Space Micro-Motion Target Recognition Method Based on Stacking Ensemble Algorithm

Abstract

The present invention discloses a method for identifying spatial micro-motion targets based on the Stacking integration algorithm, which mainly solves the problems of single feature extraction in existing spatial micro-motion target identification methods and the inability of traditional single classifiers to fully exploit the classification potential of features. The implementation scheme is as follows: 1) Use the multi-transform domain feature extraction method to extract the target time domain, frequency domain, and time-frequency domain features, and generate a training sample set and a test sample set; 2) Construct a Stacking integration classifier model in which four primary classifiers are connected in parallel and then cascaded with a secondary classifier; 3) Use the training sample set and the cross-validation method to train the Stacking integration classifier; 4) Input the test sample set into the trained Stacking integration classifier to obtain the classification result. The present invention can fully characterize the target characteristics, exploit the classification potential of features, improve the recognition rate of spatial micro-motion targets, and can be used for ballistic target recognition.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar target recognition, and particularly relates to a method for identifying spatial micro-motion targets, which can be used for ballistic target recognition. Background Art

[0002] Micro-motion is a unique motion form of ballistic targets such as mid-course warheads and decoys. The echo of a micro-motion target usually contains important characteristics such as its shape, structure, and motion. Ground-based radars can obtain the echoes of long-range spatial targets such as micro-motion ballistic targets all day and all weather, and then extract micro-motion features from them to achieve classification and recognition, that is, to realize the recognition of spatial micro-motion targets. According to different feature extraction methods, the existing methods for recognizing spatial micro-motion targets can be divided into two categories: the classification and recognition method based on the radar cross-section (RCS) of the target and the classification and recognition method based on micro-Doppler feature extraction. Among them, the classification and recognition method based on RCS estimates the motion attitude, micro-motion period, and size using the RCS sequence of the spatial micro-motion target, and finally realizes the recognition of the spatial micro-motion target by designing a classifier. The recognition method based on micro-Doppler first uses methods such as Fourier transform (FFT), short-time Fourier transform (STFT), and wavelet analysis to transform the original radar echo into the transform domain, and then extracts micro-motion features and realizes classification and recognition.

[0003] The above methods can obtain good recognition effects in specific situations, but there are problems in that the target feature extraction and classifier design are relatively single. Specifically, it is difficult for a single feature to fully describe the target characteristics, which directly affects the accuracy of classification and recognition. Traditional single classifiers such as K-nearest neighbor (KNN) and support vector machine (SVM) cannot fully exploit the potential of feature classification, and the recognition performance is limited.

[0004] A.R. Persico, C. Clemine, D. Gaglione, C.V. Ilioudis, J. Cao, and L. Pallotta proposed a method for target recognition based on micro-Doppler features in their published paper "On model, algorithms, and experiment for micro-Doppler-based recognition of ballistic targets" (IEEE Transactions on Aerospace & Electronic Systems, 2017). The specific implementation is as follows: First, the short-time Fourier transform (STFT) is used on the radar echo to obtain the time-frequency distribution of the echo. Then, the fast Fourier transform (FFT) is performed on the time-frequency distribution along the frequency axis to obtain the rhythm spectrum diagram CVD of the target. Next, statistical features are extracted from CVD, and the KNN classifier is used for recognition. This method proposed new micro-Doppler features and has good recognition performance when the micro-motion forms of the targets are quite different. However, this method has the defect of single feature representation and poor recognition performance when the micro-motion forms of the targets are similar.

