A radar multiple echo classification method based on the combination of oversampling and ensemble learning

Abstract

The present invention relates to a radar multiple echo classification method based on the combination of oversampling and ensemble learning, and relates to the technical field of radar signal processing. The method of the present invention extracts the amplitude and phase information in the radar echo data, and uses ADASYN adaptive synthetic oversampling in the data preprocessing stage to balance the number of samples of various categories of radar multiple echoes, so as to improve the recognition rate of the subsequent established classifier for the echo categories with fewer samples. The random forest algorithm in ensemble learning is selected for classification to avoid overfitting. Applying this classification method to radar signal processing can effectively distinguish the multiple echoes received by the radar, and further screen out the target traces of interest to the current radar, improving the tracking performance of radar data processing.

Description

Technical Field

[0001] The present invention relates to the technical field of radar signal processing, and particularly relates to a radar multiple echo classification method based on oversampling and ensemble learning. Background Art

[0002] A radar irradiates a target by transmitting electromagnetic waves and receives its echo to achieve target detection and ranging. Among them, a pulsed radar periodically transmits signals and receives echoes, and transmits multiple pulses into space at a certain pulse repetition period. During the reception of each pulse signal, the echo signal reflected by the target in space is received. When the distance from the target to the radar is relatively far, exceeding the maximum detection distance of the radar, that is, the maximum unambiguous distance, the echo signal reflected by the target is not received during the current pulse reception signal, but is received during the reception signal of the subsequent transmitted pulse; if the distance from the target to the radar does not exceed the maximum unambiguous distance of the radar, then the echo signal reflected by the target can be received during the current pulse reception signal. When the radar transmits the first pulse, if the echo signal reflected by a certain target in space is received during the reception signal of the first pulse, it is called a first echo; if the echo signal reflected by a certain target in space is received during the reception signal of the second pulse, it is called a second echo; if the echo signal reflected by a certain target in space is received during the reception signal of the third pulse, it is called a third echo. And so on, if the echo signal reflected by a certain target in space is received during the reception signal of the Nth pulse, it is called an Nth echo. The multiple echoes received by the radar will be detected as target traces in the subsequent target detection process, and these traces are not of interest to the current detection radar, resulting in trace redundancy and increasing the amount of computation, which brings difficulties to the subsequent radar data processing work. Especially when the radar detects the sea, due to the atmospheric duct effect, a large number of multiple echoes will appear on the sea surface, seriously interfering with the normal operation of the radar. Therefore, it is of great significance to study the classification recognition and suppression methods of multiple echoes.

[0003] Most of the existing methods classify radar multiple echoes by comparing the empirical threshold and the logarithmic amplitude difference between adjacent pulses. When the logarithmic amplitude difference between adjacent pulses is greater than the empirical threshold, it is determined as a multiple echo. However, in practical applications, the target categories are complex and diverse, and the amplitude fluctuations between adjacent pulses are relatively large. The threshold value used for multiple echo classification is usually difficult to determine, resulting in a low accuracy rate for classifying multiple echoes by the empirical threshold. It cannot significantly improve the problem of excessive computation, and there are still a large number of traces with poor quality in the detected traces, which is not conducive to improving the performance of radar data processing. Summary of the Invention

[0004] Technical Problems to be Solved

[0005] To solve the problem that only the amplitude information after pulse compression is used to classify multiple echoes in the existing method, resulting in too many false traces and still relatively difficult subsequent data processing. The present invention provides a radar multiple echo classification method that additionally incorporates phase information on the basis of amplitude information and combines oversampling with an ensemble learning algorithm, thereby improving the accuracy of classifying radar multiple echoes.

