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

Abstract

The application relates to a sea surface target recognition method using time-frequency domain multi-scale permutation entropy features, and belongs to the technical field of radar signal processing and feature recognition. The steps comprise the following steps. Step 1: target echo data buffering; step 2: time-frequency domain analysis of target echo data; step 3: target time-frequency domain multi-scale permutation entropy feature extraction; and step 4: target recognition using the time-frequency domain multi-scale permutation entropy features. The application aims to recognize sea surface targets in the case that traditional target features cannot effectively recognize targets.

Description

Technical Field

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

[0002] Rapid and accurate identification of surface targets helps determine their function and purpose, which is of great significance in both military and civilian fields. Common surface targets include floating targets (small fishing boats, buoys, ice floes, etc.) and naval targets (large cruise ships, cargo ships, and ro-ro cargo ships). These targets have significantly different structures and physical sizes, so high-resolution range profile features are currently the primary method used. However, due to the complex and changing nature of the surface, varying observation angles, and varying radar operating conditions, target features are unstable. This makes it difficult to distinguish between targets with high feature similarity, making floating target identification particularly challenging and severely impacting recognition performance.

[0003] High-resolution radars currently primarily use one-dimensional high-resolution range profiles for identification. This feature is closely related to the observation angle. When the radar beam shines along the ship's side, the two types of targets are extremely similar, severely impacting recognition performance. Low-resolution radars, on the other hand, lack the ability to obtain HRRP features, making target recognition even more difficult without relying on auxiliary information. Therefore, to overcome the bottleneck in target recognition performance, mining and utilizing more effective feature information is an effective solution. Summary of the Invention

[0004] 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 based on multi-scale permutation entropy features in the time-frequency domain, aiming to identify sea surface targets when traditional target features are difficult to effectively perform target recognition.

[0005] A method for identifying sea surface targets using multi-scale permutation entropy features in the time-frequency domain of the present invention comprises the following steps:

[0006] Step 1: Target echo data buffer

[0007] 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.

[0008] Step 2: Time-frequency domain analysis of target echo data

[0009] The target time domain echo sequence is segmented by sliding window, and each data segment is subjected to time-frequency analysis;

[0010] Step 3: Target time-frequency domain multi-scale permutation entropy feature extraction

[0011] After time-frequency domain analysis, coarse-graining processing is performed at different time scales along the Doppler channel to calculate the permutation entropy value of the corresponding scale. The minimum multi-scale permutation entropy value is then selected as the feature to finally form the target feature vector.

[0012] Step 4: Target recognition using multi-scale permutation entropy features in the time-frequency domain

[0013] The support vector machine classification algorithm is used to complete the recognition of two types of targets and test the recognition performance. The feature matrix composed of multiple targets is divided into training set and test set, and the model hyperparameter optimization is carried out, 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 is input into the model for testing to obtain the target recognition result.

[0014] Preferably, the specific steps of step 1 are:

[0015] The radar receives echo data in tracking mode. After pulse compression, clutter suppression, target detection and tracking processing, a total of k 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 z is cached. k (m), for the kth target, it is expressed as {z k (1),z k (2),...,z k (L)}, where L represents the coherent pulse train length.

[0016] Preferably, the specific steps of step 2 are:

[0017] For a target time domain echo sequence z detected in step 1) k (m), select a rectangular window function with a length of M to perform sliding window interception processing, the rectangular window function R M (n) as follows:

[0018]

[0019] The window function slides forward W pulses each time it intercepts, and we get A subsequence X of length M i , which can be expressed as X i ={z k [(i-1)W+1],z k [(i-1)W+2],…,z k [(i-1)W+M]}, where

[0020] The i-th data segment X after the sliding window is intercepted iThe short-time Fourier transform method is used to analyze the time-frequency domain to obtain information in two dimensions of time domain and frequency domain, and accumulation of target energy is realized. The Fourier transform is performed on the intercepted time domain signal each time to obtain a frequency spectrum diagram of the signal in the time period. The time domain signal is intercepted on the whole time period to obtain a set of each frequency spectrum. The result is a two-dimensional function of time-frequency. For the two-dimensional matrix, the rows of the matrix represent different Doppler channels, and the columns of the matrix represent different times. The short-time Fourier transform formula is as follows

[0021]

[0022] wherein p(t-α) is a time window with α as the center, a rectangular window is used, the number of points in the segmented Fourier transform is N, the energy of the floating target swings back and forth between different Doppler channels with time, and the energy of the ship target changes little with time. The time-frequency images of the two types of targets have obvious feature differences.