[0005] Choi, S., Park, M., Kim, M., Kang, K., and Kim, K. proposed to use the target RCS features and micro-Doppler features for efficient target recognition in their published paper "Efficient discrimination of ballistic targets with micro-motions" (IEEE Transactions on Aerospace & Electronic Systems, 2020). The steps of this method are as follows: First, five-dimensional features are extracted from the target echo, including the period of the RCS sequence, the peak difference value, the similarity between the RCS sequence and the sine waveform, the 3dB bandwidth of the echo spectrum, and the similarity between the micro-Doppler frequency curve and the sine waveform. Then, an improved KNN classifier is used to classify the five-dimensional features. Although this method combines the RCS features and micro-Doppler features to characterize the target characteristics, the extracted features are less, making it difficult to accurately and fully describe the characteristics of the micro-moving target. When the micro-motion forms of the targets are similar, the recognition accuracy is low. Summary of the Invention

[0006] The purpose of the present invention is to propose a method for recognizing spatial micro-moving targets based on the Stacking ensemble algorithm in view of the above deficiencies of the prior art, so as to fully exploit the feature classification potential of different classifiers and improve the recognition performance when the micro-motion forms of the targets are similar.

[0007] The technical idea of the present invention is as follows: using the target time domain, frequency domain, and time-frequency domain features extracted from radar echoes to form a 15-dimensional feature vector, and generating a training sample set and a test sample set respectively; constructing and training a two-layer Stacking ensemble model, and obtaining the recognition result of the spatial micro-motion target by inputting the test sample set into the trained model. Its implementation includes the following:

[0008] (1) Generate the standardized training sample set and test sample set:

[0009] (1a) Establish a model containing 4 spatial micro-motion targets, and obtain their narrowband echoes. Select 1200 narrowband echoes and corresponding labels observed by the radar at the elevation angle of 31 - 42° to form the original echo training sample set; select 1300 narrowband echoes and corresponding labels observed by the radar at the elevation angle of 43 - 55° to form the original echo test sample set;

[0010] (1b) Based on each narrowband echo in the original echo training sample set and the original echo test sample set, extract its time domain, frequency domain, and time-frequency domain features to obtain a training sample set and a test sample set composed of 15-dimensional feature vectors;

[0011] (1c) Standardize the training and test sample sets obtained by feature extraction column by column to obtain the standardized training sample set and test sample set;

[0012] (2) Construct a Stacking ensemble classifier:

[0013] (2a) Construct a Stacking ensemble classifier formed by cascading four primary classifiers, namely a random forest classifier, an extreme forest classifier, an Adaboost classifier, and a GBDT classifier, and then cascading with a secondary classifier of a logistic regression classifier;

[0014] (2b) Set the parameters of each classifier:

[0015] The number of subtrees of the random forest and the extreme forest are both set to 100, and the remaining parameters use the default values;

[0016] The number of subtrees of the Adaboost classifier is set to 100, the learning rate is 0.2, the splitter is set to "random", the maximum depth is set to 12, and the remaining parameters use the default values;

[0017] The maximum depth of the GBDT is set to 6, and the remaining parameters use the default values;

[0018] All parameters of the logistic regression use the default values;

[0019] (2c) Set all primary classifiers to use the class probability output mode;

[0020] (3) Input the standardized training sample set into the Stacking ensemble classifier, and use the cross-validation method for training to obtain a trained Stacking classifier;

[0021] (4) Input the standardized test sample set into the trained Stacking ensemble classifier for testing to obtain the classification results output by the Stacking ensemble classifier.

[0022] The present invention has the following advantages compared with the prior art:

[0023] 1. Since the present invention adopts a feature extraction method based on multiple transform domains to extract the time-domain, frequency-domain, and time-frequency domain features of the target echo, it can fully characterize the target characteristics;

[0024] 2. Since the present invention constructs a Stacking ensemble classifier, it can combine multiple different classifiers, fully explore the classification potential of the extracted features through different classifiers, not only can effectively improve the recognition performance, but also can effectively avoid the overfitting problem, and has good generalization performance and better anti-noise performance. Description of the Drawings

[0025] Figure 1 is the implementation flowchart of the present invention. Detailed Embodiment

[0026] The following further describes the embodiments and effects of the present invention with reference to the drawings.

[0027] Refer to Figure 1 , the implementation steps of this embodiment include the following.

[0028] Step 1, generate a training sample set and a test sample set.

