An intelligent target detection method based on variable dimension feature screening

Abstract

The present invention discloses an intelligent target detection method based on variable-dimensional feature screening. Since the existence of feature redundancy in the process of constructing a multi-dimensional feature space will lead to an increase in unnecessary computing costs. In view of the problem that it is impossible to adaptively select feature combinations according to actual scenarios under the condition that the feature dimensions are not restricted during the feature screening process, a variable-dimensional feature screening method based on geometric metrics and variance inflation factors is proposed. While constraining the correlation between features, this method searches for feature combinations that maximize the degree of inter-class separability, removes redundant features while retaining sufficient useful information, and establishes a comprehensive and effective feature vector. Then, based on the measured data, an intelligent detector with a controllable false alarm rate is used to complete target detection.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar target detection, and in particular relates to an intelligent target detection method based on variable dimensional feature screening. Background Art

[0002] Achieving efficient target detection in sea clutter environments is a key research area for modern radar. Extracting multidimensional features from echo signals for deep clutter perception and utilizing feature-based intelligent target detectors to achieve clutter suppression and efficient target detection is a key approach to refined signal processing in clutter environments. However, the application of this technology currently faces the following challenges:

[0003] On the one hand, as the number of features increases, the dimensionality of the feature space continues to grow, and the computational complexity increases dramatically. To maintain classifier accuracy while minimizing computational complexity, it is necessary to select an appropriate feature subset from the multi-feature set for subsequent classification. Traditional feature screening methods such as filtering, wrapping, and embedding may fail to consider the correlation characteristics between features and incur excessive computational overhead in learner training. Principal Component Analysis (PCA) is computationally simple and widely used for feature subset screening, and is not restricted by feature dimension. However, since the principal components remain orthogonal after screening, the correlation between features is completely ignored. In addition, there are feature screening methods based on Riemann distance and Bhattacharyya distance, which measure the degree of separation between features. By fixing the feature dimension, they use gradient descent or other intelligent optimization algorithms to optimize the separability. However, these methods are limited to feature subset selection within a fixed feature dimension, ignoring the diversity of data, resulting in failure to achieve optimal classification results and low robustness. Therefore, how to adaptively select feature combinations with low redundancy, good representation and unlimited feature dimensions in different scenarios and establish comprehensive and effective feature vectors is one of the key issues to be solved in clutter suppression and target detection.

[0004] On the other hand, due to the strict requirements of radar on false alarm probability, it is usually necessary to reach 10 -3 level or below, and the requirement for missed detection probability can be 10 -1 Some traditional intelligent target detection methods based on machine learning and deep learning cannot meet the required false alarm probability. Therefore, for target detection in radar, it is necessary to design a detector that can strictly control the false alarm probability while maintaining good target detection performance.

[0005] In view of this, the present invention provides an intelligent target detection method based on variable dimensional feature screening. Summary of the Invention

[0006] The present invention proposes an intelligent target detection method based on variable-dimensional feature screening. Since the existence of feature redundancy in the process of constructing a multidimensional feature space will lead to an increase in unnecessary computational costs, in the feature screening process, it is impossible to adaptively select feature combinations based on actual scenarios under the condition that the feature dimensions are not restricted. In this paper, an intelligent target detection method based on variable-dimensional feature screening is proposed. While constraining the correlation between features, this method searches for feature combinations that maximize the degree of inter-class separability, removes redundant features while retaining sufficient useful information, and establishes a comprehensive and effective feature vector. Target detection is then completed based on measured data using an intelligent detector with a controllable false alarm rate.

[0007] The technical solution of the present invention is:

[0008] An intelligent target detection method based on variable-dimensional feature screening includes the following steps:

[0009] Step S1: extract features of pure sea clutter and target units in the data in multiple processing domains, establish a feature sample space, calculate the total Kullback–Leibler divergence (tKLD) of the initial combination of feature samples, and calculate the variance inflation factor (VIF) of each feature;

[0010] Step S2: Compare the variance inflation factor in step S1 with the empirical threshold to obtain the feature index where there is severe multicollinearity between feature combinations, and take the intersection of the feature indexes selected from the target and clutter units to obtain the feature index vector where there is multicollinearity between the target and clutter units.