[0023] Preferably, the specific steps of the step 3) target time-frequency domain multi-scale permutation entropy feature extraction are as follows:

[0024] The absolute value of each data point in the time-frequency result of the data segment X i is taken, that is, y(t,f) = |stft(t,f)|. Then, the multi-scale permutation entropy value of the time-frequency sequence is calculated along the Doppler channel. The kth Doppler channel sequence in y(t,f) is taken out and represented as y(t,k) = {y(1,k),y(2,k),…,y(M,k)}. In order to obtain more feature information on different scales, coarse-grained processing is performed on the time-frequency sequence along the Doppler channel. y i (β) = y(t,k). The process is as follows:

[0025]

[0026] wherein y is the coarse-grained sequence, s is the coarse-grained scale, β is a positive integer, and the sequence length is Q after the coarse-grained processing, wherein

[0027] The phase space reconstruction matrix Y is defined as

[0028]

[0029] wherein m is the embedding dimension, a is the delay pulse number, K = Q-(m-1), and the dimension of the phase space reconstruction matrix Y is Kxm. The ascending order arrangement is performed on each row element in the matrix Y to obtain a group of symbol sequences γ1,γ2,…,γ m , which is represented as S(l), that is,

[0030] S(l) = {γ1, γ2, …, γ m}, l = 1, 2, …, k, and k ≤ m! (6)

[0031] There are m! kinds of m-dimensional phase space mapping different symbol sequences, where! represents the factorial operator, and the frequency of each permutation order appearing in the reconstructed sequence is counted to obtain the probability sequence {P1, P2, …, P k The calculation formula of the multi-scale permutation entropy H s is:

[0032]

[0033] The multi-scale permutation entropy is normalized, that is,

[0034] After normalization, its value range is [0, 1];

[0035] According to the above process, the permutation entropy value of y(t, f) at scale s is obtained, and the multi-scale permutation entropy values E1, E2, …, E k ,…E M of all Doppler channels corresponding to the time-frequency result of the data segment are calculated, and the minimum value σ i is selected as the characteristic value of this data segment, where σ i = min{E1, E2, …, E k ,…E M}, and each data segment of the current target echo sequence is processed in turn to obtain the feature vector Φ of the time-frequency domain multi-scale permutation entropy, which is represented as:

[0036]

[0037] For all target time-domain echo sequences detected in step 1), the above feature extraction method is used, and each target can extract a corresponding feature vector.

[0038] Preferably, the specific steps of step 4) using the time-frequency domain multi-scale permutation entropy feature for target recognition are:

[0039] A support vector machine classification algorithm is used, which is to find an optimal plane in the feature space to distinguish different categories. The selection criterion for this hyperplane is to maximize the interval between the two categories, that is, the minimum distance of the data points to the hyperplane. The feature matrix composed of multiple targets is divided into a training set and a test set, and the model hyperparameter optimization is constructed. Finally, the feature vectors of the test set are input into the model for testing to obtain the target recognition result.

[0040] Compared with the existing technology, the sea surface target recognition method based on multi-scale permutation entropy features in the time-frequency domain described in this technical solution has the following beneficial effects:

[0041] (1) The method proposed in this patent uses the time-frequency analysis method to comprehensively consider the time domain and frequency domain information, and fully utilizes the differences in energy distribution characteristics of the two types of targets in the time-frequency domain, thereby solving the problem that traditional target features are difficult to identify.

