Detection method based on characteristic spectrum clustering

Abstract

The invention relates to a detection method based on characteristic spectrum clustering, and belongs to the technical field of radar signal processing. Comprising the following steps: 1) feature extraction: extracting nine features of a target time domain, a frequency domain and a time-frequency domain by using echo data after pulse compression processing to obtain a target feature value; 2) calculating a Spearman correlation coefficient and a Bhattacharyya distance; 3) generalized distance conversion and spectral clustering: converting the Spearman correlation coefficient of the target features into a generalized distance, performing spectral clustering by using a distance matrix, and clustering the features with relatively strong correlation into one class; and 4) feature optimization and detection. According to the method, the correlation of the features is quantitatively converted into a Spearman correlation coefficient matrix, the matrix is converted into a generalized distance matrix, the features with high correlation are clustered into one class by using a spectral clustering method, the features with high separability in each class are selected through feature optimization, and the selected features are more accurate and higher in efficiency.

Description

Technical Field

[0001] The invention relates to a detection method based on feature spectrum clustering, and belongs to the technical field of radar signal processing. Background Art

[0002] Feature detection is one of the effective methods for radar target detection. However, as more and more features are extracted, if the feature dimension is simply increased, not only will the ideal detection results fail to be achieved, but the amount of calculation will increase dramatically and a "dimensionality disaster" will occur. Due to the limited radar echo data, fewer features can fully characterize the radar data and complete the detection. However, there are still problems in how to select features in different scenarios. The existing feature selection methods mainly select features with less strong correlation based on the correlation of features for joint detection. Such selection is not accurate enough and requires manual participation.

[0003] Therefore, there is an urgent need for a detection method that can solve the above-mentioned technical problems. Summary of the invention

[0004] The purpose of the present invention is to solve the shortcomings of the above-mentioned prior art and provide a detection method based on feature spectral clustering, which quantitatively converts the correlation of features into a Spearman correlation coefficient matrix, and converts the matrix into a generalized distance matrix, uses the spectral clustering method to cluster features with strong correlation into one category, and selects features with strong separability in each category through feature optimization, so that feature selection is more accurate and efficient.

[0005] The detection method based on feature spectrum clustering of the present invention is special in that it comprises the following steps:

[0006] Step 1, feature extraction: using the echo data after pulse compression processing, extract 9 features of the target in time domain, frequency domain, and time-frequency domain to obtain the target feature value;

[0007] Step 2, calculation of Spearman correlation coefficient and Bhattacharyya distance: Calculate the Spearman correlation coefficient of the target feature and the Bhattacharyya distance between the target and clutter features;

[0008] Step 3, generalized distance conversion and spectral clustering: convert the Spearman correlation coefficient of the target feature into a generalized distance, and use the distance matrix to perform spectral clustering to cluster the features with strong correlation into one category;

[0009] Step 4, feature optimization and detection: extract the features with the largest Bhattacharyya distance in each category and sort them from high to low according to the Bhattacharyya distance. Select the three features with the largest Bhattacharyya distance and use the convex hull algorithm to perform target detection.

[0010] Preferably, the specific steps of step 1 are:

[0011] The radar echo data after pulse compression is used to extract 9 features from the time domain, frequency domain, and time-frequency domain. The extraction method is as follows:

[0012] 1) RAA is defined as the ratio of the average amplitude of the Clutter Under Test (CUT) to the average amplitude of the reference cell (RC), which is used for the amplitude difference between CUT and RC. The specific calculation is as follows:

[0013]

[0014] Among them, x(n) and x p (n) are the echo data of CUT and RC respectively, K is the number of cells;

[0015] 2) RPH is defined as the ratio of the peak value of the CUT pulse echo to the average amplitude of adjacent pulses. It is used to reflect the energy proportion of the target and sea clutter echo peaks in the total signal and the difference in peak fluctuation. The specific calculation is as follows:

[0016]