[0011] Step S3: Eliminate the features corresponding to the index vectors in step S2 one by one. After each elimination operation, calculate the tKLD of the remaining feature combination. If the largest tKLD is not less than the initial value, the feature elimination is considered effective and the feature combination completes a screening operation.

[0012] Step S4, recalculate tKLD for the remaining feature combinations in step S3, and repeat steps S1-S3 until the termination condition is reached, that is, the screening of the entire feature combination is completed;

[0013] Step S5: Label the filtered feature combinations, divide them into training set and test set in a ratio of 7:3, and input them into the intelligent detector for detection;

[0014] Step S6, initialize the parameters of the intelligent detector; construct a Classification and Regression Tree (CART), count the number of misjudgments of clutter samples under different pruning parameters, and if the threshold is reached, output the weight factor at this time and update the decision tree. If the threshold is not reached, adjust the step direction of the weight factor according to the overall data distribution until the update is completed and the detection is completed.

[0015] The step S1 includes the following steps:

[0016] Step S11, extracting features from the target unit and pure sea clutter unit in the data, wherein the extracted features include fractal, time domain, and Doppler domain features. The fractal features include the time domain Hurst exponent and the frequency domain Hurst exponent; the time domain features include: relative average amplitude, relative amplitude variance, relative time domain waveform entropy, and decorrelation time; the Doppler domain features include: relative Doppler peak-to-average ratio, relative Doppler skewness, relative Doppler kurtosis, relative Doppler vector entropy, relative Doppler coefficient of variation, and average Doppler spectrum radius; the above features are sorted and numbered in order to facilitate subsequent indexing;

[0017] Step S12: Consider all features as a feature combination, calculate the inter-class separability tKLD of the target and clutter in the feature combination, denoted as tKLD0, and the variance inflation factor VIF of each feature;

[0018] The step S3 includes the following steps:

[0019] Step S31: Based on the feature index obtained in step S2, a subset of candidate features is obtained. The features in the subset are eliminated one by one. First, the first feature is eliminated, and the tKLD of the remaining feature combination is recalculated. Then, the first feature is retained, the second feature is eliminated, and the tKLD of the new feature combination is calculated again, until all features in the candidate feature subset are traversed.

[0020] Step S32: Calculate the maximum value of tKLD for all feature combinations in step S31. If the maximum value is greater than the initial value tKLD0, find the feature index corresponding to the maximum value and delete the feature.

[0021] The step S6 includes the following steps:

[0022] Step S61, setting the initial parameters of the decision tree model, the parameters include false alarm probability, initial weight factors of target and clutter, weight factor step value and minimum number of samples of leaf nodes, etc., wherein the minimum number of samples of leaf nodes msl ranges from [1, msl max ].

[0023] Step S62: Based on the pruning parameters of step S61, a loop is set to generate a decision tree under different msl, and the number of clutter misjudgments output by the model is measured. If it is the same as the theoretical number of misjudgments α under the set false alarm probability, the weight factor of the target and clutter at this time is output. If they are different, the loop is continued until all values ​​of msl are traversed to obtain the misjudgment number vector δ;

[0024] Step S63: If the number of clutter misjudgments that meets the condition still does not exist in step S62, then compare the value in δ with α. If most of the values ​​are greater than α, then increase the clutter weight according to the step value; otherwise, decrease the weight. Repeat step S62 until δ contains the same value as α.

[0025] Step S64, update the decision tree under the weight factors and corresponding MSL that meet the conditions, use the test set to test, complete target detection, and output the actual false alarm probability and target detection probability results.

[0026] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0027] (1) The feature screening method based on total KL divergence and variance inflation factor proposed in the present invention can maximize the separability between categories while removing feature redundancy. It can adaptively select the best feature combination for measured data in different scenarios, and the feature dimension within the combination is not limited.