[0042] (2) The permutation entropy value is used to measure the mutation degree and randomness of the Doppler channel energy amplitude sequence, and the features are selected based on the mechanism of the energy distribution of the two types of targets to achieve the separability of the features of the two types of samples; and the coarse-grained processing at different time scales is used to optimize the multi-scale permutation entropy features to further improve the recognition performance.

[0043] (3) Under the existing data conditions, the method proposed in this patent has an average accuracy of 97.05% in identifying ship targets and floating targets within an observation time of 0.25s, which is 3.08% and 5.65% higher than the two existing methods respectively. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 are time-frequency domain images of two types of targets in the embodiment of the present invention. DETAILED DESCRIPTION

[0045] For better understanding and implementation, the following specific embodiments are given to explain in detail a method for recognizing sea surface targets based on multi-scale permutation entropy features in the time-frequency domain. The method for recognizing sea surface targets is as follows:

[0046] 1) Target echo data cache

[0047] The radar receives echo data in tracking mode. After pulse compression, clutter suppression, target detection and tracking processing, a total of k targets are detected in the current scene. The range unit where the strongest scattering point of the target is located is selected and its pulse echo data sequence is cached. For the kth target, it is expressed as {z k (1),z k (2),...,z k (L)}, where L represents the coherent pulse train length.

[0048] 2) Time-frequency domain analysis of target echo data

[0049] For a target time domain echo sequence z detected in step 1) k (m), select a rectangular window function with a length of M to perform sliding window interception processing, the rectangular window function R M (n) as follows:

[0050]

[0051] The window function slides forward W pulses each time it intercepts, and we get A subsequence X of length M i , represented by X i ={z k [(i-1)W+1],z k [(i-1)W+2],…,z k [(i-1)W+M]}, where

[0052] The i-th data segment X after the sliding window is intercepted i Perform time-frequency domain analysis, and the main analysis methods include fractional Fourier transform, Wigner-Ville distribution, smoothed pseudo-Wigner-Ville distribution, wavelet transform, etc. This patent uses the short-time Fourier transform method for time-frequency analysis. This method can obtain information in both the time domain and frequency domain dimensions of the signal, and can achieve coherent accumulation of target energy. It has been widely used in the time-frequency analysis of time-varying and non-stationary signals. First, perform Fourier transform on the time domain signal of a certain length each time to obtain the spectrum diagram of the signal in that time period, and then slide the time domain signal over the entire time period to obtain a collection of each spectrum segment. The result is a two-dimensional function of time-frequency. For this two-dimensional matrix, the rows of the matrix represent different Doppler channels, and the columns of the matrix represent different times. The formula for the short-time Fourier transform is as follows

[0053]

[0054] Among them, p(t-α) is the time window centered on α. ​​In order to utilize the difference in energy amplitude and avoid the reduction of target main lobe energy, this patent uses a rectangular window and the number of points when performing Fourier transform in segments is N. The time-frequency domain images of the two types of targets are as follows: Figure 1 As shown, Figure 1 (a) is the time-frequency domain image of the ship target, Figure 1 (b) is the time-frequency domain image of a floating target. The energy of the floating target swings back and forth between different Doppler channels over time, while the energy of the ship target changes less over time and is relatively stable. There are obvious characteristic differences between the time-frequency images of the two types of targets.

[0055] 3) Target time-frequency domain multi-scale permutation entropy feature extraction

[0056] First, the data segment X iThe absolute value of each data point in the time-frequency result is taken, that is, y(t,f)=|stft(t,f)|, and then the multi-scale permutation entropy value of the time-frequency sequence is calculated along the Doppler channel. For example, the k-th Doppler channel sequence in y(t,f) is taken out and expressed as y(t,k)={y(1,k),y(2,k),…,y(M,k)}. In order to obtain more feature information at different scales, the time-frequency sequence calculated along the Doppler channel is coarse-grained, and y is set. i (β)=y(t,k), the process is as follows:

[0057]

[0058] in, is the 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] 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. Arranging the elements of each row in the matrix Y in ascending order, we can obtain a set of symbol sequences γ1,γ2,…,γ consisting of the index of the position of each element in the row vector. m , expressed as S(l), that is

