A method for sea surface target recognition using multi-scale permutation entropy features in time domain

Abstract

This invention relates to a method for identifying sea surface targets using time-domain multi-scale permutation entropy features, belonging to the field of radar signal processing technology. The method comprises the following steps: 1) caching target echo data; 2) extracting target multi-scale permutation entropy features; 3) feature optimization; and 4) identifying targets using multi-scale permutation entropy features. The invention aims to identify sea surface targets when traditional target features are ineffective.

Description

Technical Field

[0001] The present invention relates to a sea surface target recognition method using time domain multi-scale permutation entropy features, and belongs to the technical field of radar signal processing. Background Art

[0002] The ocean is vast and its environment is highly variable. Due to differences in longitude and latitude, as well as distance from land, temperature, salinity, and meteorological conditions create diverse marine environments. Surface targets are numerous and diverse in characteristics, including primary ship targets such as merchant ships, civilian vessels, and warships, as well as offshore platforms, aquaculture farms, navigation buoys, and ice floes. Recognizing surface targets in complex marine environments is a pressing need for navigation, production, and law enforcement. High-resolution range profiles (HRRPs) are currently the primary feature source for radar target identification, focusing primarily on target energy, size, and shape. Due to the complex material composition of different targets and the diverse nature of their presentation, target features fluctuate under different observation angles, natural environments, and radar operating conditions. This makes it difficult to distinguish between targets with highly similar features, leading to frequent radar misidentification and limited adaptability to complex and changing marine environments. Summary of the Invention

[0003] The purpose of the present invention is to solve the shortcomings of the above-mentioned prior art and propose a sea surface target recognition method using time domain multi-scale permutation entropy features, aiming to identify sea surface targets when traditional target features are difficult to effectively perform target recognition.

[0004] The present invention provides a method for identifying sea surface targets using time-domain multi-scale permutation entropy features, which is characterized in that it includes the following steps:

[0005] Step 1: Target echo data buffer

[0006] Target detection is completed using the echo data after pulse compression processing. The target time domain echo sequence is intercepted and cached according to the target range unit information.

[0007] Step 2: Target multi-scale permutation entropy feature extraction

[0008] The target time domain echo sequence is segmented and the permutation entropy value of each data segment is calculated. The data segments are coarse-grained at different time scales to obtain the permutation entropy value of the corresponding scale. The obtained entropy value information is integrated to obtain the multi-scale permutation entropy feature vector of each data segment to form the target feature matrix.

[0009] Step 3: Feature Optimization

[0010] For various parameters that appear when calculating multi-scale permutation entropy, including the embedding dimension m of the phase space reconstruction matrix, the time coarse-graining scale s, and the number of pulses N per data segment, the appropriate parameter range is determined based on the degree of influence of different parameter settings on the separability of the two types of targets, and feature optimization is achieved to improve recognition performance;

[0011] Step 4: Target recognition using multi-scale permutation entropy features

[0012] Using the classification algorithm, two types of targets were identified and the recognition performance was tested. The support vector machine method was used to divide the feature matrix composed of multiple targets into training set and test set, and the model hyperparameters were optimized, including the penalty factor c and the RBF kernel function parameter gamma, to obtain better recognition effect. Finally, the feature vector of the test set was input into the model for testing to obtain the target recognition result.

[0013] Preferably, the specific steps of step 1 are as follows:

[0014] The radar receives echo data in tracking mode. After pulse compression, clutter suppression, target detection and tracking processing, a total of J targets are detected in the current scene. The range unit where the target's strongest scattering point is located is selected and its pulse echo data sequence is cached. For the jth target, it is expressed as {z j (1),z j (2),…,z j (L)}, where L represents the coherent pulse train length.