[0029] 1.1) Establish four types of target models: flat-bottom cone, spherical-bottom cone, cone-column body, and spherical-bottom cone-column body. Use the standard PO method to obtain their full-angle domain static electromagnetic echoes, and then extract the static electromagnetic echoes according to the target micro-motion form to generate the dynamic electromagnetic echoes of the target. The carrier frequency of the radar is set to 10 GHz, the target pitch angle ranges from 31° to 55°, changing at intervals of 1°, the precession frequency is equally spaced at 5 values and the precession angle is equally spaced at 5 values at each pitch angle, and 625 echo samples are generated for each type of target;

[0030] 1.2) For each type of target, take the echo samples with pitch angles from 31° to 42°, and extract the time-domain, frequency-domain, and time-frequency domain features:

[0031] 1.2.1) Extract time-domain features from the target echo, including the micro-Doppler period F1, the mean value F2 of the RCS sequence, the peak difference value F3 of the RCS sequence, and the mean square error F4 between the RCS sequence and the sine curve. Among them, the micro-Doppler period is extracted by the autocorrelation function method;

[0032] 1.2.2) Perform Fourier transform on the target echo, and extract the frequency-domain features of the target according to the obtained frequency-domain signal, that is, take the modulus of the frequency-domain signal, and then calculate the interval length above the threshold as the frequency-domain micro-Doppler modulation bandwidth F5 by setting the threshold;

[0033] 1.2.3) Use short-time Fourier transform on the target echo to obtain a time-frequency diagram, and extract time-frequency domain features according to the time-frequency diagram, including the average value F6 of the micro-Doppler curve, the mean square error F7 between the micro-Doppler curve and the sine curve, and the standardized standard deviation F8 of the micro-Doppler signal intensity;

[0034] 1.2.4) Perform Fourier transform on the time-frequency diagram in 1.2.3) along the time dimension to obtain a prosody spectrum diagram CVD, calculate the mean Doppler frequency at each prosody frequency point of it, convert CVD into a one-dimensional sequence, denoted as ACVD, and standardize it. Extract time-frequency features according to the standardized ACVD, including the mean value F9, the standard deviation F 10 of the ACVD sequence, the kurtosis F 11 of the ACVD sequence, the skewness F 12 of the ACVD sequence, the peak sidelobe ratio F of the ACVD normalized autocorrelation function 13 of the ACVD sequence, the integrated sidelobe ratio 1F 14 of the ACVD sequence, and the integrated sidelobe ratio 2F 15 of the ACVD sequence;

[0035] 1.2.5) Combine the features extracted in the above steps into a 15-dimensional feature vector. A total of 1200 training sample sets composed of feature vectors are generated for these four types of targets, and then all training samples are standardized column by column and combined with the corresponding labels to obtain a standardized training sample set S1. The standardization process is carried out through the following formula:

[0036]

[0037] where, F d ∈ R N×1 is the d-th feature component of all samples, N is the number of samples, and are the statistical mean and standard deviation of F d respectively, and is the standardized feature component.

[0038] 1.3) For each type of target, take the echo samples with the elevation angle ranging from 43° to 55°, and extract time-domain, frequency-domain, and time-frequency domain features:

[0039] 1.3.1) Extract time-domain features from the target echo, including the micro-Doppler period F1, the mean value F2 of the RCS sequence, the peak difference value F3 of the RCS sequence, and the mean square error F4 between the RCS sequence and the sine curve. The micro-Doppler period is extracted using the autocorrelation function method;

[0040] 1.3.2) Perform Fourier transform on the target echo, and extract the frequency-domain features of the target according to the obtained frequency-domain signal, that is, take the modulus of the frequency-domain signal, and then calculate the interval length higher than the threshold value by setting the threshold as the frequency-domain micro-Doppler modulation bandwidth F5;

[0041] 1.3.3) Use the short-time Fourier transform on the target echo to obtain the time-frequency diagram, and extract the time-frequency domain features according to the time-frequency diagram, including the mean value F6 of the micro-Doppler curve, the mean square error F7 between the micro-Doppler curve and the sine curve, and the standardized standard deviation F8 of the micro-Doppler signal intensity;