[0017] Δ=[-τ1,-τ2]∪[τ2,τ1] (4)

[0018] Among them, #Δ represents the number of units in the set Δ, and τ1 and τ2 together define the pulse range involved in the ratio operation;

[0019] 3) TEM is defined as the average value of the CUT echo information entropy. Since the two echoes fluctuate differently, it reflects the difference in the degree of confusion between the target and sea clutter signal waveforms. The specific calculation is as follows:

[0020] Use a rectangular window of width W and a step of width S to slide the N echo signals x of the CUT to obtain A short time-domain sequence {s i}

[0021]

[0022] in, represents rounding up, TE(i) represents the i-th short sequence s i The time domain entropy value of .

[0023] 4) FPAR is defined as the ratio of the Doppler peak value of CUT to the Doppler channel mean value, which is calculated as follows:

[0024]

[0025] Among them, f dis the Doppler frequency, N is the number of Doppler channels, and DAS stands for the discrete Fourier transform of the echo.

[0026] 5) RDPH is defined as the ratio of the Doppler peak value of CUT to the Doppler peak mean value of the reference channel, which can reflect the difference in the peak energy proportion and mutation degree of the two types of echo frequencies. The specific calculation is as follows:

[0027]

[0028] Among them, T r is the pulse repetition period, f d is the Doppler frequency. #γ represents the number of channels in the set γ, and Together they define the Doppler channel range involved in the ratio calculation.

[0029] 6) RVE is defined as the ratio of CUT to RC information entropy, which can reflect the degree of confusion of the signal waveform. The specific calculation is as follows:

[0030]

[0031] 7) RI is defined as the cumulative value of the time-frequency ridge on the CUT time-frequency spectrum, which reflects the energy intensity of the signal time-frequency ridge. The specific calculation is as follows:

[0032]

[0033] Among them, SPWVD(n,f d |x) represents the smoothed pseudo-Wigner-Wiley distribution time-frequency transform of the CUT received sequence x;

[0034] 8) MS is defined as the maximum number of points that pass the threshold in a single connected area in the time-frequency spectrum binary graph. The specific calculation is as follows:

[0035] MS(x;x p )==max{num(Γ1),...,num(Γ C )} (20)

[0036] Among them, Γ c is a connected region set Γ=[Γ1,...,Γ C ] is a connected region, num(Γ c ) represents the connected region Γ c The number of internal time-frequency ridges;

[0037] 9) NR is defined as the number of connected regions in the time-frequency binary graph, reflecting the discrete degree of time-frequency ridge energy. The specific calculation is as follows:

[0038] NR(x;x p)=C (21).

[0039] Preferably, the specific steps of step 2 are:

[0040] Spearman correlation is a nonparametric statistical method used to measure the monotonic relationship between two variables. It does not require the variables to follow a specific distribution, accurately describes the correlation strength and properties of interval feature data that do not conform to the normal distribution assumption, and is suitable for feature distribution in difficult scenarios. For two feature sequences x = (x1, x2, ..., x n ), y=(y1,y2,...,y n ) The calculation method is as follows:

[0041]

[0042] d i =rank(x i )-rank(y i ) (twenty three)

[0043]

[0044] Where, II(x) is an indicator function, which is 1 when the conditions in the brackets are met, otherwise it is 0. The range of the Spearman correlation coefficient ρ is -1 to 1. When it is positive, it indicates positive correlation, when it is negative, it indicates negative correlation, and when it is 0, it indicates that there is no correlation between the data. The larger the absolute value of ρ, the greater the correlation between the data. |ρ|>0.8 indicates strong correlation, |ρ|<0.2 indicates weak correlation, and the rest are general correlations.

[0045] Bhattacharyya distance (BD) is used as a reference for feature selection after clustering. The calculation formula of BD is as follows:

[0046]

[0047] Where p(x) and q(x) are the probability density functions (PDF) of the features in the clutter and target data, respectively. In physical terms, they represent the overlap of the two distribution functions. When BD is larger, it means that the overlap between the target and the clutter is smaller, and the feature separability is better, and vice versa.