[0015] Preferably, the specific steps of step 2 are:

[0016] For a target echo sequence z detected in step 1 j (n), and segment it. The number of pulses in each segment is N. After segmentation, I subsequences can be obtained, which are expressed as {x1(n),x2(n),...,x I (n)1≤n≤N}, where

[0017] For the i-th data segment x extracted above i (n), in order to obtain more feature information at different time scales, it is first coarsened, and the process is as follows:

[0018]

[0019] in, is the echo sequence after coarse-graining, s is the coarse-graining scale, β is a positive integer, and after coarse-graining, the sequence length is Q, where

[0020] For coherent radar, the target pulse echo data sequence is complex. In order to compare the size in the subsequent permutation entropy calculation, the absolute value of each data point should be taken, that is, let y s (β) = x s (β), define the phase space reconstruction matrix Y as:

[0021]

[0022] Where m is the embedding dimension, a is the number of delayed pulses, K = Q-(m-1), the dimension of the phase space reconstruction matrix Y is K×m, and the radar pulse repetition interval is expressed as T r , then the corresponding delay time t is expressed as asT r , arrange the elements of each row in the matrix Y in ascending order, and obtain a set of symbol sequences consisting of the indices of the positions of each element in the row vector, denoted as S(l). There are k kinds of arrangement patterns, namely:

[0023] S(l)={γ1,γ2,…,γ m}, l=1,2,…,k, and k≤m! (3)

[0024] There are m! different symbol sequences mapped in the m-dimensional phase space, where ! represents the factorial operator. The frequency of each permutation order in the reconstructed sequence is counted, and the probability sequence {P1, P2, ..., P k}, then the multi-scale permutation entropy H s The calculation formula is:

[0025]

[0026] Normalize the multi-scale permutation entropy, that is:

[0027]

[0028] After normalization, its value range is [0,1];

[0029] According to the above process, we get the i-th data segment x i (n) Permutation entropy H at scale s s , select different coarse-grained scales s, that is, let s=[s1,s2,…,s λ ], and obtain the permutation entropy values ​​at different coarse-grained scales, which form a feature vector with a dimension of λ;

[0030] Process each data segment of the current target echo sequence in turn to obtain its multi-scale permutation entropy feature matrix Φ, which is expressed as:

[0031]

[0032] The number of rows of the matrix Φ is related to the number of data segments I, and the number of columns is related to the number of coarse-grained scales λ;

[0033] For all target echo sequences detected in step 1), the above feature extraction method is used to extract I eigenvectors corresponding to each target, and these eigenvectors are integrated to obtain the feature matrix of the target.

[0034] Preferably, the specific steps of step 3 are:

[0035] The Bhattacharyya distance (B-distance), which can measure the similarity between two types of samples, is used as an indicator to quantitatively analyze the separability between the ship target sample space S1 and the floating target sample space S2 in the feature space. The appropriate parameter range is selected through analysis and optimization. The Bhattacharyya distance estimation method is as follows:

[0036] Calculate the mean vector μ and covariance matrix C of the two sample spaces:

[0037] μ1=mean(S1),C1=cov(S1) (7)

[0038] μ2=mean(S2),C2=cov(S2) (8)

[0039] In the formula, mean and cov represent the mean and covariance operators respectively, and Bhattacharyya distance d B The estimated value of is obtained by the following formula:

[0040]

[0041] Where, C = (C1 + C2) / 2, d μ =(μ1-μ2) / chol(C), where det represents the determinant operator and chol represents the Cholesky decomposition operator.

[0042] The beneficial effects of the present invention are as follows:

[0043] (1) The method proposed in the present invention uses the multi-scale permutation entropy of the target as a feature, obtains more target feature information through multiple time scales, and can study the mutation and randomness characteristics of the echo signal, solving the problem of relying on traditional target features but being difficult to achieve identification.

[0044] (2) Taking the Bhattacharyya distance as a quantitative indicator to measure the similarity between the two types of target samples, the feature extraction parameters were optimized to improve the feature separability of the two types of samples.