[0042] 1.3.4) Perform Fourier transform on the time-frequency diagram in 1.3.3) along the time dimension to obtain the prosody frequency spectrum diagram CVD, calculate the mean Doppler frequency at each prosody frequency point of it, convert CVD into a one-dimensional sequence, denoted as ACVD, and standardize it. Extract time-frequency features according to the standardized ACVD, including the mean value F9 of the ACVD sequence, the standard deviation F 10 、kurtosis F 11 、skewness F 12 、the peak sidelobe ratio F of the ACVD normalized autocorrelation function 13 、integrated sidelobe ratio 1F 14 and integrated sidelobe ratio 2F 15 ;

[0043] 1.3.5) Combine the features extracted in the above steps into a 15-dimensional feature vector. These four types of targets altogether generate a test sample set containing 1300 feature vectors. Then, standardize all the test samples column by column and combine them with the corresponding labels to obtain the standardized test sample set S2, where the standardization process uses the same formula as in 1.2.5).

[0044] Step 2, construct a Stacking ensemble classifier.

[0045] 2.1) Establish a random forest classifier in the python environment, and set the parameters of the random forest classifier: the number of subtrees is set to 100, and the class probability mode is used for output, and the remaining parameters use the default values;

[0046] 2.2) Establish an extreme forest classifier in the python environment, and set the parameters of the extreme forest classifier: the number of subtrees is set to 100, and the class probability mode is used for output, and the remaining parameters use the default values;

[0047] 2.3) Establish an Adaboost classifier in the Python environment and set the Adaboost classifier parameters: the base classifier is set as a decision tree, the number of subtrees is set to 100, the learning rate is set to 0.2, the decision tree splitter parameter is set to "random", the maximum depth is set to 12, and the class probability mode is adopted for output;

[0048] 2.4) Establish a GBDT classifier in the Python environment and set the GBDT classifier parameters: the maximum depth is set to 6, the class probability mode is adopted for output, and the remaining parameters adopt the default values;

[0049] 2.5) Establish a logistic regression classifier in the Python environment, and the logistic regression classifier adopts the default parameter settings;

[0050] 2.6) Connect the four primary classifiers, namely the random forest classifier, the extreme forest classifier, the Adaboost classifier, and the GBDT classifier, in parallel as the first layer of the Stacking ensemble classifier, use the logistic regression classifier as the secondary classifier for the second layer of the Stacking ensemble classifier, and then cascade the first layer and the second layer to form the final Stacking ensemble classifier.

[0051] Step 3, train the Stacking ensemble classifier.

[0052] Input the standardized training sample set S1 into the constructed Stacking ensemble classifier, and use the K-fold cross-validation method for training to obtain the trained Stacking ensemble classifier. In the example of the present invention, K is taken as 10;

[0053] The specific steps of training are as follows:

[0054] 3.1) Adopt the 10-fold cross-validation method to randomly divide S1 into 10 equal parts;

[0055] 3.2) Each time, use 9 of them to train all the primary classifiers in the first layer, and the remaining one is used as the validation set;

[0056] After 10 times like this, each type of classifier can obtain 10 trained models and the class probability vectors of 10 validation sets;

[0057] For each type of classifier, horizontally splice the class probability vectors of the 10 obtained validation sets to obtain a data set with the same number of samples as S1;

[0058] 3.3) Horizontally splice the data sets obtained by each type of classifier to obtain a secondary training set that can be used for the training of the secondary classifier, and the labels remain unchanged;

[0059] 3.4) Input the secondary training set obtained in step 3 into the logistic regression classifier, and the trained Stacking integrated classifier can be obtained.

[0060] Step 4, output the prediction results of the test sample set.

[0061] 4.1) Input the standardized test sample set S2 into the first layer of the trained Stacking integrated classifier, so that each primary classifier obtains 10 prediction results of S2;

[0062] 4.2) Take the average of the 10 prediction results of S2 for each primary classifier to obtain a data set with the same number of samples as S2;

[0063] 4.3) Horizontally splice the data sets obtained by each primary classifier and input them into the trained secondary classifier to obtain the final prediction results, and complete the recognition of spatial micro-motion targets.

