weak signal of group 54 from the time domain and requires combined analysis in the frequency domain. However, the proposed method can detect weak signals directly from the time domain, and the number of iterations used in the training model is significantly reduced from 10,000 to 250. The required training set length is reduced from 10,000 points to 5,000 points. Compared with LSTM and WOA-LSTM, the detection capabilities of the method of the present invention are improved by 78% and 22%, respectively.

[0068] The present invention designs an NRBO-BiLSTM model to perform short-term predictions on reconstructed sea clutter signals and plots a relative amplitude difference curve. The proposed model predicts the sea clutter chaotic system with low error, while failing to predict weak target signals superimposed on the clutter. While the model's prediction error for the sea clutter system is small, it generates a larger error for the target echo. Therefore, the error peak can be used to intuitively distinguish signals containing target echoes from pure sea clutter signals. The presence of a target signal can be determined when a clear error spike appears in the difference curve. While the error peak determination method is intuitive and effective, it relies on manual observation and is inefficient. Therefore, a RMSE quantification error is defined to improve detection efficiency. Comparing the RMSE values ​​of different range gates reveals that range gates containing target echoes have higher RMSE values. Based on this difference, a threshold value is determined to distinguish between target and clutter gates. In summary, the present invention solves the problem of weak signal detection in strong sea clutter, lowers the detection threshold, and improves detection accuracy.

[0069] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting faint targets on the sea surface based on an improved bidirectional long short-term memory network, characterized in that: The following steps are involved: S1: Acquire the signal to be detected, which includes sea clutter signal and target echo signal, divide the signal to be detected into training data and test data, and reconstruct the chaotic phase space of the signal to be detected; S2: Use the Bi-LSTM algorithm to initially model the signal and obtain an LSTM model. S3: Optimize the initial model using the improved Newton-Raphson optimization algorithm NRBO; The specific steps of the NRBO optimization algorithm are as follows: S31: Set initial parameters: determine the population size, set the maximum number of iterations, and define the decision factor; S32: Initialize the population: randomly generate the position vector of the initial population as the set of initial solutions; S33: Evaluate the fitness function, calculate the fitness value of each individual in the population, record the best individual and this best individual, and select RMSE as the fitness function value; S34: Apply the NRSR rule to approximate the solution and gradually determine the list of hyperparameters with the best effect: Through the iterative optimization process, NRSR can accurately adjust the hyperparameters to reduce the model prediction error and improve the overall performance; S35: Trap Avoidance Operation: During the NRSR iteration process, the TAO method is used to detect and avoid falling into local optimal solutions. When a local optimal solution is detected, TAO combines the positions of the current vector and the optimal vector to explore a better solution, thereby effectively escaping the local optimal solution dilemma. S4: The signal to be detected enters the NRBO-BiLSTM model, which outputs a predicted signal. The predicted signal is denormalized, the error between the predicted signal and the actual signal is calculated, and a relative amplitude difference curve is drawn. When the error curve reaches a peak, it can be determined that a weak signal exists. Calculate the RMSE value of each range gate, determine the minimum RMSE value of the target gate and the maximum RMSE value of the clutter gate, and the interval between them is the detection threshold selectable range. The larger the selectable range, the better the weak detection effect. Due to the small RMSE value, define the ratio of the minimum RMSE value of the target gate to the maximum RMSE value of the clutter gate as the measurement indicator, which is called the "RMSE maximum value ratio". The formula is as follows: Where, Indicates the RMSE maximum value ratio, Indicates the minimum RMSE value of the target gate, Indicates the maximum RMSE of the clutter gate; the maximum RMSE ratio can quantify the discrimination between the target and the clutter. The larger the value, the higher the discrimination between the weak signal and the sea clutter, and the better the fault tolerance and detection effect of the model.

2. The method for detecting faint targets on the sea surface based on the improved bidirectional long short-term memory network according to claim 1 is characterized in that: In S1, the chaotic phase space reconstruction uses the CC method to solve the embedding dimension m and delay time τ of the sea clutter time series. Reconstructed space sequence , The correlation dimension of the chaotic characteristic strange attractor is solved using the correlation integral, and the delay time τ and the embedding dimension m are solved according to the critical radius when the correlation index is saturated.

3. The method for detecting faint targets on the sea surface based on the improved bidirectional long short-term memory network according to claim 1 is characterized in that: In S2, the specific calculation formula of the LSTM model is as follows: The symbol is the input gate , the Forget Gate , output gate , the hidden layer state at the previous moment , the current hidden layer state , input at the current moment , temporary unit status , the current unit state , the unit state at the previous moment , Sigmoid activation function , hyperbolic tangent activation function tanh, forget gate, input gate, candidate memory unit and output gate bias vector 、 、 、 and the weight matrix 、 、 、 The forget gate is used to determine the information to be forgotten and discarded. The memory gate determines the units that need to be updated and the updated unit information through the sigmoid function and the tanh function. The output gate determines the output value through the sigmoid function and the tanh function according to the unit state. Compared with the traditional unidirectional LSTM, the bidirectional long short-term memory network can capture the context information of the sequence data by introducing both forward and reverse LSTM modules at the same time.

4. The method for detecting faint targets on the sea surface based on the improved bidirectional long short-term memory network according to claim 1 is characterized in that: In S4, the NRBO-BiLSTM model can predict the sea clutter chaotic system with a low error, but it cannot predict the weak target signal superimposed on the clutter. For the sea clutter system, the model's prediction error is small, while for the target echo, the model will produce a large prediction error. The error peak can be used to intuitively distinguish between the signal containing the target echo and the pure sea clutter signal. When a clear error peak appears in the difference curve, it can be determined that the target signal exists.

5. The method for detecting faint targets on the sea surface based on the improved bidirectional long short-term memory network according to claim 1 is characterized in that: In S1, the distribution ratio of training data to test data is 7:

3.

6. The method for detecting faint targets on the sea surface based on the improved bidirectional long short-term memory network according to claim 2, characterized in that: The delay time τ =11, and the embedding dimension m=3.

Images