[0048] Preferably, the specific steps of step 3 are:

[0049] 1) Convert the Spearman correlation coefficient matrix into a generalized distance matrix: Assuming that the Spearman correlation coefficient matrix calculated in step 2) is C, the generalized distance matrix between the features is defined as:

[0050] Z=-log|C| (26)

[0051] 2) Convert the generalized distance matrix into a similarity matrix:

[0052]

[0053] Among them, σ is the width parameter of the Gaussian kernel, which determines the range of similarity;

[0054] 3) Construct the normalized Laplace matrix:

[0055]

[0056] Where D is a diagonal matrix with diagonal elements D ij is the sum of the elements in the i-th row of the similarity matrix W, that is, D ii =∑ j W ij , I is the identity matrix;

[0057] 4) Calculate eigenvectors: Calculate the eigenvectors of the Laplacian matrix. Usually, the eigenvectors corresponding to the first k smallest non-zero eigenvalues ​​are selected, where k is the number of clusters you want to obtain;

[0058] 5) Clustering: The feature vectors obtained in the previous step are used as new data points, and then these feature vectors are clustered using k-means to obtain k clustered feature groups.

[0059] Preferably, the specific steps of step 4 are:

[0060] Based on the Bhattacharyya distance, the clustered features are screened and the convex hull algorithm is used for target detection. The specific steps are as follows:

[0061] 1) If the number of clusters is equal to 3, the feature with the largest Bhattacharyya distance in each category is selected to obtain the three features used;

[0062] 2) If the number of clusters is greater than 3, first select the feature with the largest Bhattacharyya distance in each category, and then select the three features with the largest Bhattacharyya distance from the screened features to obtain the three features used;

[0063] 3) The number of feature vectors of clutter in the feature set S of the three features used is W, the false alarm rate is set to PFA, and the number of false alarms is calculated to be N F =W*PFA;

[0064] 4) Find the N in S that is farthest from the clutter target sample gathering area F Sample points are removed;

[0065] 5) Generate the convex hull decision area Ω using the clutter feature points after removing the sample points;

[0066] 6) Calculate the number of target feature points that fall outside the decision area and divide it by the number of target feature points to obtain the detection probability Pd.

[0067] Compared with the prior art, the detection method based on feature spectrum clustering described in the present invention has the following beneficial effects:

[0068] (1) The method proposed in the present invention utilizes the multi-domain and multi-dimensional characteristics of radar to reflect the echo data more fully.

[0069] (2) The correlation and separability of features are used to comprehensively consider the optimal feature combination in the current scenario, which has strong scene adaptability and better detection performance.

[0070] (3) It essentially uses a combination of three-dimensional features, which has less computational complexity and higher practicality than high-dimensional feature detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is a detection flow chart of a detection method based on feature spectrum clustering of the present invention;

[0072] Figure 2 It is the Spearman correlation coefficient matrix diagram of level 2 sea state characteristics;

[0073] Figure 3 is the data Bhattacharyya distance diagram; in the figure, (a) is the data Bhattacharyya distance diagram under different sea conditions, and (b) is the data Bhattacharyya distance diagram under different ground-grazing angles;

[0074] Figure 4 It is the flow chart of k-means algorithm;

[0075] Figure 5 It is the echo graph of different sea state data;

[0076] Figure 6 It is the echo graph of the flight data;

[0077] Figure 7 It is a graph of the ground contact angle and signal-to-noise ratio of the flight data;

[0078] Figure 8 is the cluster number-detection probability diagram; in the figure, (a) is the detection probability diagram of different sea state data; (b) is the detection probability diagram of different ground-grazing angle data;

[0079] Fig. 9 It is a detection method-detection probability diagram; in the figure, (a) is the detection method-detection probability diagram under different sea conditions, and (b) is the detection method-detection probability diagram at different ground-grabbing angles. DETAILED DESCRIPTION

[0080] For better understanding and implementation, a specific implementation is given below in conjunction with the accompanying drawings to describe in detail a method for identifying floating targets on the sea surface based on a recursive graph of the present invention. It should be noted that this embodiment introduces the method proposed by the present invention by taking the identification of ship targets and buoy targets as an example.

