Small target detection method on sea surface based on multi-scale envelope spectrum entropy and random forest

Abstract

The present invention belongs to the field of radar signal processing technology, and specifically relates to a method for detecting small targets on the sea surface based on multi-scale envelope spectrum entropy and random forest, comprising: obtaining a signal to be detected, the signal to be detected includes a sea clutter signal and a target echo signal; using an improved adaptive noise set empirical mode decomposition algorithm to decompose the signal to obtain an IMF component; calculating the correlation coefficient and variance contribution rate between the decomposed IMF component and the original signal, and performing signal screening and reconstruction; calculating the multi-scale envelope spectrum entropy value of the reconstructed signal, and normalizing the entropy value; constructing a random forest classifier model, and finely controlling the classification performance during training; and judging the final prediction value according to the decision threshold during judgment. The present invention effectively improves the probability and performance of detecting small targets on the sea surface.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar signal processing, and in particular relates to a method for detecting small sea surface targets based on multi-scale envelope spectral entropy and random forest. Background Art

[0002] Sea clutter is a common natural phenomenon in ocean radar detection. It is a radar echo generated by sea surface fluctuations. Sea clutter has complex spatial and temporal variation characteristics, as well as non-stationarity under different sea conditions and radar observation conditions. Since the multi-dimensional non-Gaussian characteristics of sea clutter are particularly obvious at high resolution, and the target echo is usually weak and the signal-to-clutter ratio (SCR) is low, small target detection on the sea surface has always been a problem faced by researchers.

[0003] At present, many scholars at home and abroad have conducted extensive research on the problem that sea clutter and target echoes are difficult to distinguish effectively. Among them, the empirical mode decomposition (EMD) algorithm can decompose the signal into intrinsic mode functions (IMFs) according to the signal's own scale. However, EMD is prone to modal aliasing and endpoint effects when processing non-stationary signals, and the time scale is single. Some existing technologies combine variational mode decomposition with deep forests, which can solve the modal aliasing problem of empirical mode decomposition to a certain extent. However, for sea clutter, which contains more noise, more accurate modal separation is required and the parameters in the algorithm are extremely difficult to adjust. Therefore, the noise suppression method in sea clutter still needs to be further improved, and the existing algorithm structure needs to be optimized. Summary of the invention

[0004] The purpose of the present invention is to provide a method for detecting small targets on the sea surface based on multi-scale envelope spectral entropy and random forest, which is used to solve the technical problems in the prior art that noise suppression in sea clutter is still insufficient, parameter adjustment is difficult, the time scale is single, and modal aliasing is easy to occur.

[0005] The method for detecting small targets on the sea surface based on multi-scale envelope spectral entropy and random forest includes the following steps.

[0006] The signals to be detected are obtained, and the signals to be detected include sea clutter signals and target echo signals.

[0007] The signal is decomposed using an improved adaptive noise ensemble empirical mode decomposition algorithm to obtain IMF components.

[0008] The correlation coefficient and variance contribution rate between the IMF components obtained by decomposition and the original signal are calculated, and signal screening and reconstruction are performed based on the two indicators.

[0009] Calculate the multi-scale envelope spectrum entropy of the reconstructed signal and normalize the entropy value.

[0010] A random forest classifier model is constructed, and the multi-scale envelope spectrum entropy value is used as the true value of the corresponding sample during training. During training, the prediction effect of the random forest classifier model is evaluated by comparing the final predicted value with the true value. The false alarm rate of the random forest model is controlled by updating the decision threshold, thereby achieving fine control of the classification performance. During judgment, the final predicted value is judged according to the decision threshold.

[0011] Preferably, the method of using the improved adaptive noise set empirical mode decomposition algorithm to decompose the signal includes: adding Gaussian white noise to the original signal; calculating the The residual components and The inherent mode components are obtained by iterative calculation of signal decomposition until the decomposition of the original signal is completed.