[0006] Technical solution

[0007] A radar multiple echo classification method based on the combination of oversampling and ensemble learning, characterized by the following steps:

[0008] S1: First, receive the radar multi-pulse signal echo through a radar receiver, and after performing preprocessing operations such as channel compensation, beam synthesis, and pulse compression on the radar multi-pulse signal echo, obtain the radar multi-pulse echo complex data X N×L = [x1, x2,... x l ,... x L N×L ; where N represents the number of pulses received by the radar, L represents the number of range cells included in the radar data, represents the column vector composed of N pulse echoes of the l-th range cell in the radar echo;

[0009] S2: Process the radar multi-pulse echo complex data X N×L = [x1, x2,..., x L N×L ; through a radar target detection method to obtain the radar detection trace data y M×1 = [y1, y2,... y m ,... y M M×1 , which contains the traces of the target reflections of the first echo and the traces of the target reflections of the multiple echoes; where M represents the number of radar detection traces, y m , m = 1, 2,..., M represents the range cell number where the m-th radar trace is located;

[0010] S3: According to the range cell where the radar detection trace data y M×1 = [y1, y2,... y M M×1 is located, extract the radar echo complex data of N pulses at the corresponding range cell from the radar multi-pulse echo complex data X N×L = [x1, x2,... x L N×L in S1, and convert it into logarithmic amplitude data S' N×M = [s'1, s'2,... s' N N×M and phase data S″​​​​​​N×M = [s″1, s″2, …, s″ N N×M , merge S′ N×M and S″ N×M and then transpose and store as S M×2N = [S′ N×M , S″ N×M T ; Determine the track category k of the radar track through empirical threshold decision and manual auxiliary judgment M×1 = [k1, k2, … k m , … k M M×1 , where k m , m = 1, 2, …, M is a value between 1 and N, representing the current track category, also known as the sample label;

[0011] S4: Design an oversampling method to preprocess the logarithmic amplitude and phase data of N pulses of the radar detection track according to their track categories, balance the number of samples in each category, and obtain the logarithmic amplitude data and phase data after data preprocessing and their corresponding category labels where M′ is the total number of track samples after sampling processing;

[0012] S5: Design an ensemble learning algorithm to obtain a multi-classifier, with the input being the logarithmic amplitude and phase data of N pulses of the radar detection track after oversampling processing and the output being the category k to which the radar detection track belongs M′×1 = [k1, k2, …, k M′ M′×1 ;

[0013] S6: Use the logarithmic amplitude and phase data of N pulses at the corresponding range cell of the radar track after oversampling processing in S4 and the sample labels to train the multi-classifier obtained in S5, and verify the performance of the classifier through the test set.

[0014] A further technical solution of the present invention: The oversampling method described in S4 selects the "ADASYN" method, that is, adaptive synthetic oversampling. The basic idea is to use a weighted distribution for samples of different minority classes according to the different learning difficulties, and generate more synthetic data for samples of minority echoes that are difficult to learn.

[0015] A further technical solution of the present invention: The ensemble learning algorithm described in S5 selects "random forest". In the present invention, the random forest contains 50 decision tree classifiers, and finally the "voting method" is used to integrate the output results;

[0016] ​​​​In the designed random forest algorithm, the method for generating the training data set of each decision tree is the Bootstrap method. In the present invention, the sample set is the logarithmic amplitude data and phase data of N pulses at the corresponding range cell of the radar detection points after sampling preprocessing. Resample from the training sample set S using the Bootstrap method with replacement. Each time, m < M training samples are taken out to obtain 50 sample subsets. That is, the training data set of each tree is different and contains duplicate training samples.

[0017] The designed random forest algorithm further introduces random attribute selection in the training process of the decision tree on the basis of constructing a bagging method integration with the decision tree as the base learner: Instead of using all features to build a single decision tree, the features are also sampled. Each time, n < 2N features are drawn to train a decision tree.

[0018] A further technical solution of the present invention: S6 is specifically as follows:

[0019] S61: The logarithmic amplitude data and phase data of N pulses at the corresponding range cell of the radar point trace data after oversampling preprocessing in S4 Are randomly divided into a training data set And a test data set In two parts along the range dimension according to the ratio of p:q; where p + q = 1 and p > q. Q = M - P, P is the number of the training data set, and Q is the number of the test data set. Denotes rounding down.

[0020] S62: According to the logarithmic amplitude data and phase data of N pulses at the corresponding range cell of the radar detection points after preprocessing at S4, use the empirical threshold and manual auxiliary judgment to determine the sample labels corresponding to the detection points, and obtain the sample label vector Corresponding to the training data set And the sample label vector Corresponding to the test data set Where Represents the sample class label of the detection point corresponding to the p-th sample in the training data set. If the point trace belongs to the target point trace of the first echo reflection, the sample label is 1. If the point trace belongs to the target point trace of the second echo reflection, the sample label is 2, and so on. If the point trace belongs to the target point trace of the N-th echo reflection, the sample label is Represents the sample class label of the detection point corresponding to the q-th sample in the test data set.