[0045] (3) Under the existing data conditions, the method proposed in this patent has an average accuracy of 93.49% in identifying ship targets and floating targets within a 0.5s observation time, which is 39.17% higher than the existing methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 : is the comparison diagram of Bhattacharyya distance between two types of target samples under different embedding dimensions m;

[0047] Figure 2 : is the comparison diagram of Bhattacharyya distance between two types of target samples under different coarse-grained scales s;

[0048] Figure 3 : The comparison diagram of Bhattacharyya distance between two types of target samples under different pulse numbers N;

[0049] Figure 4 : Scatter envelope diagram of two-dimensional multi-scale permutation entropy features of two types of targets before and after feature optimization. DETAILED DESCRIPTION

[0050] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.

[0051] A method for recognizing sea surface targets using time domain multi-scale permutation entropy features is described in the attached Figure 1-4 , including the following steps:

[0052] 1) Target echo data cache

[0053] The radar receives echo data in tracking mode. After pulse compression, clutter suppression, target detection and tracking processing, a total of J targets are detected in the current scene. The range unit where the target's strongest scattering point is located is selected and its pulse echo data sequence is cached. For the jth target, it is expressed as {z j (1),z j (2),...,z j (L)}, L represents the coherent pulse train length;

[0054] 2) Target multi-scale permutation entropy feature extraction

[0055] For a target echo sequence z detected in step 1) j (n), and segment it. The number of pulses in each segment is N. After segmentation, I subsequences can be obtained, which are expressed as {x1(n),x2(n),...,x I (n)|1≤n≤N}, where

[0056] For the i-th data segment x extracted above i(n), in order to obtain more feature information at different time scales, it is first coarsened, and the process is as follows:

[0057]

[0058] in, is the echo sequence after coarse-graining, s is the coarse-graining scale, β is a positive integer, and after coarse-graining, the sequence length is Q, where

[0059] For coherent radar, the target pulse echo data sequence is complex. In order to compare the size in the subsequent permutation entropy calculation, the absolute value of each data point should be taken, that is, let y s (β) = x s (β). Define the phase space reconstruction matrix Y as:

[0060]

[0061] Where m is the embedding dimension, a is the number of delayed pulses, K = Q-(m-1), and the dimension of the phase space reconstruction matrix Y is K×m. The radar pulse repetition interval is expressed as T r , then the corresponding delay time t is expressed as asT r By arranging the elements of each row in the matrix Y in ascending order, we can obtain a set of symbol sequences consisting of the indexes of the positions of each element in the row vector, expressed as S(l), that is,

[0062] S(l)={γ1,γ2,…,γ m}, l=1,2,…,k, and k≤m! (3)

[0063] There are m! different symbol sequences mapped in the m-dimensional phase space, where ! represents the factorial operator. By counting the frequency of each permutation order in the reconstructed sequence, we can get the probability sequence {P1, P2, ..., P k}, then the multi-scale permutation entropy H s The calculation formula is:

[0064]

[0065] Normalize the multi-scale permutation entropy, that is:

[0066]

[0067] After normalization, its value range is [0,1].

[0068] According to the above process, the i-th data segment x can be obtained i (n) Permutation entropy H at scale s s, select different coarse-grained scales s, that is, let s=[s1,s2,…,s λ ], we can get the permutation entropy values ​​at different coarse-grained scales, and these values ​​form a feature vector with dimension λ. For example, when λ = 3, for the i-th data segment, we can get a three-dimensional feature vector, expressed as

[0069] Process each data segment of the current target echo sequence in turn to obtain its multi-scale permutation entropy feature matrix Φ, which is expressed as:

[0070]

[0071] The number of rows of the matrix Φ is related to the number of data segments I, and the number of columns is related to the number of coarse-grained scales λ.