[0081] A detection method based on feature spectrum clustering in this embodiment, the detection flow chart is as follows Figure 1 As shown, the following steps are included:

[0082] Step 1, feature extraction: using the echo data after pulse compression processing, extract 9 features of the target in time domain, frequency domain, and time-frequency domain to obtain the target feature value;

[0083] Step 2, calculation of Spearman correlation coefficient and Bhattacharyya distance: Calculate the Spearman correlation coefficient of the target feature and the Bhattacharyya distance between the target and clutter features;

[0084] Step 3, generalized distance conversion and spectral clustering: convert the Spearman correlation coefficient of the target feature into a generalized distance, and use the distance matrix to perform spectral clustering to cluster the features with strong correlation into one category;

[0085] Step 4, feature optimization and detection: extract the features with the largest Bhattacharyya distance in each category and sort them from high to low according to the Bhattacharyya distance. Select the three features with the largest Bhattacharyya distance and use the convex hull algorithm to perform target detection.

[0086] The specific steps are as follows:

[0087] 1) Feature extraction

[0088] This embodiment uses the radar echo data after pulse compression to extract 9 features from the time domain, frequency domain, and time-frequency domain. The extraction method is as follows:

[0089] 1. RAA is defined as the ratio of the average amplitude of the Clutter Under Test (CUT) to the average amplitude of the reference cell (ReferenceCell), which can be used to calculate the amplitude difference between CUT and RC. The specific calculation is as follows:

[0090]

[0091]

[0092] Among them, x(n) and x p (n) are the echo data of CUT and RC respectively, and K is the number of cells.

[0093] 2. RPH is defined as the ratio of the peak value of the CUT pulse echo to the average amplitude of adjacent pulses. It can be used to reflect the energy proportion of the target and sea clutter echo peaks in the total signal and the difference in peak fluctuation. The specific calculation is as follows:

[0094]

[0095] Δ=[-τ1,-τ2]∪[τ2,τ1] (4)

[0096] Among them, #Δ represents the number of units in the set Δ, and τ1 and τ2 together define the pulse range participating in the ratio operation.

[0097] 3. TEM is defined as the average value of the CUT echo information entropy. Since the two echoes fluctuate differently, it reflects the difference in the degree of confusion between the target and sea clutter signal waveforms. The specific calculation is as follows:

[0098] Use a rectangular window of width W and a step of width S to slide the echo signal x of the CUT to obtain A short time-domain sequence {s i}

[0099]

[0100] in, represents rounding up, TE(i) represents the i-th short sequence s i The time domain entropy value of .

[0101] 4. FPAR is defined as the ratio of the Doppler peak value of the CUT to the Doppler channel mean value, and is calculated as follows:

[0102]

[0103] Among them, f d is the Doppler frequency, N is the number of Doppler channels, and DAS stands for the discrete Fourier transform of the echo.

[0104] 5. RDPH is defined as the ratio of the Doppler peak value of CUT to the Doppler peak value average of the reference channel, which can reflect the difference in the peak energy proportion and mutation degree of the two types of echo frequencies. The specific calculation is as follows:

[0105]

[0106] Among them, T r is the pulse repetition period, f d is the Doppler frequency. #γ represents the number of channels in the set γ, and Together they define the Doppler channel range involved in the ratio calculation.