[0021] S63: Use the training data set Divided in S61 and the sample class labels corresponding to the training data set obtained in S62 Input the training data set into the random forest multi-classifier, and the output is the corresponding sample class label. In the present invention, the random forest contains 50 decision tree classifiers, and finally the "voting method" is used to integrate the output results;

[0022] S64: Use the test data set divided in S61 and the sample class labels corresponding to the test data set obtained in S62 to test the random forest multi-classifier obtained in S63, calculate the accuracy rate and loss rate of this classifier on the test set, and optimize the network model according to the test results;

[0023] S65: Use the optimized multi-classifier combining oversampling and random forest algorithm for radar multiple echo classification and discrimination.

[0024] A computer system, characterized in that it includes: one or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors are caused to implement the above method.

[0025] A computer-readable storage medium, characterized in that it stores computer-executable instructions, and the instructions are used to implement the above method when executed.

[0026] Beneficial effects

[0027] A method for classifying radar multiple echoes based on the combination of oversampling and ensemble learning provided by the present invention has the following beneficial effects:

[0028] 1. On the basis that the existing method only uses the amplitude information after pulse compression to classify radar multiple echoes, the present invention introduces phase information, so that the designed classification algorithm can learn more feature information from the input data, thereby further improving the classification accuracy of radar multiple echoes

[0029] 2. The present invention is a method for classifying radar multiple echoes based on the combination of oversampling and ensemble learning. For the first time, the oversampling data preprocessing method in machine learning is combined with the ensemble learning classification algorithm and applied to the classification of radar multiple echoes. After preprocessing the data using the adaptive synthetic oversampling method, the ensemble learning classification algorithm of random forest is used to construct multiple independent decision trees, and the classification results are integrated through the voting method to classify the radar multiple echoes. Compared with the existing radar multiple echo classification method based on empirical threshold discrimination, the recognition rate of echo categories with a relatively small proportion in radar multiple echoes is improved, thereby enhancing the classification accuracy of radar multiple echoes, and it does not depend on empirical thresholds. The present invention is an important processing link in the field of radar data processing, which can improve the detection performance while reducing the computational complexity of radar target detection algorithms.

[0030] 3. The oversampling method for data preprocessing proposed in the present invention adopts "ADASYN" adaptive synthetic oversampling, which balances the number of various categories of radar multiple echoes in the sample set, so that the subsequent constructed classifier will not be overly biased towards the majority category, thereby further improving the classification accuracy of echo categories with a relatively small proportion in the original sample set.

[0031] 4. The ensemble learning method proposed in the present invention is random forest. Random forest has a fast training speed and strong anti-overfitting ability. The double randomness of random forest makes it have good anti-interference ability, the variance of the trained model is small, and the generalization ability is strong. It can process high-dimensional data and does not require feature selection.

[0032] 5. The method for classifying radar multiple echoes based on the combination of oversampling and ensemble learning obtained by the present invention can be applied to radar signal processing. It can effectively suppress radar multiple echoes without setting empirical thresholds, improve the quality of radar detection points, and further enhance the performance of subsequent radar data processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The drawings are only for the purpose of showing specific embodiments and are not considered as limitations of the present invention. Throughout the drawings, the same reference numerals represent the same components.

[0034] Figure 1 is a schematic diagram of the implementation process of the present invention;

[0035] Figure 2 is a schematic diagram of the random forest classification algorithm process adopted by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0036] To make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0037] See Figure 1 , the present invention is a radar multiple echo classification method based on the combination of oversampling and ensemble learning, including the following methods:

[0038] S1: First, the radar multi-pulse signal echo is received by a radar receiver, and after preprocessing operations such as channel compensation, beam synthesis, and pulse compression on the radar multi-pulse signal echo, the radar multi-pulse echo complex data X N×L = [x1, x2,..., x L N×L ; where N represents the number of pulses received by the radar, N represents the number of range cells included in the radar data, represents the column vector composed of N pulse echoes of the l-th range cell in the radar echo;