[0072] For all target echo sequences detected in step 1), the above feature extraction method is used to extract I eigenvectors corresponding to each target. These eigenvectors are then integrated to form the target feature matrix. For example, if two targets are detected in the echo data, the number of data segments obtained after truncation of each target echo data is A, and the number of coarse-grained scales is λ = B, then the multi-scale permutation entropy feature matrices Φ1 and Φ2 of the targets can be obtained, both with A × B dimensions.

[0073] 3) Feature Optimization

[0074] The feature extraction process in step 2) involves many parameters, including the embedding dimension m of the phase space reconstruction matrix Y, the number of delayed pulses a, the coarse-graining scale s, and the number of pulses per data segment N. The number of delayed pulses a is typically set to 1 to ensure the continuity of the target information features in the time domain. The other three parameters should be optimized based on their impact on the separability of different targets to achieve feature optimization.

[0075] The Bhattacharyya distance (B-distance), which can measure the similarity between two types of samples, is used as an indicator to quantitatively analyze the separability between the ship target sample space S1 and the floating target sample space S2 in the feature space. The appropriate parameter range is selected through analysis and optimization. The Bhattacharyya distance estimation method is as follows:

[0076] Calculate the mean vector μ and covariance matrix C of the two sample spaces

[0077] μ1=mean(S1),C1=cov(S1) (7)

[0078] μ2=mean(S2),C2=cov(S2) (8)

[0079] In the formula, mean and cov represent the mean and covariance operators respectively, and Bhattacharyya distance d B The estimated value of is obtained by the following formula:

[0080]

[0081] Where, C = (C1 + C2) / 2, d μ =(μ1-μ2) / chol(C). Here det represents the determinant operator and chol represents the Cholesky decomposition operator;

[0082] The following uses measured data to study the effects of the three parameters on the separability of the two types of targets in order to determine the appropriate parameter range.

[0083] (1) The embedding dimension m of the phase space reconstruction matrix Y

[0084] m determines the number of columns in the phase space reconstruction matrix Y. The larger the value of m, the more symbol sequences are formed by the position index of each element when the elements in each row of the matrix are arranged in ascending order, and the higher the accuracy of sorting the elements. The Bhattacharyya distance comparison of the two types of target samples obtained from the measured data under different embedding dimensions m is as follows: Figure 1 Obviously, when the embedding dimension m is 2, the separability of the features is poor. When m is 3, the separability of the features is generally good. When m is 4, 5, or 6, the separability of the features is close, but slightly worse than when m is 3.

[0085] (2) Coarse-grained scale s

[0086] s determines the length of the time series after coarse-graining. Different s values ​​represent different time scales from which to observe the target features. The Bhattacharyya distance comparison of the two types of target samples obtained from the measured data at different coarse-graining scales s is as follows: Figure 2 As shown in the figure, when the coarse-graining scale s is 2 or 3, the difference in sea surface target features is better. However, as the value of s increases, the feature separability deteriorates. This is because averaging too many data points obscures the target's characteristic information, making it difficult to achieve good differentiation. During the feature extraction process, the multi-scale permutation entropy is calculated for each data segment with coarse-graining scales of 2 and 3, and these two values ​​are used to form the feature vector of each data segment.

[0087] (3) Number of pulses per segment of data N

[0088] N determines the amount of data required for an entropy calculation when processing the echo signal. The Bhattacharyya distance comparison of the two types of target samples obtained from the measured data under different pulse numbers N is as follows: Figure 3Obviously, the more pulses there are, the better the separability of the multi-scale permutation entropy features of the two types of targets. However, it is necessary to consider the needs of target recognition in actual radar systems and make a compromise between recognition accuracy and real-time performance to achieve better performance.

[0089] After feature optimization, the appropriate parameter range is selected to increase the Bhattacharyya distance between the two types of samples and improve the feature separability. The comparison of the extracted two-dimensional features before and after optimization is as follows: Figure 4 As shown, Figure 4 (a) Select any two sets of feature parameters to form a two-dimensional feature, and the feature envelope areas of the two types of targets overlap. Figure 4 In (b), two sets of optimized feature parameters are selected, and the overlap between the two types of target features is significantly reduced. This shows that feature optimization is very helpful for subsequent target recognition.