[0028] (2) The intelligent detector model with controllable false alarm probability based on decision tree in multi-dimensional space proposed in the present invention improves the traditional CART decision tree. By adjusting the weighting factor and optimizing the pruning parameters, it realizes precise control of the false alarm probability in high-dimensional feature space, and has better performance in terms of target detection performance and computational complexity. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a schematic diagram of the process of the present invention;

[0030] Figure 2 Schematic diagram of the process of the feature screening module of the present invention;

[0031] Figure 3 Schematic diagram of the process of the intelligent detector module of the present invention;

[0032] Figure 4 The target detection performance of the present invention under four measured data scenarios with different observation durations;

[0033] Figure 5 Comparison of the average detection probability of the present invention and other detectors under observation durations of 0.512s and 1.024s; DETAILED DESCRIPTION

[0034] The accompanying drawings are for illustrative purposes only and are not to be construed as limiting the present invention;

[0035] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0036] An intelligent target detection method based on variable dimension feature screening, the flow chart is as follows Figure 1 As shown, the method includes the following steps:

[0037] Step S1: Extract 12 features from the pure sea clutter and target units in the measured data in the fractal, time and Doppler domains and perform zero mean conversion. The number of features is recorded as θ. The above features are numbered in order to facilitate subsequent indexing. Initialize the projection matrix D for feature screening, that is, initialize D to the index matrix of θ features: D = E θ ;

[0038] Then, the feature sample space is established, and the feature sample space of clutter and target is denoted as F c and F t , then the covariance matrix of the two types of samples can be expressed as:

[0039]

[0040]

[0041] Among them, I0 and I1 are the number of samples of clutter and target respectively; based on the covariance matrix and The tKLD of the initial combination of two types of features is:

[0042]

[0043] in, |·| is the matrix determinant operation, and tr(·) is the matrix trace operation.

[0044] Then, the variance inflation factor (VIF) of each feature in the pure sea clutter and target cells is calculated:

[0045]

[0046] The variance inflation factors of the clutter and target units are denoted as VIF c,p and VIF t,p , where p represents the feature number, p = 1, 2, ..., θ; is the coefficient of determination obtained by regressing feature p on all other variables, representing the square of the correlation between the independent variable and other independent variables, which can be obtained by the multiple correlation coefficient; if the feature sample F tThe Pearson correlation coefficient matrix is ​​represented as S t , then VIF can be further written as:

[0047]

[0048] Among them, M pp Indicates S t Remove the remainder obtained by removing the p-th row and p-th column;

[0049] Step S2: Compare the variance inflation factor in step S1 with the empirical threshold. It is generally believed that if VIF is greater than or equal to 10, there is severe multicollinearity. Select the feature numbers of the target and clutter units with severe multicollinearity and take the intersection to obtain the vector H. 1×n , where n is the number of features;

[0050] Step S3, vector H in step S2 1×n The corresponding features are eliminated one by one, that is, for H 1×n For each feature number in , remove the corresponding column in the projection matrix D and obtain n projection matrices D1,...,D n , at this time, the feature space is compressed and expressed as:

[0051]

[0052] Then the covariance matrix of the feature space is updated as:

[0053]

[0054] Calculate the inter-class separability tKLD of feature combinations under all projection matrices i , if the maximum value is not less than the initial value tKLD0, it is considered that the elimination of this feature is effective for the overall feature combination, which increases the separability between the target and clutter samples;

[0055] In step S4, the features that meet the conditions in step S3 are eliminated. At this time, the feature dimension is reduced to θ-1, and tKLD0 is updated using the maximum separability solved in step S3, that is:

[0056] tKLD0=max{tKLD i},i=1,2,...,n

[0057] Repeat steps S1-S3, gradually reducing the feature dimension until the screening stops. The conditions for stopping the screening are as follows:

[0058] Stop screening when the target and clutter samples are different and have serious multicollinearity characteristics;

[0059] Stop screening when tKLD no longer increases;

[0060] Step S5: Label the filtered feature combinations, divide them into training set and test set in a ratio of 7:3, and input them into the intelligent detector for detection;

[0061] Step S6, initialize the parameters of the intelligent detector; construct a decision tree CART, count the number of misjudgments of clutter samples under different pruning parameters, and if the threshold is reached, output the weight factor at this time and update the decision tree. If the threshold is not reached, adjust the step direction of the weight factor according to the overall data distribution until the update is completed and the detection is completed.