[0039] S2: The radar multi-pulse echo complex data X N×L = [x1, x2,..., x L N×L is processed through a radar target detection method to obtain the radar detection point data y M×1 = [y1, y2,..., y M M×1 , which contains the points of the first echo target reflection and the points of the multiple echo target reflection; where M represents the number of radar detection points, y m , m = 1, 2,..., M represents the range cell number where the m-th radar point is located;

[0040] S3: According to the range cell where the radar detection point data y M×1 = [y1, y2,..., y M M×1 is located, the radar echo complex data of N pulses at the corresponding range cell is extracted from the radar multi-pulse echo complex data X N×L = [x1, x2,..., x L N×L in S1, and is converted into logarithmic amplitude data S' N×M = [s'1, s'2,..., s' N N×M and phase data S'' N×M = [s''1, s''2,..., s'' N ​N×M Merge S′ N×M with S″ N×M and then transpose and store it as S M×2N = [S′ N×M , S″ N×M T . Determine the track category k of the radar echo through empirical threshold decision and manual assistance M×1 = [k1, k2, …, k M M×1 , where k m , m = 1, 2, …, M is a value between 1 and N, representing the current track category, also known as the sample label;

[0041] S4: Design an oversampling method to preprocess the logarithmic amplitude and phase data of N pulses of the radar detection echo according to their track categories, balance the number of samples in each category, and obtain the logarithmic amplitude data and phase data after data preprocessing and their corresponding category labels where M′ is the total number of track samples after sampling processing;

[0042] S5: Design an ensemble learning algorithm, with the input being the logarithmic amplitude and phase data of N pulses of the radar detection echo after oversampling processing and the output being the category k to which the radar detection echo belongs M′×1 = [k1, k2, …, k M′ M′×1 ;

[0043] S6: Use the logarithmic amplitude and phase data of N pulses at the corresponding range cell of the radar echo after oversampling processing in S4 and the sample labels to train the multi-classifier obtained in S5 and verify the performance of the classifier through the test set;

[0044] S7: According to a radar multiple echo classification method based on the combination of oversampling and ensemble learning obtained in S6, applying this classification method to radar signal processing can effectively distinguish the multiple echoes received by the radar, further screen out the required target echoes, and improve the tracking performance of radar data processing.

[0045] The oversampling method described in S4 preprocesses the logarithmic amplitude and phase data of N pulses of the radar detection echo, balances the number of samples of radar multiple echoes, and obtains the logarithmic amplitude data and phase data after data preprocessing also includes the following methods:

[0046] ​​​S01: The method of designed oversampling selects the "ADASYN" method, that is, adaptive synthetic oversampling. The basic idea is to use a weighted distribution for samples of different minority classes according to different learning difficulties, and generate more synthetic data for samples of minority class echoes that are difficult to learn. The ADASYN method improves the learning of data distribution in two ways: (1) reducing the bias introduced by class imbalance, and (2) adaptively shifting the classification decision boundary to difficult samples. Thereby improving the recognition ability for minority class echoes.

[0047] S02: According to the logarithmic amplitude data and phase data of N pulses at the distance unit corresponding to the radar detection trace in S3, use an empirical threshold and manual assistance to judge the category corresponding to the detection trace. Using the ADASYN oversampling method for data preprocessing means artificially synthesizing new training samples for the trace category with a relatively small proportion, and obtaining the balanced logarithmic amplitude data and phase data and their corresponding class labels Thereby improving the classification accuracy of the subsequent trained multi-classifier for the echo classes with a small number in the original sample set.

[0048] The designed ensemble learning classification algorithm described in S5 also includes the following methods:

[0049] S001: The designed ensemble learning classification algorithm selects "random forest". As the name implies, it is to build a forest in a random way, consisting of many decision trees, and there is no association between the decision trees. In the present invention, fifty decision trees are trained, and finally through the "voting method", that is, the principle of the minority obeying the majority, the classification results of each decision tree are integrated;

[0050] S002: The method of generating the training data set for each decision tree in the designed random forest algorithm is the Bootstrap method. In the present invention, the sample set is the logarithmic amplitude data and phase data of N pulses at the distance unit corresponding to the radar detection trace after sampling preprocessing Resample with replacement from the training sample set S using the Bootstrap method, and take out (m < M) training samples each time sampling, obtaining 50 sample subsets, that is, the training data set for each tree is different and contains repeated training samples;

[0051] S003: The designed random forest algorithm further introduces random attribute selection in the training process of the decision tree on the basis of constructing a bagging method integration with the decision tree as the base learner: not using all features to build a single decision tree, but also sampling the features, and extracting (n < 2N) features each time to train a decision tree.