[0107] 6. RVE is defined as the ratio of CUT to RC information entropy, which can reflect the degree of confusion of the signal waveform. The specific calculation is as follows:

[0108]

[0109]

[0110] 7. RI is defined as the cumulative value of the time-frequency ridge on the CUT time-frequency spectrum, which reflects the energy intensity of the signal time-frequency ridge. The specific calculation is as follows:

[0111]

[0112] Among them, SPWVD(n,f d |x) represents the smoothed pseudo-Wigner-Wiley distribution time-frequency transform of the CUT received sequence x.

[0113] 8. MS is defined as the maximum number of points that pass the threshold in a single connected area in the time-frequency spectrum binary graph. The specific calculation is as follows:

[0114] MS(x;x p )==max{num(Γ1),...,num(Γ C )} (20)

[0115] Among them, Γ c is a connected region set Γ=[Γ1,...,Γ C ] is a connected region, num(Γ c ) represents the connected region Γ c The number of internal time-frequency ridges.

[0116] 9. NR is defined as the number of connected regions in the time-frequency binary graph, reflecting the discrete degree of time-frequency ridge energy. The specific calculation is as follows:

[0117] NR(x;x p )=C (21)

[0118] 2) Calculation of Spearman correlation coefficient and Bhattacharyya distance

[0119] Spearman correlation is a nonparametric statistical method used to measure the monotonic relationship between two variables. Spearman correlation is based on the rank of the variables, which makes it effective for any type of monotonic relationship. It does not require the variables to follow a specific distribution, accurately describes the correlation strength and properties of interval feature data that do not conform to the normal distribution assumption, and is suitable for feature distributions in difficult scenarios. For two feature sequences x = (x1, x2, ..., x n ), y=(y1,y2,...,y n ) The calculation method is as follows:

[0120]

[0121] d i =rank(x i )-rank(y i ) (twenty three)

[0122]

[0123] Among them, II(x) is an indicator function, which is 1 when the conditions in the brackets are met, otherwise it is 0. The range of the Spearman correlation coefficient ρ is -1 to 1. When it is positive, it indicates positive correlation, when it is negative, it indicates negative correlation, and when it is 0, it indicates that there is no correlation between the data. The larger the absolute value of ρ, the greater the correlation between the data. When |ρ|>0.8, it is a strong correlation, when |ρ|<0.2, it is a weak correlation, and the rest are general correlations. Taking the Naval Aviation University shared data set level 2 sea conditions as an example, the Spearman correlation coefficients of the 9 features are calculated as follows Figure 2 shown.

[0124] Separability refers to the ability of a feature to distinguish a target from clutter, and is an important indicator of detection. The patent uses Bhattacharyya distance (BD) as a reference for feature selection after clustering. The calculation formula of BD is as follows:

[0125]

[0126] Where p(x) and q(x) are the probability density functions (PDF) of the features in the clutter and target data, respectively. In physical terms, they represent the overlap of the two distribution functions. When BD is larger, it means that the overlap between the target and the clutter is smaller, and the feature separability is better, and vice versa. The Bhattacharyya distances under different ground-grazing angles of the Naval Aviation University shared data set and the Yantai flight data are calculated as follows: Figure 3 shown.

[0127] 3) Generalized distance transformation and spectral clustering

[0128] After obtaining the echo features and the Bhattacharyya distance and Spearman correlation coefficient of the features in the data in step 1) and step 2), generalized distance transformation and spectral clustering can be performed. The specific steps are as follows:

[0129] (1) Convert the Spearman correlation coefficient matrix into a generalized distance matrix: Assume that the Spearman correlation coefficient matrix calculated in step 2) is C, and define the generalized distance matrix between features as:

[0130] Z=-log|C| (26)

[0131] (2) Convert the generalized distance matrix into a similarity matrix:

[0132]

[0133] Among them, σ is the width parameter of the Gaussian kernel, which determines the range of similarity.