[0064] The effect of the present invention can be illustrated by the following simulation experiments.

[0065] 1. Simulation experiment conditions:

[0066] The simulation experiment of the present invention uses the standard PO method to generate radar echo data of 4 types of spatial micro-motion targets. These 4 types of targets are: flat-bottom cone, ball-bottom cone, cone-column body, and ball-bottom cone-column body, and the micro-motion form is precession. Set the radar carrier frequency to 10 GHz, the pitch angle of each type of target ranges from 31° to 55°, with a 1° interval change. At each pitch angle, the precession frequency is equally spaced at 5 values and the precession angle is equally spaced at 5 values. 625 echo samples are generated for each type of target, a total of 2500 echo samples, and Gaussian white noise is added to the echo to generate echo samples with signal-to-noise ratios of 0 dB, 5 dB, 10 dB, and 15 dB respectively. Take the echo samples with pitch angles from 31° to 42°, extract time-domain, frequency-domain, and time-frequency domain features, and perform standardization processing to obtain the standardized training sample set; take the echo samples with pitch angles from 43° to 55°, extract time-domain, frequency-domain, and time-frequency domain features, and perform standardization processing to obtain the standardized test sample set.

[0067] The simulation experiment hardware platform is an Intel(R) Core(TM) i7-6700@3.40GHz CPU and 8GB RAM;

[0068] The simulation experiment software platform is MATLAB 2016b, Python 3.6, and sklearn 0.18.1.

[0069] 2. Simulation experiment content and result analysis:

[0070] Simulation Experiment 1: For the training sample sets and test sample sets with different signal-to-noise ratios obtained by the feature extraction method of the present invention, use traditional single classifiers and classical ensemble classifiers to classify and identify the target, and calculate the recognition rates of the two methods respectively through the following formula:

[0071]

[0072] Among them, c represents the recognition rate of the test sample set, M represents the number of samples in the test sample set, h(·) represents the classification discrimination function, and t i represents the true category of the i-th test sample in the test sample set, and y i represents the output result of the classifier corresponding to the i-th test sample in the test sample set. When t i and y i are equal, h(t i , y i ) is equal to 1; otherwise, h(t i , y i ) is equal to 0.

[0073] In the classical ensemble classifier method, four classical ensemble classifiers, namely Adaboost, GBDT, random forest, and extreme forest, are used for comparison. First, use the training sample set to train the classical ensemble classifier to obtain a trained classical ensemble classifier model. Then, use the test sample set to test on the trained classical ensemble classifier model, calculate the recognition rate of the classical ensemble classifier, repeat the experiment 100 times, and take the average of the recognition rates of the 100 experiments as the final result.

[0074] In the traditional single classifier method, an SVM classifier and a decision tree classifier are used for comparison. First, use the training sample set to train the single classifier to obtain a trained single classifier. Then, use the test sample set to test on the trained single classifier, calculate the recognition rate of the traditional single classifier, repeat the experiment 100 times, and take the average of the recognition rates of the 100 experiments as the final result.

[0075] The results of the two methods under different signal-to-noise ratio data sets are compared as follows:

[0076] Table 1 Comparison of the recognition rates (%) of each classifier at each signal-to-noise ratio

[0077]

[0078] The minimum improvement refers to the difference between the lowest recognition rate among the four ensemble classifiers and the highest recognition rate among SVM and decision tree at each signal-to-noise ratio data set.

[0079] It can be seen from the simulation results that at each signal-to-noise ratio, the lowest recognition rate of the integrated classifier is 5.08% higher than the highest recognition rate of the traditional single classifier. This fully demonstrates that the integrated classifier can fully exploit the classification potential of features and improve the target recognition rate. In addition, at 15 dB, the recognition rate of the extreme forest classifier has reached 92.15%, indicating that the 15-dimensional feature vector extracted by the present invention can fully characterize the target characteristics.