[0012] Preferably, the method for decomposing a signal using an improved adaptive noise set empirical mode decomposition algorithm specifically includes: adding Gaussian white noise to the original signal. The expression is as follows:

[0013]

[0014] in, represents the original signal, Indicates The signal to be decomposed, represents the noise standard deviation of the signal at the first decomposition, represents the first modal component generated by empirical mode decomposition (EMD), Indicates A zero-mean unit-variance white noise is added.

[0015] No. The residual components and The calculation formula of the natural mode components is as follows:

[0016]

[0017]

[0018] in, Indicates The residual components, Indicates The residual components, Indicates The intrinsic mode components, Indicates The noise standard deviation of the signal at the decomposition time, represents the first modal components, It means taking the average value, It means to calculate the local mean of the signal.

[0019] The iterative calculation stops when the obtained residual is a monotonic function.

[0020] Preferably, when the correlation coefficient is greater than threshold one and the variance contribution rate is greater than threshold two, it indicates that the corresponding IMF component contains the main signal feature, and the signal of the IMF component is screened out for signal reconstruction.

[0021] Preferably, the method for calculating the multi-scale envelope spectrum entropy value of the reconstructed signal includes: performing Hilbert transform on the screened IMF component signal, obtaining the envelope spectrum using fast Fourier transform, and segmenting the envelope spectrum sequence to obtain a new sequence; calculating the correlation between different vector elements in the new sequence; taking the sample entropy set at multiple time scales as the multi-scale envelope spectrum entropy, and then normalizing the calculated multi-scale envelope spectrum entropy.

[0022] Preferably, the method for calculating the multi-scale envelope spectrum entropy value of the reconstructed signal specifically includes: the envelope spectrum sequence is , the new sequence is: ,in, Indicates that the envelope spectrum sequence contains data points, represents the time scale, Express The value obtained by rounding down is to divide the original envelope spectrum sequence into part.

[0023] The calculation formula for the correlation between different vector elements is as follows:

[0024]

[0025] in, Express satisfaction The element-wise probability of the condition, represents the distance between all different vector elements, is the similarity coefficient, is the embedding dimension.

[0026] Multi-scale envelope spectral entropy, the formula is as follows:

[0027]

[0028] in, represents the multi-scale envelope spectral entropy, Indicates that based on parameters Calculate the sample entropy of .

[0029] Preferably, during training, after obtaining the signal to be detected, the signal to be detected is normalized and preprocessed, and the sea clutter signal and the target echo signal are labeled with "0" and "1" respectively.

[0030] Preferably, during training, after obtaining the signal to be detected, the detection problem is classified as a binary hypothesis test, and the expression is as follows:

[0031]

[0032] in, represents the sea clutter signal, Indicates the target echo signal, Assume that there is only sea clutter signal in the echo signal. It is assumed that the echo signal contains the target echo.

[0033] Preferably, the expression of the random forest model is:

[0034]

[0035] in, represents the number of decision trees in the forest, For the Tree pair samples The predicted value of For sample The true value of For sample During training, a target false alarm rate is set and the decision threshold closest to the target false alarm rate is adjusted.

[0036] Preferably, find the decision threshold closest to the target false alarm rate The expression is:

[0037]

[0038] in, Represents the false positive rate, which is used to measure the error rate. is the target false alarm rate.

[0039] When judging, if , it is judged that it does not contain the target echo, belonging to Assume; if , it is judged to contain the target echo, belonging to Assumption.

[0040] The present invention has the following advantages: the present invention uses an improved adaptive noise ensemble empirical mode decomposition (ICEEMDAN) algorithm to decompose intrinsic mode components, and selects components with high correlation for signal reconstruction. The IMF components after denoising and reconstruction are subjected to Hilbert transform, the time domain signal is mapped to the frequency domain, and the multi-scale envelope spectrum entropy of the signal is calculated, which can increase the difference between the sea clutter and the target echo characteristics. Then the features are normalized and input into a random forest classifier, which effectively improves the detection probability and enhances the detection performance. By comparing the predicted value with the decision threshold, the decision threshold is updated in real time to achieve the purpose of controlling the false alarm rate to be constant. Experimental results show that the detection probability of this method is better than several common detection methods at present as a whole. The detection rate is significantly improved under polarization mode. The present invention solves the problems of difficult design of target detection methods for feature extraction and unstable models, and effectively improves the detection probability and detection performance of small targets on the sea surface. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 The figure is a flow chart of the method for detecting small targets on the sea surface based on multi-scale envelope spectral entropy and random forest in the present invention.