[0052] The logarithmic amplitude and phase data of N pulses at the distance unit corresponding to the radar trace after oversampling processing in S4 in S6 described above and sample labels The multi-classifier obtained by training S5, and the performance of the classifier is verified through the test set; The following methods are also included:

[0053] S0001: The logarithmic amplitude data and phase data of N pulses at the corresponding range cell of the radar echo data after oversampling preprocessing in S4 are randomly divided into a training data set and a test data set along the range dimension according to the ratio of p:q; where p + q = 1, p > q, Q = M - P, P is the number of the training data set, Q is the number of the test data set, represents rounding down;

[0054] S0002: According to the logarithmic amplitude data and phase data of N pulses at the corresponding range cell of the radar detection echo preprocessed in S4, the sample labels corresponding to the detection echoes are obtained by using the empirical threshold and manual assistance judgment, and the sample label vectors corresponding to the training data set and the sample label vectors corresponding to the test data set are obtained; where represents the sample category label of the detection echo corresponding to the p-th sample in the training data set. If the echo belongs to the target echo of the first echo reflection, the sample label is 1. If the echo belongs to the target echo of the second echo reflection, the sample label is 2, and so on. If the echo belongs to the target echo of the N-th echo reflection, the sample label is represents the sample category label of the detection echo corresponding to the q-th sample in the test data set;

[0055] S0003: Using the training data set divided by S0001 and the sample category labels corresponding to the training data set obtained by S0002 input the training data set into the random forest multi-classifier, and the output is the corresponding sample category label. In the present invention, the random forest contains 50 decision tree classifiers, and finally the "voting method" is used to integrate the output results;

[0056] S0004: Using the test data set divided by S0001 and the sample category labels corresponding to the test data set obtained by S0002 test the random forest multi-classifier obtained by S0003, calculate the accuracy rate and loss rate of the classifier on the test set, and optimize the network model according to the test results.

[0057] S0005: Use the optimized oversampling combined with the multi-classifier of the random forest algorithm for the classification and discrimination of multiple radar echoes.

[0058] As described above, it is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art in the technical field disclosed by the present invention can easily think of various equivalent modifications or substitutions, and these modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. A radar multiple echo classification method based on the combination of oversampling and ensemble learning, characterized in that The steps are as follows: S1: First, the radar multi-pulse signal echo is received by the radar receiver. After performing preprocessing operations such as channel compensation, beam synthesis, and pulse compression on the radar multi-pulse signal echo, the radar multi-pulse echo complex data X is obtained. N×L = [x1, x2, … x l , … x L N×L ; where N represents the number of pulses received by the radar, L represents the number of range cells contained in the radar data, represents the column vector composed of N pulse echoes of the l-th range cell in the radar echo.​ S2: Process the complex data X of radar multi-pulse echoes through a radar target detection method N×L =[x1,x2,…,x L N×L ; and obtain the radar detection point data y M×1 =[y1,y2,…y m ,…y M M×1 , which contains the point traces of the first echo target reflection and the point traces of the multiple echo target reflections; where M represents the number of radar detection point traces, and y m ,m = 1,2,…,M represents the range cell number where the mth radar point trace is located;​​ S3: According to the radar detection point data y M×1 = [y1, y2, …, y M M×1 in the corresponding range cell, extract the radar echo complex data of N pulses at the corresponding range cell from the radar multi-pulse echo complex data X N×L = [x1, x2, …, x L N×L in S1, and convert it into logarithmic amplitude data S' N×M = [s'1, s'2, …, s' N N×M and phase data S'' N×M = [s''1, s''2, …, s'' N N×M . Combine S' N×M and S'' N×M , and then transpose and store them as S M×2N = [S' N×M , S'' N×M T ;​​​​​ Determine the track category k of the radar track through empirical threshold decision and manual assistance M×1 =[[k1,k2,…k m ,…k M M×1 , where k m ,m = 1,2,…,M is a value between 1 and N, representing the current track category, also known as the sample label;​ S4: Design an oversampling method to preprocess the logarithmic amplitude and phase data of N pulses of radar detection points according to their respective point - track categories, balance the number of samples in each category, and obtain the logarithmic amplitude data and phase data after data preprocessing and their corresponding class labels where M' is the total number of point - track samples after sampling processing; S5: Design an ensemble learning algorithm to obtain a multi-classifier. The input is the logarithmic amplitude and phase data of N pulses of the radar detection points after oversampling processing The output is the category k to which the radar detection points belong M′×1 =[k1, k2, …, k M′ M′×1 ;​ S6: Using the logarithmic amplitude and phase data of N pulses at the range cell corresponding to the radar traces after the oversampling process in S4 and the sample labels Train the multi-classifier obtained in S5 and verify the performance of the classifier through the test set.