[0080] Simulation Experiment 2: Using the training sample sets and test sample sets with different signal-to-noise ratios obtained by the feature extraction method of the present invention, the Stacking integrated classifier constructed by the present invention is used to classify and identify the target, and the best performance BestClassifier of the classical integrated classifiers in Table 1 at each signal-to-noise ratio and the recognition results of the best-performing SVM among the traditional single classifiers are compared.

[0081] In this method, the constructed Stacking integrated classifier is trained using the training sample set to obtain a trained Stacking integrated classifier, and then the test sample set is used to test on the trained Stacking integrated classifier to calculate the recognition rate of the Stacking integrated classifier. The experiment is repeated 100 times, and the average of the recognition rates of the 100 experiments is taken as the final result. The results are as follows:

[0082] Table 2 Comparison results of the recognition rates of each classifier at different signal-to-noise ratios

[0083]

[0084] It can be seen from Table 2 that the Stacking integrated classifier proposed by the present invention has a better recognition rate than the best result BestClassifier that combines the existing four classical integrated classifiers, and the highest recognition rate reaches 94.36%. This indicates that the spatial micro-motion target recognition method based on the Stacking integration method proposed by the present invention can fully characterize the target characteristics, and at the same time can combine the advantages of multiple classifiers, exploit the classification potential of features, and effectively improve the recognition rate of spatial micro-motion targets. In addition, it can also be seen that compared with the SVM, the recognition rates of the Stacking integrated classifier of the present invention are increased by 20.99%, 12.53%, 10.14%, and 9.44% respectively from 0 dB to 15 dB, and the lowest recognition rate also reaches 85.61%, indicating that the method proposed by the present invention has high anti-noise performance.

[0085] Combined with Simulation Experiments 1 and 2, it shows that the spatial micro-motion target recognition method based on the Stacking integration method proposed by the present invention can fully characterize the target characteristics, combine the advantages of multiple classifiers, tap the feature classification potential, effectively improve the recognition rate of spatial micro-motion targets, and has high anti-noise performance, which has important theoretical significance and application value.

Claims

1. A method for identifying spatial micro-motion targets based on the Stacking integration algorithm, characterized in that Including: (1) Generate a standardized training sample set and a test sample set: (1a) Establish a model containing 4 space micro-motion targets, obtain their narrowband echoes, select 1200 narrowband echoes observed by the radar at elevation angles of 31 - 42° and their corresponding labels to form the original echo training sample set; select 1300 narrowband echoes observed by the radar at elevation angles of 43 - 55° and their corresponding labels to form the original echo test sample set; (1b) Based on each narrowband echo in the original echo training sample set and the original echo test sample set, extract its time-domain, frequency-domain, and time-frequency-domain features to obtain a training sample set and a test sample set composed of 15-dimensional feature vectors, as follows: (1b1) Extract time-domain features from the narrowband echo, including the micro-Doppler period F1, the mean value of the RCS sequence F2, the peak difference value of the RCS sequence F3, and the mean square error between the RCS sequence and the sine curve F4; (1b2) Perform Fourier transform on the narrowband echo to extract the frequency-domain features of the target, that is, by setting a threshold and calculating the interval length higher than the threshold value as the frequency-domain micro-Doppler modulation bandwidth F5; (1b3) Use the short-time Fourier transform on the narrowband echo to obtain the time-frequency diagram, and extract the time-frequency-domain features from the time-frequency diagram, including the average value of the micro-Doppler curve F6, the mean square error between the micro-Doppler curve and the sine curve F7, and the standardized standard deviation of the micro-Doppler signal intensity F8; (1b4) Perform a Fourier transform on the time-frequency diagram obtained in (1b3) along the time dimension to obtain the prosody frequency spectrum diagram CVD. Calculate the mean Doppler frequency at each prosody frequency point of it, convert CVD into a one-dimensional sequence, denoted as ACVD, and standardize it. Extract time-frequency features based on the standardized ACVD, including the mean F9 of the ACVD sequence, the standard deviation F 10 , the kurtosis F 11 , the skewness F 12 , the peak sidelobe ratio F of the ACVD normalized autocorrelation function 13 , the integrated sidelobe ratio 1F 14 and the integrated sidelobe ratio 2F 15 ; (1b5) Extract the features shown in (1b1), (1b2), (1b3), and (1b4) according to each echo sample in the original echo training sample set and the original echo test sample set to form a 15-dimensional feature vector, and finally obtain a training sample set and a test sample set composed of feature vectors; (1c) Standardize the training and test sample sets obtained by feature extraction column by column to obtain a standardized training sample set and a test sample set; (2) Construct a Stacking ensemble classifier: (2a) Construct a Stacking ensemble classifier composed of four primary classifiers, namely a random forest classifier, an extreme forest classifier, an Adaboost classifier, and a GBDT classifier, connected in parallel and then cascaded with a secondary classifier of a logistic regression classification; (2b) Set the parameters of each classifier: The number of subtrees of the random forest and the extreme forest are both set to 100, and the remaining parameters use the default values; The number of subtrees of the Adaboost classifier is set to 100, the learning rate is 0.2, the splitter is set to "random", the maximum depth is set to 12, and the remaining parameters use the default values; The maximum depth of the GBDT is set to 6, and the remaining parameters use the default values; All parameters of the logistic regression use the default values; (2c) Set all primary classifiers to use the class probability output mode; (3) Input the standardized training sample set into the Stacking ensemble classifier and use the cross-validation method for training to obtain a trained Stacking classifier; (4) Input the standardized test sample set into the trained Stacking ensemble classifier for testing, and obtain the classification result output by the Stacking ensemble classifier.