[0134] (3) Construct the normalized Laplace matrix:

[0135]

[0136] Where D is a diagonal matrix with diagonal elements D ij is the sum of the elements in the i-th row of the similarity matrix W, that is, D ii =∑ j W ij , I is the identity matrix.

[0137] (4) Calculate eigenvectors: Calculate the eigenvectors of the Laplacian matrix. Usually, the eigenvectors corresponding to the first k smallest non-zero eigenvalues ​​are selected, where k is the number of clusters you want to obtain.

[0138] (5) Clustering: The feature vectors obtained in the previous step are used as new data points, and then these feature vectors are clustered using k-means. The k-means flow chart is as follows: Figure 4 As shown, the k clustered feature groups are obtained.

[0139] 4) Feature optimization and target detection

[0140] Based on the Bhattacharyya distance, the clustered features are screened and the convex hull algorithm is used for target detection. The specific steps are as follows:

[0141] (1) If the number of clusters is equal to 3, the feature with the largest Bhattacharyya distance in each cluster is selected to obtain the three features used.

[0142] (2) If the number of clusters is greater than 3, first select the feature with the largest Bhattacharyya distance in each category, and then select the three features with the largest Bhattacharyya distance from the screened features to obtain the three features used.

[0143] (3) The number of feature vectors of clutter in the feature set S of the three features used is W, the false alarm rate is set to PFA, and the number of false alarms is calculated to be N F =W*PFA.

[0144] (4) Find the N in S that is farthest from the clutter target sample gathering area F sample points and remove them.

[0145] (5) Generate the convex hull decision area Ω using the clutter feature points after removing the sample points.

[0146] (6) Calculate the number of target feature points that fall outside the decision area and divide it by the number of target feature points to obtain the detection probability Pd.

[0147] The method in this embodiment is verified by using the 2-5 level sea state data and Yantai flight data from the Naval Aviation University shared data set. The time domain echo diagram of the 2-5 level sea state data is as follows: Figure 5 As shown, the time domain echo diagram of Yantai flight data is as follows Figure 6 To facilitate the analysis of the flight data, every 40960 pulses are divided into a group of data, a total of 10 groups, each group of data corresponds to the ground angle and signal noise Figure 7 shown.

[0148] When clustering, the number of clusters is an important parameter. If the number of clusters is too small, one type of feature will contain more than one type of highly separable features, thus missing the better features. If the number of clusters is too large, the relevant information between features will not be fully utilized, and ideal detection results cannot be obtained. Set the number of clusters to 3-7, the false alarm rate PFA to 0.001, and calculate the detection probability of different sea conditions and different ground-grazing angle data. Figure 8 As shown in the figure, it is found that the number of clusters has little effect on the detection effect of different sea state data, while the detection effect is best when the number of clusters k=6 in the large ground-grabbing angle data. Considering comprehensive considerations, k=6 is set in the detection.

[0149] Keep the settings, the same as RAA-RDPH-RVE in [1] (SHUIPenglang, LIDongchen, and XU Shuwen. Tri-featurebased detection of floating small targets in sea clutter [J]. IEEE Transactions on Aerospace&Electronic Systems, 2014, 50(2): 1416–1430. doi: 10.1109 / TAES.2014.120657.), and [2] (SHI SAINAN, SHUIPENGLANG. Sea-surface floating small target detection by one-class classifier in time-frequency feature space [J]. IEEE Transactions on Geoscience and Remote Sensing, 2018, 56(11): 6395-6411.), the detection effects of RAA-FPAR-RDPH and the combination of three-feature RAA-RPH-TEM in the literature [3] (Dong Yunlong, Zhang Zhaoxiang, Ding Hao, et al. Detection Method for Small Targets in Sea Clutter Based on Three-feature Prediction[J]. Journal of Radars, 2023, 12(4): 762-775. doi: 10.12000 / JR23037.). Fig. 9 It can be seen that the detection performance of the method proposed in this embodiment is better than that of other detectors in high sea conditions and large ground-grabbing angles. After calculation, the detection probability of the detector proposed in this article is 3.95% higher than that of the best detector compared in the data of high sea conditions of level 4 to 5. In the data of ground-grabbing angles above 60°, the detection probability of the detector proposed in this embodiment is 4.53% higher than that of the best detector compared.