[0062] The step S6 includes the following steps:

[0063] Step S61, setting the initial parameters of the decision tree model, the parameters include false alarm probability, initial weight factors of target and clutter, weight factor step value and minimum number of samples of leaf nodes, etc., wherein the minimum number of samples of leaf nodes msl ranges from [1, msl max ];

[0064] Step S62: Based on the pruning parameters of step S61, a loop is set to generate a decision tree under different msl, and the number of clutter misjudgments output by the model is measured. If it is the same as the theoretical number of misjudgments α under the set false alarm probability, the weight factor of the target and clutter at this time is output. If they are different, the loop is continued until all values ​​of msl are traversed to obtain the misjudgment number vector δ;

[0065] Step S63: If the number of clutter misjudgments that meets the condition still does not exist in step S62, then compare the value in δ with α. If most of the values ​​are greater than α, then increase the clutter weight according to the step value; otherwise, decrease the weight. Repeat step S62 until δ contains the same value as α.

[0066] Step S64, update the decision tree under the weight factors and corresponding MSL that meet the conditions, use the test set to test, complete target detection, and output the actual false alarm probability and target detection probability results.

[0067] The following is an example of target detection using the above method and the results in measured data.

[0068] Example

[0069] Using the method of the present invention, 10 groups of measured sea clutter data sets were segmented into four low-observation time series. Features were extracted from multiple dimensions, and all feature types were averaged under four polarization modes to obtain a large number of labeled target and pure sea clutter samples. The observation time after segmentation was 0.512s, 1.024s, 2.048s, and 4.096s, respectively. The number of target samples and clutter samples under each observation time is shown in Table 1:

[0070] The initial parameters of the intelligent detector are as follows: the false alarm probability is 0.001, the initial weight factors of the target and clutter are both set to 0.5, the weight factor step value is 0.05, and the minimum number of leaf node samples is in the range of [1, 20]. The value of α is calculated based on the false alarm probability and the number of clutter samples in the test set to be 6 (datasets 1, 4-10) or 7 (datasets 2 and 3).

[0071] Table 1 The number of samples in each data set under four observation durations

[0072]

[0073] The target detection experiment was conducted using the present invention. The feature screening results and tKLD changes of 10 data sets under four observation times are shown in Table 2. The eliminated features are sorted in the order of front and back, corresponding to the tKLD changes. Figure 4 The method of the present invention is compared with some current intelligent detectors with better performance, including: a detector based on three features in the time domain and frequency domain, a detector based on three features in the time and frequency domain, a detector based on feature compression, a detector based on controllable false alarm K nearest neighbor, and a clutter proportion tree detector based on a biased Gini coefficient. Under observation times of 0.512s and 1.024s, the average detection probability of the ten data sets is as follows: Figure 5 shown.

[0074] Table 2 Feature screening results

[0075]

[0076]

[0077] From the results in Table 2, it can be seen that the feature screening method of the present invention can adaptively remove redundant features according to the environment and intrinsic characteristics of the data itself, without being restricted by dimension. Figure 4 and Figure 5 The results in

[15] show that compared with traditional detection methods, the present invention has better detection performance and stronger robustness under different data sets.

