Hrrp target recognition method based on joint feature extraction and stacking model classification

Abstract

The application discloses a HRRP target recognition method based on joint feature extraction and stacking model classification, and comprises the following steps: step 1, acquiring HRRP data, taking HRRP data modulo and performing normalization processing; step 2, extracting non-parametric features of normalized HRRP data; step 3, inputting the normalized HRRP data into a trained 1D-CNN model to obtain deep convolution features of the HRRP data; and step 4, fusing several non-parametric features of the normalized HRRP and ConvHRRP features, using a stacking integrated model to complete classification, and obtaining HRRP target recognition results. The application combines various non-parametric features of HRRP to perform feature-level fusion, thereby obtaining more comprehensive feature representation, adopts a stacking integrated model to perform decision-level fusion on multiple basic classifiers, and further improves the recognition accuracy and robustness of the model. The application jointly uses HRRP parametric features extracted by the 1D-CNN and various HRRP non-parametric features, and combines a stacking integrated classification method, thereby improving the HRRP target recognition accuracy.

Description

Technical Field

[0001] This invention relates to the field of radar target recognition technology, and in particular to the HRRP target recognition method based on joint feature extraction and Stacking model classification. Background Technology

[0002] With the development of radar technology, high-resolution bandwidth radar can extract high-resolution range profiles (HRRPs) of targets, and HRRP radar target recognition methods have always been one of the research hotspots in automatic radar target recognition. Among these methods, extracting non-parametric features from the HRRP sequence and feeding them into a classifier to distinguish unknown targets is a classic and effective approach. This method does not require constructing a mathematical model containing the parameters to be estimated; instead, it directly calculates the statistical features of the HRRP in the time and frequency domains. These non-parametric features are relatively simple to calculate, highly interpretable, clearly defined, and robust. However, they rely heavily on the designer's experience in judging feature quality and data understanding, resulting in poor flexibility and limited separability in classification tasks, which has encountered bottlenecks in recent years. Meanwhile, deep learning methods are increasingly widely used in target recognition, and deep artificial neural networks can perform automatic feature extraction and classification tasks.

[0003] Convolutional Neural Networks (CNNs), as a deep learning model, are widely used in pattern recognition. Traditional CNNs are mostly used for recognition and regression of two-dimensional data. One-dimensional Convolutional Neural Networks (1D-CNNs) are a variation of traditional CNNs, mainly used for feature extraction and classification tasks on one-dimensional data with certain structural characteristics. It treats the input data as a one-dimensional matrix, uses convolution operations to extract local features in the temporal or spatial dimensions, compresses and reduces the dimensionality of these features through pooling layers, and finally performs classification through fully connected layers. The deep abstract features of HRRP extracted using 1D-CNNs are actually feature vectors obtained after pooling and fully connected layers, derived from the convolution kernel weight parameters extracted from HRRP. (Hereinafter referred to as ConvHRRP features).

[0004] ConvHRRP features and non-parametric features of HRRP are highly complementary: ConvHRRP features do not require manual extraction and can reflect local features and structural relationships of HRRP, but 1D-CNN requires a large amount of data, has too many parameters, and has high training costs during training; non-parametric features of HRRP can reflect global features and statistical regularities, are simple to compute, easy to interpret, and have strong robustness, but have limited separability, rely on experience, and lack flexibility. The combined use of these two types of features can obtain a more comprehensive and accurate HRRP feature representation.

[0005] With the rapid development of machine learning, various classification and regression models are now available for automated radar target identification. Choosing a model with high accuracy and robustness is crucial to the success of radar target identification tasks. Wolpert first proposed the Stacking ensemble model algorithm in 1992. The Stacking method organizes the results of multiple base classifiers, fully leveraging the strengths of each classifier to achieve stronger classification performance.

[0006] However, in current research and practice, the Stacking method is relatively rarely used in the field of radar target recognition. Summary of the Invention

[0007] 1. Technical problems to be solved

[0008] The purpose of this invention is to address the relatively limited application of the Stacking method in the field of radar target recognition in existing technologies, and to propose an HRRP target recognition method based on joint feature extraction and Stacking model classification.

[0009] 2. Technical Solution

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] The HRRP target recognition method based on joint feature extraction and Stacking model classification includes the following steps:

[0012] Step 1: Obtain HRRP data, perform modulo and normalization on the HRRP data to limit the values ​​to between 0 and 1, thereby eliminating the amplitude sensitivity of the HRRP data;

[0013] Step 2: Extract non-parametric features from the normalized HRRP data, including power spectrum, center moment, echo energy, normalized standard deviation, etc.

[0014] Step 3: Input the normalized HRRP data into the trained 1D-CNN model, and use the output of the first fully connected layer in the network as feature data to obtain the deep convolutional features of the HRRP data, hereinafter referred to as ConvHRRP;

[0015] Step 4: Fuse the normalized non-parametric features of HRRP with the ConvHRRP features, and input the joint feature data into the trained Stacking ensemble model to obtain the HRRP target recognition result.