[0042] Figure 2 This is the signal diagram obtained after decomposition using the ICEEMDAN algorithm in the present invention.

[0043] Figure 3 It is a comparison diagram of detection probabilities of the present invention under four polarizations: HH, HV, VH, and VV.

[0044] Figure 4 The following are the detection performance result diagrams of five detection methods of the present invention and other prior arts under four polarizations. DETAILED DESCRIPTION

[0045] The specific implementation modes of the present invention are further explained in detail below by describing the embodiments with reference to the accompanying drawings, so as to help those skilled in the art to have a more complete, accurate and in-depth understanding of the inventive concept and technical solution of the present invention.

[0046] like Figure 1 , Figure 2 As shown, the present invention provides a method for detecting small targets on the sea surface based on multi-scale envelope spectrum entropy and random forest, comprising the following steps:

[0047] S1: Acquire the signal to be detected, which includes the sea clutter signal and the target echo signal.

[0048] This step also performs normalization preprocessing on the signal to be detected, and labels the sea clutter signal and the target echo signal with "0" and "1" respectively.

[0049] This step reduces the detection problem to a binary hypothesis test, expressed as follows:

[0050]

[0051] in, represents the sea clutter signal, Indicates the target echo signal, Assume that there is only sea clutter signal in the echo signal. It is assumed that the echo signal contains the target echo.

[0052] S2: Decompose the signal using the improved adaptive noise ensemble empirical mode decomposition (ICEEMDAN) algorithm to obtain IMF (intrinsic modal component) components. This step specifically includes the following sub-steps:

[0053] S21: Add Gaussian white noise to the original signal. The expression is as follows:

[0054]

[0055] in, represents the original signal, Indicates The signal to be decomposed, represents the noise standard deviation of the signal at the first decomposition, represents the first modal component generated by empirical mode decomposition (EMD), Indicates A zero-mean unit-variance white noise is added.

[0056] S22: The residual components and The natural mode components are calculated as follows:

[0057]

[0058]

[0059] in, Indicates The residual components, Indicates The residual components, Indicates The intrinsic mode components, Indicates The noise standard deviation of the signal at the decomposition time, represents the first modal components, It means taking the average value, It means to calculate the local mean of the signal.

[0060] S23: Perform iterative calculation of signal decomposition until the decomposition of the original signal is completed, that is, stop when the obtained residual is a monotonic function.

[0061] S3: Calculate the correlation coefficient and variance contribution rate (VCR) between the IMF components obtained by decomposition and the original signal, and perform signal screening and reconstruction based on the two indicators.

[0062] After the IMF components are calculated in the previous step, the correlation coefficient and variance contribution rate between the IMF components and the original signal are calculated. When the correlation coefficient is greater than threshold 1 (0.1) and the variance contribution rate is greater than threshold 2 (0.01), it means that the corresponding IMF component contains the main signal features, which can be used to represent the characteristics of the original sea clutter signal. Therefore, the signal of the IMF component is screened out for signal reconstruction. In this embodiment, the correlation coefficient of components IMF1-IMF4 is greater than 0.1 and the variance contribution rate is greater than 0.01, so the first 4 IMF components are retained for reconstruction to obtain a new sequence.

[0063] S4: Calculate the multi-scale envelope spectrum entropy value of the reconstructed signal and normalize the entropy value. This step includes the following sub-steps.

[0064] S41: Perform Hilbert Transform on the preprocessed IMF component signal, use Fast Fourier Transform (FFT) to obtain the envelope spectrum, and convert the envelope spectrum sequence into , the new sequence is: ,in, Indicates that the envelope spectrum sequence contains data points, represents the time scale, Express The value obtained by rounding down is to divide the original envelope spectrum sequence into part.