[0090] 4) Target recognition using multi-scale permutation entropy features

[0091] In step 3), we obtained the optimized feature vector sets for all targets. We then used classification algorithms from machine learning and deep learning to identify the two types of targets and tested their performance. These algorithms included decision trees, support vector machines, naive Bayes, convex hull algorithms, and convolutional neural networks.

[0092] Three sets of measured data were used for feature extraction, and their basic information is shown in Table 1. The echo data were segmented, with each segment consisting of 1024 pulses, corresponding to a duration of approximately 0.5 seconds. A total of 128 feature vectors were extracted for each target, and a total of 256 feature vectors for both target types were obtained from each data set, forming a feature matrix.

[0093] Table 1 Introduction to measured data

[0094]

[0095] During recognition, first, each feature vector is assigned a label value, for example, a ship target is 1 and a floating target is 2, and the feature matrix is ​​divided into a training set and a test set. The recognition process is explained using the support vector machine method as an example. The training set is used to build an SVM classifier model, and the hyperparameters in the SVM are optimized using a cyclic iteration method, including the penalty factor c and the RBF kernel function parameter gamma. The range of both parameters is from the minimum value of 2 -8 Start, maximum value 2 8 At the end, the loop step is increased exponentially by one, and a five-fold cross-validation is performed on the model under each set of parameters. The parameters with the best recognition effect are selected to obtain the classifier model. Then the test set data is input into the model for recognition. The following four situations will appear in the test results:

[0096] ①True Positive (TP): The model correctly predicts a sample that is actually a ship target as a ship target;

[0097] ②False Negative (FN): The model mistakenly predicts a sample that is actually a ship target as a floating target;

[0098] ③False Positive (FP): The model mistakenly predicts a floating target as a ship target;

[0099] ④True Negative (TN): The model correctly predicts a sample that is actually a floating target as a floating target.

[0100] Table 2 Statistics of target recognition results of time domain multi-scale permutation entropy features

[0101]

[0102] Table 3 Statistics of target recognition results based on time domain waveform entropy features

[0103]

[0104] The recognition results are summarized in Table 2. For comparison purposes, the recognition results using time domain waveform entropy as a feature are also shown in Table 3. It can be seen that the average recognition accuracy achieved by existing methods is 54.32%, while the average accuracy of the proposed method reaches 93.49%. Under the existing data conditions, the recognition accuracy is improved by 39.17% compared to existing methods.

Claims

1. A method for sea surface target recognition using time domain multi-scale permutation entropy features, characterized in that The following steps are involved: Step 1: Target echo data buffer Target detection is completed using the echo data after pulse compression processing. The target time domain echo sequence is intercepted and cached according to the target range unit information. Step 2: Target multi-scale permutation entropy feature extraction The target time domain echo sequence is segmented and the permutation entropy value of each data segment is calculated. The data segments are coarse-grained at different time scales to obtain the permutation entropy value of the corresponding scale. The obtained entropy value information is integrated to obtain the multi-scale permutation entropy feature vector of each data segment to form the target feature matrix. Step 3: Feature Optimization For various parameters that appear when calculating multi-scale permutation entropy, including the embedding dimension m of the phase space reconstruction matrix, the time coarse-graining scale s, and the number of pulses N per data segment, the appropriate parameter range is determined based on the degree of influence of different parameter settings on the separability of the two types of targets, and feature optimization is achieved to improve recognition performance; Step 4: Target recognition using multi-scale permutation entropy features Using the classification algorithm, two types of targets were identified and the recognition performance was tested. The support vector machine method was used to divide the feature matrix composed of multiple targets into training set and test set, and the model hyperparameters were optimized, including the penalty factor c and the RBF kernel function parameter gamma, to obtain better recognition effect. Finally, the feature vector of the test set was input into the model for testing to obtain the target recognition result.