[0078] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An intelligent target detection method based on variable dimensional feature screening, characterized in that: The method comprises the following steps: Step S1, extracting features from pure sea clutter and target units in the data in multiple processing domains, establishing a feature sample space, and calculating the total KL divergence tKLD of the initial combination of feature samples and the variance inflation factor VIF of each feature; Step S2: Compare the variance inflation factor in step S1 with the empirical threshold to obtain the feature index where there is severe multicollinearity between feature combinations, and take the intersection of the feature indexes selected from the target and clutter units to obtain the feature index vector where there is multicollinearity between the target and clutter units. Step S3, the features corresponding to the index vectors in step S2 are eliminated one by one; after each elimination operation, the tKLD of the remaining feature combinations is calculated. If the largest tKLD is not less than the initial value, the feature elimination is considered effective, and the feature combination completes a screening operation; Step S4, recalculate tKLD for the remaining feature combinations in step S3, and repeat steps S1-S3 until the termination condition is reached, that is, the screening of the entire feature combination is completed; Step S5: Label the filtered feature combinations, divide them into training set and test set in a ratio of 7:3, and input them into the intelligent detector for detection; Step S6, initialize the parameters of the intelligent detector; construct a classification and regression tree CART, count the number vectors of misjudgments of clutter samples under different pruning parameters, and if the set threshold is reached, output the weight factor at this time and update the decision tree; if the threshold is not reached, adjust the step direction of the weight factor according to the overall data distribution until the update is completed and the detection is completed.

2. The intelligent target detection method based on variable dimension feature screening according to claim 1, characterized in that: In step S1, multiple features are extracted from the pure sea clutter and target units in the data in the fractal, time and Doppler domains and zero-meaned. The number of features is denoted as θ; the corresponding feature screening projection matrix is ​​D, that is, D is the index matrix of θ features: D=E θ , where E is the unit matrix; After all features are extracted, a feature sample space is established. The feature sample spaces of the clutter and target are F c and F t , the total KL divergence of the initial combination of feature samples is recorded as tKLD0, that is: in, |·| is the matrix determinant operation, tr(·) is the matrix trace operation, ln(·) is the logarithm operation, and Represent the covariance matrices of clutter and target feature samples, namely: Where I0 and I1 are the number of clutter and target samples, u = 1, 2, ..., I0, v = 1, 2, ..., I1, and T is the transposed symbol; The variance inflation factor VIF of each feature is expressed as: The variance inflation factors of the clutter and target units are denoted as VIF c,p and VIF t,p , where p represents the feature number, p = 1, 2, ..., θ; is the coefficient of determination obtained by regressing feature p on all other variables.

3. The intelligent target detection method based on variable dimension feature screening according to claim 1, characterized in that: In step S2, the variance inflation factor in step S1 is compared with the empirical threshold. If VIF is greater than or equal to 10, there is severe multicollinearity. The feature numbers of the target and clutter units with severe multicollinearity are selected and the intersection is taken to obtain the vector H. 1×n , where n is the number of features.

4. The intelligent target detection method based on variable dimension feature screening according to claim 1, characterized in that: In step S3, the vector H in step S2 is 1×n The corresponding features are eliminated one by one, that is, for H 1×n For each feature number in , remove the corresponding column in the projection matrix D and get n new projection matrices: D1,...,D n At this time, the feature sample space of clutter and target is compressed and updated to Z c and Z t : Among them, i is the feature number that needs to be eliminated; Then the covariance matrix of the feature space is updated as: The inter-class separability tKLD of the remaining feature combination after removing the i-th feature i Expressed as: Among them, θ s is the number of remaining features, After each culling operation, tKLD i Compare with the initial value tKLD0. If the maximum value is not less than the initial value tKLD0, it is considered that the elimination of this feature is effective for the overall feature combination, which increases the separability between the target and clutter samples. Update the initial value tKLD0 according to the current maximum inter-class separability, that is: tKLD0=max{tKLD i },i=1,2,...,n。 5. The intelligent target detection method based on variable dimension feature screening according to claim 1, characterized in that: In step S4, the feature sample space is re-established for the remaining feature combinations in step S3 and tKLD and VIF are calculated. Steps S1-S3 are repeated, and the feature dimension is gradually reduced until the screening stops. The conditions for stopping the screening are as follows: 1) Stop screening when the target and clutter samples are different and have serious multicollinearity characteristics; 2) Stop screening when tKLD no longer increases.

Images