[0062] S(l)={γ1,γ2,…,γ m}, l=1,2,…,k, and k≤m! (6)

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

[0067] According to the above process, the permutation entropy value of y(t,f) at scale s can be obtained, and the multi-scale permutation entropy values ​​E1, E2…, E corresponding to all Doppler channels of the time-frequency results of the data segment can be obtained.k ,…E M , and select the minimum value σ i As the eigenvalue of this data segment, where σ i =min{E1,E2…,E k ,…E M By sequentially processing each data segment of the current target time domain echo sequence, the characteristic vector Φ of its time-frequency domain multi-scale permutation entropy is obtained, which is expressed as

[0068]

[0069] For all target time-domain echo sequences detected in step 1), the above feature extraction method can be used to extract a corresponding feature vector for each target. For example, if two targets are detected in the echo data, and the number of data segments obtained after sliding window truncation of the echo data of each target is M, then the target's time-frequency domain multi-scale permutation entropy feature vectors Φ1 and Φ2 can be obtained, and the number of columns of each vector is M.

[0070] 4) Target recognition using multi-scale permutation entropy features in the time-frequency domain

[0071] In step 3), we obtained the feature vectors of 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.

[0072] Three sets of measured data were used for feature extraction, and their basic information is shown in Table 1. The echo data was segmented, with each segment consisting of 4096 pulses. Each forward slide was 512 pulses, corresponding to a duration of approximately 0.25 seconds. A total of 249 eigenvalues ​​were extracted for each target, and a total of 498 eigenvalues ​​for both target types were obtained from each data set, forming a feature matrix.

[0073] Table 1 Introduction to measured data

[0074]

[0075] The present invention adopts the recognition method of support vector machine. First, each feature vector is assigned a label value during recognition, 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. Then, 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 2 -8 Start, maximum value 2 8At the end, the loop step size 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. Finally, the test set data is input into the model for recognition. The following four situations will appear in the test results:

[0076] ①True Positive (TP): The model correctly predicts a sample that is actually a ship target as a ship target;

[0077] ②False Negative (FN): The model mistakenly predicts a sample that is actually a ship target as a floating target;

[0078] ③False Positive (FP): The model mistakenly predicts a floating target as a ship target;

[0079] ④True Negative (TN): The model correctly predicts a sample that is actually a floating target as a floating target.

[0080] Table 2 Statistics of target recognition results of multi-scale permutation entropy features in time-frequency domain

[0081]

[0082] Table 3 Statistics of target recognition results based on waveform entropy features in the time-frequency domain

[0083]

[0084] Table 4 Statistics of target recognition results using entropy features proposed in reference [2]

[0085]

[0086] The recognition results are statistically analyzed, and the results are shown in Table 2. For the sake of comparison, the recognition results of multi-scale permutation entropy in the time-frequency domain as a feature are given in Table 3, and the recognition results of the entropy feature proposed in the literature [2] (Literature [2]: Zhao Yue, Chen Zhichun, Jiu Bo, et al. A method for extracting features for narrowband radar aircraft target classification based on time-frequency analysis [J]. Journal of Electronics and Information Technology. 2017, 39(9): 2225-2231) are shown in Table 4. It can be seen that the average recognition accuracy achieved by the existing methods is 93.97% (using a single feature) and 91.4% (using four features), while the average recognition accuracy of the present invention using a single feature is 97.05%. Under the existing data conditions, the recognition accuracy is improved by 3.08% and 5.65% compared with the existing methods.

Claims

1. A method for sea surface target recognition using multi-scale permutation entropy features in the time-frequency domain, 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: Time-frequency domain analysis of target echo data The target time domain echo sequence is segmented by sliding window, and each data segment is subjected to time-frequency analysis; Step 3: Target time-frequency domain multi-scale permutation entropy feature extraction After time-frequency domain analysis, coarse-graining processing is performed at different time scales along the Doppler channel to calculate the permutation entropy value of the corresponding scale. The minimum multi-scale permutation entropy value is then selected as the feature to finally form the target feature vector. Step 4: Target recognition using multi-scale permutation entropy features in the time-frequency domain The support vector machine classification algorithm is used to complete the recognition of two types of targets and test the recognition performance. The feature matrix composed of multiple targets is divided into training set and test set, and the model hyperparameter optimization is carried out, 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 is input into the model for testing to obtain the target recognition result.