2. A method for sea surface target recognition using time domain multi-scale permutation entropy features according to claim 1, characterized in that The specific steps of step 1 are as follows: The radar receives echo data in tracking mode. After pulse compression, clutter suppression, target detection and tracking processing, a total of J targets are detected in the current scene. The range unit where the target's strongest scattering point is located is selected and its pulse echo data sequence is cached. For the jth target, it is expressed as {z j (1),z j (2),…,z j (L)}, where L represents the coherent pulse train length.

3. A method for identifying sea surface targets using time-domain multi-scale permutation entropy features according to claim 2, characterized in that The specific steps of step 2 are: For a target echo sequence z detected in step 1 j (n), and segment it. The number of pulses in each segment is N. After segmentation, I subsequences can be obtained, which are expressed as {x1(n),x2(n),...,x I (n)|1≤n≤N}, where For the i-th data segment x extracted above i (n), in order to obtain more feature information at different time scales, it is first coarsened, and the process is as follows: in, is the echo sequence after coarse-graining, s is the coarse-graining scale, β is a positive integer, and after coarse-graining, the sequence length is Q, where For coherent radar, the target pulse echo data sequence is complex. In order to compare the size in the subsequent permutation entropy calculation, the absolute value of each data point should be taken, that is, let y s (β)=|x s (β)|, define the phase space reconstruction matrix Y as: Where m is the embedding dimension, a is the number of delayed pulses, K = Q-(m-1), the dimension of the phase space reconstruction matrix Y is K×m, and the radar pulse repetition interval is expressed as T r , then the corresponding delay time t is expressed as asT r , arrange the elements of each row in the matrix Y in ascending order, and obtain a set of symbol sequences consisting of the indices of the positions of each element in the row vector, denoted as S(l). There are k kinds of arrangement patterns, namely: S(l) = {γ1, γ2, …, γ m}, l = 1, 2, …, k, and k ≤ m! (3) There are m! different symbol sequences mapped in the m-dimensional phase space, where ! represents the factorial operator. The frequency of each permutation order in the reconstructed sequence is counted, and the probability sequence {P1, P2, ..., P k }, then the multi-scale permutation entropy H s The calculation formula is: Normalize the multi-scale permutation entropy, that is: After normalization, its value range is [0,1]; According to the above process, we get the i-th data segment x i (n) Permutation entropy H at scale s s , select different coarse-grained scales s, that is, let s=[s1,s2,…,s λ ], and obtain the permutation entropy values ​​at different coarse-grained scales, which form a feature vector with a dimension of λ; Process each data segment of the current target echo sequence in turn to obtain its multi-scale permutation entropy feature matrix Φ, which is expressed as: The number of rows of the matrix Φ is related to the number of data segments I, and the number of columns is related to the number of coarse-grained scales λ; For all target echo sequences detected in step 1), the above feature extraction method is used to extract I eigenvectors corresponding to each target, and these eigenvectors are integrated to obtain the feature matrix of the target.

4. A method for sea surface target recognition using time domain multi-scale permutation entropy features according to claim 1, characterized in that The specific steps of step 3 are: The Bhattacharyya distance, which can measure the similarity of two types of samples, is used as an indicator to quantitatively analyze the separability between the ship target sample space S1 and the floating target sample space S2 in the feature space. The appropriate parameter range is selected through analysis and optimization. The Bhattacharyya distance estimation method is as follows: Calculate the mean vector μ and covariance matrix C of the two sample spaces: μ1=mean(S1),C1=cov(S1) (7) μ2=mean(S2),C2=cov(S2) (8) In the formula, mean and cov represent the mean and covariance operators respectively, and Bhattacharyya distance d B The estimated value of is obtained by the following formula: Where, C = (C1 + C2) / 2, d μ =(μ1-μ2) / chol(C), where det represents the determinant operator and chol represents the Cholesky decomposition operator.

Images