Claims

1. A detection method based on feature spectrum clustering, characterized in that The following steps are involved: Step 1, feature extraction: using the echo data after pulse compression processing, extract 9 features of the target in time domain, frequency domain, and time-frequency domain to obtain the target feature value; Step 2, calculation of Spearman correlation coefficient and Bhattacharyya distance: Calculate the Spearman correlation coefficient of the target feature and the Bhattacharyya distance between the target and clutter features; Step 3, generalized distance conversion and spectral clustering: convert the Spearman correlation coefficient of the target feature into a generalized distance, and use the distance matrix to perform spectral clustering to cluster the features with strong correlation into one category; Step 4, feature optimization and detection: extract the features with the largest Bhattacharyya distance in each category and sort them from high to low according to the Bhattacharyya distance. Select the three features with the largest Bhattacharyya distance and use the convex hull algorithm to perform target detection.

2. A detection method based on feature spectrum clustering according to claim 1, characterized in that The specific steps of step 1 are: The radar echo data after pulse compression is used to extract 9 features from the time domain, frequency domain, and time-frequency domain. The extraction method is as follows: 1) RAA is defined as the ratio of the average amplitude of the CUT to be tested and the reference cell (Reference Cell), which is used for the amplitude difference between CUT and RC. The specific calculation is as follows: Among them, x(n) and x p (n) are the echo data of CUT and RC respectively, K is the number of cells; 2) RPH is defined as the ratio of the peak value of the CUT pulse echo to the average amplitude of adjacent pulses. It is used to reflect the energy proportion of the target and sea clutter echo peaks in the total signal and the difference in peak fluctuation. The specific calculation is as follows: Δ=[-τ1,-τ2]∪[τ2,τ1] (4) Among them, #Δ represents the number of units in the set Δ, and τ1 and τ2 together define the pulse range involved in the ratio operation; 3) TEM is defined as the average value of the CUT echo information entropy. Since the two echoes fluctuate differently, it reflects the difference in the degree of confusion between the target and sea clutter signal waveforms. The specific calculation is as follows: Use a rectangular window of width W and a step of width S to slide the N echo signals x of the CUT to obtain A short time-domain sequence {s i }; in, represents rounding up, TE(i) represents the i-th short sequence s i The time domain entropy value of 4) FPAR is defined as the ratio of the Doppler peak value of CUT to the Doppler channel mean value, which is calculated as follows: Among them, f d is the Doppler frequency, N is the number of Doppler channels, and DAS represents the discrete Fourier transform of the echo; 5) RDPH is defined as the ratio of the Doppler peak value of CUT to the mean Doppler peak value of the reference channel, reflecting the difference in the peak energy proportion and mutation degree of the two types of echo frequencies. The specific calculation is as follows: Among them, T r is the pulse repetition period, f d is the Doppler frequency, #γ is the number of channels in the set γ, and Together, the Doppler channel range involved in the ratio calculation is defined; 6) RVE is defined as the ratio of CUT to RC information entropy, reflecting the degree of confusion of the signal waveform. The specific calculation is as follows: 7) RI is defined as the cumulative value of the time-frequency ridge on the CUT time-frequency spectrum, which reflects the energy intensity of the signal time-frequency ridge. The specific calculation is as follows: Among them, SPWVD(n,f d |x) represents the smoothed pseudo-Wigner-Wiley distribution time-frequency transform of the CUT received sequence x; 8) MS is defined as the maximum number of points that pass the threshold in a single connected area in the time-frequency spectrum binary graph. The specific calculation is as follows: MS(x;x p )==max{num(Γ1),...,num(Γ C )}(20) Among them, Γ c is a connected region set Γ=[Γ1,...,Γ C ] is a connected region in num(Γ c ) represents the connected region Γ c The number of internal time-frequency ridges; 9) NR is defined as the number of connected regions in the time-frequency binary graph, reflecting the discrete degree of time-frequency ridge energy. The specific calculation is as follows: NR(x;x p )=C(21)。 3. A detection method based on feature spectrum clustering according to claim 1, characterized in that The specific steps of step 2 are: Spearman correlation is a nonparametric statistical method used to measure the monotonic relationship between two variables. It does not require the variables to follow a specific distribution, accurately describes the correlation strength and properties of interval feature data that do not conform to the normal distribution assumption, and is suitable for feature distribution in difficult scenarios. For two feature sequences x = (x1, x2, ..., x n ), y=(y1,y2,...,y n ) The calculation method is as follows: d i =rank(x i )-rank(y i )(23) Where, II(x) is an indicator function, which is 1 when the conditions in the brackets are met, otherwise it is 0. The range of the Spearman correlation coefficient ρ is -1 to 1. When it is positive, it indicates positive correlation, when it is negative, it indicates negative correlation, and when it is 0, it indicates that there is no correlation between the data. The larger the absolute value of ρ, the greater the correlation between the data. |ρ|>0.8 indicates strong correlation, |ρ|<0.2 indicates weak correlation, and the rest are general correlations. Bhattacharyya distance BD is used as a reference for feature selection after clustering. The calculation formula of BD is as follows: Where p(x) and q(x) are the probability distribution functions PDF of the feature in the clutter and target data, respectively. In physical terms, they represent the overlap of the two distribution functions. When BD is larger, it means that the overlap between the target and the clutter is smaller, and the feature separability is better, and vice versa.