[0065] S42: Calculate the correlation between different vector elements:

[0066]

[0067] in, Express satisfaction The element-wise probability of the condition, represents the distance between all different vector elements, is the similarity coefficient, is the embedding dimension.

[0068] S43: Defining multiple scales The sample entropy set under is the multi-scale envelope spectral entropy:

[0069]

[0070] in, represents the multi-scale envelope spectral entropy, Indicates that based on parameters Calculate the sample entropy of .

[0071] S5: Build a random forest classifier model and compare the predicted values Update the decision threshold in real time with the size of the decision threshold , the extracted features are input into the classifier, and by comparing with the decision threshold, it is determined whether there is a target in the echo signal.

[0072] In step S5, a random forest classifier model is constructed to determine whether there is a target in the echo signal. This step includes the following sub-steps:

[0073] S51: Construct a random forest classifier model. The expression of the random forest model is:

[0074]

[0075] in, represents the number of decision trees in the forest, For the Tree pair samples The predicted value of For sample The true value of For sample The final predicted value of the random forest model can be evaluated by comparing the final predicted value with the true value.

[0076] S52: Set a target false alarm rate , and by finding the decision threshold closest to the target false alarm rate To adjust and find the decision threshold closest to the target false alarm rate The expression is:

[0077]

[0078] in, Represents the false positive rate, which is used to measure the error rate.

[0079] S53: During training, the false alarm rate of the random forest model is controlled by updating the decision threshold, thereby achieving fine control of the classification performance. During judgment, the final prediction value is judged according to the decision threshold. , it is judged that it does not contain the target echo, belonging to Assume; if , it is judged to contain the target echo, belonging to Assumption.

[0080] To verify the effect of this method, the embodiment is verified by experiments, and the data used is from the IPIX radar target database collected on the east coast of Canada in 1993. The radar operates in the x-band, the operating frequency is 9.39 GHz, and the resolution is 30 m. The target to be detected is a small ball wrapped in a wire mesh with a diameter of about 1 m. The data contains 4 polarization modes, namely HH, HV, VH, and VV. Each polarization mode consists of 14 distance units, and each distance unit contains 217 (131.072s) pulses. Figure 3-4 As shown, Figure 3 This is a comparison chart of the detection probability of the MSESEn-RFCFA algorithm under four polarizations. ,It can be seen that the detection performance under HH, HV, VH, and VV polarization modes is good. Figure 4 The following are the detection performance results of the five detection methods under four polarizations. It can be seen from the figure that under HH and VV polarizations, this method (abbreviated as MSESEn-RFCFA) is significantly better than other detection methods (Fourier change, Hurst index, three-feature detector, average energy ratio), and the detection rate of this method under HV polarization is increased by an average of 18%.

[0081] The present invention is described above by way of example in conjunction with the accompanying drawings. It is obvious that the specific implementation of the present invention is not limited to the above-mentioned method. As long as various non-substantial improvements are made using the inventive concept and technical solution of the present invention, or the inventive concept and technical solution are directly applied to other occasions without improvement, they are all within the protection scope of the present invention.

Claims

1. A method for detecting small targets on the sea surface based on multi-scale envelope spectrum entropy and random forest, characterized by: The following steps are involved: Acquire the signal to be detected, the signal to be detected includes sea clutter signal and target echo signal; The signal is decomposed using the improved adaptive noise ensemble empirical mode decomposition algorithm to obtain the IMF components; The correlation coefficient and variance contribution rate between the IMF components obtained by decomposition and the original signal are calculated, and the signal is screened and reconstructed based on the two indicators; Calculate the multi-scale envelope spectrum entropy value of the reconstructed signal and normalize the entropy value; Construct a random forest classifier model, use the multi-scale envelope spectrum entropy value as the true value of the corresponding sample during training, evaluate the prediction effect of the random forest classifier model by comparing the final prediction value with the true value during training, and control the false alarm rate of the random forest model by updating the decision threshold, so as to achieve fine control of the classification performance; during judgment, judge the final prediction value according to the decision threshold; The expression of the random forest model is: in, represents the number of decision trees in the forest, For the Tree pair samples The predicted value of For sample The true value of For sample The final predicted value; during training, a target false alarm rate is set, and the decision threshold closest to the target false alarm rate is adjusted; Find the decision threshold closest to the target false alarm rate The expression is: in, Represents the false positive rate, which is used to measure the error rate. is the target false alarm rate; When judging, if , it is judged that it does not contain the target echo, belonging to Assume; if , it is judged to contain the target echo, belonging to Assumptions; During training, after obtaining the signal to be detected, the detection problem is classified as a binary hypothesis test, and the expression is as follows: in, represents the sea clutter signal, Indicates the target echo signal, Assume that there is only sea clutter signal in the echo signal. It is assumed that the echo signal contains the target echo.