2. The radar multiple echo classification method based on the combination of oversampling and ensemble learning according to claim 1, wherein: The oversampling method described in S4 selects the "ADASYN" method, that is, adaptive synthetic oversampling. The basic idea is to use a weighted distribution for different minority-class samples according to the different learning difficulties, and generate more synthetic data for the samples of minority-class echoes that are difficult to learn.

3. The radar multiple echo classification method based on the combination of oversampling and ensemble learning according to claim 1, characterized in that: The ensemble learning algorithm described in S5 selects "random forest". In the present invention, the random forest contains 50 decision tree classifiers, and finally the "voting method" is used to integrate the output results; In the designed random forest algorithm, the method for generating the training data set of each decision tree is the Bootstrap method. In the present invention, the sample set is the logarithmic amplitude data and phase data of N pulses at the corresponding range cell of the radar detection points after sampling preprocessing. Resample from the training sample set S using the Bootstrap method with replacement. Each time, m < M training samples are taken out to obtain 50 sample subsets. That is, the training data set of each tree is different and contains repeated training samples. The designed random forest algorithm further introduces random attribute selection in the training process of decision trees on the basis of constructing a bagging method integration with decision trees as the base learners: instead of using all features to build a single decision tree, the features are also sampled, and each time n < 2N features are selected to train a decision tree.

4. The radar multiple echo classification method based on the combination of oversampling and ensemble learning according to claim 1, characterized in that: S6 is specifically as follows: S61: The logarithmic amplitude data and phase data of N pulses at the corresponding range cells of the radar echo data after the over-sampling preprocessing in S4 are randomly divided into a training data set and a test data set along the range dimension according to the ratio of p:q; where p + q = 1, p > q, Q = M - P, P is the number of the training data set, Q is the number of the test data set, denotes rounding down; S62: According to the logarithmic amplitude data and phase data of N pulses at the distance unit corresponding to the preprocessed radar detection points at S4, use the empirical threshold and manual auxiliary judgment to obtain the sample labels corresponding to the detection points, and obtain the training data set The corresponding sample label vector and the test data set The corresponding sample label vector where represents the sample class label of the detection point corresponding to the p-th sample in the training data set. If the point belongs to the target point of the first echo reflection, the sample label is 1. If the point belongs to the target point of the second echo reflection, the sample label is 2, and so on. If the point belongs to the target point of the N-th echo reflection, the sample label is N; represents the sample class label of the detection point corresponding to the q-th sample in the test data set; S63: Training data set divided by S61 and sample class labels corresponding to the training data set obtained by S62 Input the training data set into the random forest multi-classifier, and the output is the corresponding sample class label. In the present invention, the random forest contains 50 decision tree classifiers, and finally the "voting method" is used to integrate the output results; S64: The test data set divided by S61 and the sample class labels corresponding to the test data set obtained by S62 Test the random forest multi-classifier obtained by S63, calculate the accuracy rate and loss rate of the classifier on the test set, and optimize the network model according to the test results; S65: Use the optimized oversampling combined with the multi-classifier of the random forest algorithm for the classification and discrimination of radar multiple echoes.

5. A computer system, characterized in that Including: One or more processors, a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in claim 1.

6. A computer-readable storage medium, characterized in that Stored with computer-executable instructions, the instructions are used to implement the method described in claim 1 when executed.

Images