4. A detection method based on feature spectrum clustering according to claim 1, characterized in that The specific steps of step 3 are: 1) Convert the Spearman correlation coefficient matrix into a generalized distance matrix: Assuming that the Spearman correlation coefficient matrix calculated in step 2) is C, the generalized distance matrix between the features is defined as: Z=-log|C| (26) 2) Convert the generalized distance matrix into a similarity matrix: Among them, σ is the width parameter of the Gaussian kernel, which determines the range of similarity; 3) Construct the normalized Laplace matrix: Where D is a diagonal matrix with diagonal elements D ij is the sum of the elements in the i-th row of the similarity matrix W, that is, D ii =Σ j W ij , I is the identity matrix; 4) Calculate eigenvectors: Calculate the eigenvectors of the Laplacian matrix. Usually, the eigenvectors corresponding to the first k smallest non-zero eigenvalues ​​are selected, where k is the number of clusters you want to obtain; 5) Clustering: The feature vectors obtained in the previous step are used as new data points, and then these feature vectors are clustered using k-means to obtain k clustered feature groups.

5. A detection method based on feature spectrum clustering according to claim 1, characterized in that The specific steps of step 4 are: Based on the Bhattacharyya distance, the clustered features are screened and the convex hull algorithm is used for target detection. The specific steps are as follows: 1) If the number of clusters is equal to 3, the feature with the largest Bhattacharyya distance in each category is selected to obtain the three features used; 2) If the number of clusters is greater than 3, first select the feature with the largest Bhattacharyya distance in each category, and then select the three features with the largest Bhattacharyya distance from the screened features to obtain the three features used; 3) The number of feature vectors of clutter in the feature set S of the three features used is W, the false alarm rate is set to PFA, and the number of false alarms is calculated to be N F =W*PFA; 4) Find the N in S that is farthest from the clutter target sample gathering area F Sample points are removed; 5) Generate the convex hull decision area Ω using the clutter feature points after removing the sample points; 6) Calculate the number of target feature points that fall outside the decision area and divide it by the number of target feature points to obtain the detection probability Pd.

Images