2. The method for detecting small sea targets based on multi-scale envelope spectral entropy and random forest according to claim 1, characterized in that: The method of decomposing a signal using an improved adaptive noise ensemble empirical mode decomposition algorithm includes: adding Gaussian white noise to the original signal; calculating the The residual components and The inherent mode components are obtained by iterative calculation of signal decomposition until the decomposition of the original signal is completed.

3. The method for detecting small sea targets based on multi-scale envelope spectral entropy and random forest according to claim 2 is characterized in that: The method of using the improved adaptive noise set empirical mode decomposition algorithm to decompose the signal specifically includes: adding Gaussian white noise to the original signal. The expression is as follows: in, represents the original signal, Indicates The signal to be decomposed, represents the noise standard deviation of the signal at the first decomposition, represents the first modal component generated by empirical mode decomposition, Indicates A zero-mean unit-variance white noise is added; No. The residual components and The calculation formula of the natural mode components is as follows: in, Indicates The residual components, Indicates The residual components, Indicates The intrinsic mode components, Indicates The noise standard deviation of the signal at the decomposition time, represents the first modal components, It means taking the average value, Indicates the calculation of the local mean of the signal; The iterative calculation stops when the obtained residual is a monotonic function.

4. The method for detecting small sea targets based on multi-scale envelope spectral entropy and random forest according to claim 1, characterized in that: When the correlation coefficient is greater than threshold one and the variance contribution rate is greater than threshold two, it means that the corresponding IMF component contains the main signal features, and the signal of the IMF component is screened out for signal reconstruction.

5. The method for detecting small sea targets based on multi-scale envelope spectral entropy and random forest according to claim 1, characterized in that: The method for calculating the multi-scale envelope spectrum entropy value of the reconstructed signal includes: performing Hilbert transform on the screened IMF component signal, obtaining the envelope spectrum by using fast Fourier transform, segmenting the envelope spectrum sequence to obtain a new sequence; calculating the correlation between different vector elements in the new sequence; taking the sample entropy set at multiple time scales as the multi-scale envelope spectrum entropy, and then normalizing the calculated multi-scale envelope spectrum entropy.

6. The method for detecting small sea targets based on multi-scale envelope spectral entropy and random forest according to claim 5, characterized in that: The method for calculating the multi-scale envelope spectrum entropy value of the reconstructed signal specifically includes: the envelope spectrum sequence is , i =1,2,…, N , the new sequence is: , j =1,2,…, T ,in, Indicates that the envelope spectrum sequence contains data points, represents the time scale, Express The value obtained by rounding down is to divide the original envelope spectrum sequence into part; The calculation formula for the correlation between different vector elements is as follows: in, Express satisfaction The element-wise probability of the condition, represents the distance between all different vector elements, is the similarity coefficient, is the embedding dimension; Multi-scale envelope spectral entropy, the formula is as follows: in, represents the multi-scale envelope spectral entropy, Indicates that based on parameters The sample entropy calculation of .

7. The method for detecting small sea targets based on multi-scale envelope spectral entropy and random forest according to claim 1, characterized in that: During training, after obtaining the signal to be detected, the signal to be detected is normalized and preprocessed, and the sea clutter signal and the target echo signal are labeled "0" and "1" respectively.

Images