2. The method according to claim 1, wherein (1c) the training and test sample sets obtained by feature extraction are standardized by columns using the following formula: Among them, F d ∈R N×1 is the d-th feature component of all samples, and N is the number of samples. and are the statistical mean and standard deviation of F d respectively, and is the standardized feature component.

3. The method according to claim 1, wherein, (2) The Stacking ensemble classifier adopts a two-layer structure, that is, the four primary classifiers, random forest, extreme forest, Adaboost, and GBDT, constitute the first layer of the Stacking ensemble classifier, and the logistic regression classifier, as a secondary classifier, constitutes the second layer of the Stacking.

4. The method according to claim 1, wherein in (3), the standardized training sample set is input into the Stacking ensemble classifier and trained using a cross-validation method, which is implemented as follows: (3a) Use K-fold cross-validation method to train the primary classifier of the first layer: The training sample set is randomly shuffled and divided into K parts. K-1 parts are used to train the four primary classifiers each time, and the remaining part is used as a validation set. After K rounds, each primary classifier obtains K trained models and K validation set prediction results; The K validation set prediction results of each primary classifier are vertically spliced ​​to obtain a data set with the same number of samples as the training sample set; The data sets obtained by each primary classifier are horizontally spliced, with the labels unchanged, to obtain the secondary training set for training the second-level classifier; (3b) The secondary training set obtained in the first layer is input into the secondary classifier of the second layer for training, and a trained secondary classifier can be obtained. At this point, the training of the Stacking ensemble classifier is completed.

5. The method according to claim 1, wherein in (4), the standardized test sample set is input into the trained Stacking ensemble classifier for testing to obtain the classification result output by the Stacking ensemble classifier, which is implemented as follows: (4a) Input the standardized test sample set into the first layer of the trained Stacking ensemble classifier so that each primary classifier obtains K test sample set prediction results, and then averages the K test set prediction results to obtain a data set with the same number of samples as the test sample set; (4b) horizontally splicing the data sets obtained by each primary classifier without changing the labels, so as to obtain the secondary test set for testing the second-level classifier; (4c) The secondary test set obtained in the first layer is input into the trained secondary classifier in the second layer for testing, and the classification result output by the Stacking ensemble classifier is obtained.

Images