2. A method for sea surface target recognition using multi-scale permutation entropy features in the time-frequency domain according to claim 1, characterized in that The specific steps of step 1 are: The radar receives echo data in tracking mode. After pulse compression, clutter suppression, target detection and tracking processing, it detects k targets in the current scene, selects the range unit where the strongest scattering point of the target is located, and caches its pulse echo data sequence z. k (m), for the kth target, it is expressed as {z k (1),z k (2),...,z k (L)}, where L represents the coherent pulse train length.

3. A method for sea surface target recognition using multi-scale permutation entropy features in the time-frequency domain according to claim 1, characterized in that The specific steps of step 2 are: For a target time domain echo sequence z detected in step 1 k (m), select a rectangular window function with a length of M to perform sliding window interception processing, the rectangular window function R M (n) as follows: The window function slides forward W pulses each time it intercepts, and we get A subsequence X of length M i , represented by X i ={z k [(i-1)W+1],z k [(i-1)W+2],…,z k [(i-1)W+M]}, where The i-th data segment X after the sliding window is intercepted i , the short-time Fourier transform method is used for time-frequency domain analysis to obtain information in both the time domain and frequency domain dimensions of the signal, and to achieve the accumulation of target energy. Each time the intercepted time domain signal is Fourier transformed, the spectrum diagram of the signal in that time period is obtained. The time domain signal is intercepted slidingly over the entire time period to obtain a collection of each spectrum segment. The result is a two-dimensional function of time-frequency. For this two-dimensional matrix, the rows of the matrix represent different Doppler channels, and the columns of the matrix represent different times. The short-time Fourier transform formula is as follows Among them, p(t-α) is the time window centered on α, a rectangular window is used, and the number of points when performing the segmented Fourier transform is N. The energy of the floating target swings back and forth between different Doppler channels as time changes, while the energy of the ship target changes less with time. There are obvious characteristic differences between the time-frequency images of the two types of targets.

4. A method for identifying sea surface targets using multi-scale permutation entropy features in the time-frequency domain according to claim 3, characterized in that The specific steps of step 3 target time-frequency domain multi-scale permutation entropy feature extraction are: First, the data segment X i Take the absolute value of each data point in the time-frequency result, that is, y(t,f)=|stft(t,f)|, and then calculate the multi-scale permutation entropy value of the time-frequency sequence along the Doppler channel, and take out the k-th Doppler channel sequence in y(t,f), expressed as y(t,k)={y(1,k),y(2,k),…,y(M,k)}. In order to obtain more feature information at different scales, the time-frequency sequence calculated along the Doppler channel is coarse-grained, and y i (β)=y(t,k), the process is as follows: in, is the 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 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 each row of the matrix Y is sorted in ascending order to obtain a set of symbol sequences γ1,γ2,…,γ consisting of the index of each element position in the row vector. m , expressed as S(l), that is S(l) = {γ1, γ2, …, γ m}, l = 1, 2, …, k, and k ≤ m! (6) 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, the permutation entropy value of y(t,f) at scale s is obtained, and the multi-scale permutation entropy values ​​E1, E2…, E corresponding to all Doppler channels of the time-frequency results of the data segment are obtained. k ,…E M , and select the minimum value σ i As the eigenvalue of this data segment, where σ i =min{E1,E2…,E k ,…E M }, process each data segment of the current target time domain echo sequence in turn to obtain the characteristic vector Φ of its time-frequency domain multi-scale permutation entropy, which is expressed as: For all target time domain echo sequences detected in step 1, the above feature extraction method is applied to extract the corresponding feature vector for each target.

Images