[0016] Preferably, the process of training the 1D-CNN model and extracting ConvHRRP features in step 3 is as follows:

[0017] S301: Modulo and normalize the HRRP data used for training to limit the values ​​to between 0 and 1;

[0018] S302: Input the normalized HRRP training data into the 1D-CNN network model, use 'sgdm' as the optimizer, set the maximum number of training epochs to 100, the size of each mini-batch to 16, and the initial learning rate to 0.001 to obtain the trained 1D-CNN network model.

[0019] S303: Use the feature vector output by the fully connected layer 1 as ConvHRRP features.

[0020] Preferably, the 1D-CNN model includes a convolutional layer 1, a pooling layer 1, a convolutional layer 2, a pooling layer 2, a convolutional layer 3, a pooling layer 3, a fully connected layer 1, and a fully connected layer 2 connected in sequence.

[0021] Preferably, the joint feature data construction process in step 4 is as follows: the extracted HRRP non-parametric feature vectors, including power spectrum features, central moment features, normalized standard deviation features and echo energy features, are serially concatenated with the extracted ConvHRRP features to obtain joint feature data, which will be used for target identification in HRRP.

[0022] Preferably, the training process of the Stacking ensemble model in step 4 is as follows:

[0023] S401: Divide the joint feature data used for training into 5 parts, denoted as tr1, tr2, tr3, tr4 and tr5;

[0024] S402: Select the base model: SVM, KNN, fully connected neural network, and decision tree;

[0025] S403: For each base model, use one of tr1, tr2, tr3, tr4 and tr5 as the validation set in sequence, and the other four as the training set to perform 5-fold cross-validation to train the base model. Each base model can be trained to obtain 5 sub-models with different built-in parameters.

[0026] S404: Use the trained sub-model to classify and identify the validation set, and stack the results into a new feature; similarly, each base model can obtain a new feature by following this step, and finally obtain 4 sets of new features. This is because 4 base models are selected, and the number of samples for the new features is the same as the number of samples for the total training data.

[0027] S405: Use a logistic regression (LR) model to train the new features to obtain a trained meta-classifier model.

[0028] Preferably, the process of using a trained Stacking ensemble model to identify HRRP for targets with unknown labels is as follows:

[0029] S1: Normalize the HRRP of unknown labels, and then extract the joint features obtained by fusing non-parametric features and ConvHRRP features;

[0030] S2: Use the five sub-models of each base model in the trained Stacking model to classify and identify the joint features respectively, and average the five identification results under each base model. The four base models will yield four new features.

[0031] S3: Input the new features into the trained logistic regression model (meta-classifier) ​​to obtain the final recognition result.

[0032] 3. Beneficial effects

[0033] Compared with the prior art, the advantages of this invention are:

[0034] (1) In this invention, deep abstract features extracted by 1D-CNN and non-parametric features extracted directly from HRRP data are combined at the feature level and the joint features are used for HRRP target recognition. These two types of features are complementary in HRRP target recognition. By combining parametric and non-parametric features, a more comprehensive and accurate HRRP feature representation can be obtained.

[0035] (2) In this invention, the Stacking method is used to integrate multiple classifier models, and decision-level fusion is used to reduce the false recognition rate of a single classifier and improve the robustness and recognition accuracy of the classification model. Attached Figure Description

[0036] Figure 1 This is the spatial model of tank scattering points proposed in this invention.

[0037] Figure 2 This is the spatial model of truck scattering points proposed in this invention.

[0038] Figure 3 This is the spatial model of the scattering points of the interfering object proposed in this invention.

[0039] Figure 4 This is an example of the training and testing method for the base model in the Stacking method proposed in this invention.

[0040] Figure 5 This is an example of the training and testing method for the meta-classifier model in the Stacking method proposed in this invention.

[0041] Figure 6 This is a flowchart illustrating the extraction of HRRP from the MSTAR dataset proposed in this invention.

[0042] Figure 7 This is a block diagram illustrating the specific structure of the HRRP target recognition method based on joint feature extraction and Stacking model classification proposed in this invention. Detailed Implementation

[0043] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0044] Example 1:

[0045] Reference Figure 1-5 The HRRP target recognition method based on joint feature extraction and Stacking model classification includes the following steps:

[0046] Step 1: Obtain HRRP data, perform modulo and normalization on the HRRP data to limit the values ​​to between 0 and 1, thereby eliminating the amplitude sensitivity of the HRRP data;

[0047] Step 2: Extract non-parametric features from the normalized HRRP data, including power spectrum, center moment, echo energy, normalized standard deviation, etc.

[0048] Step 3: Input the normalized HRRP data into the trained 1D-CNN model, and use the output of the first fully connected layer in the network as the feature data to obtain the deep convolutional features of the HRRP data, hereinafter referred to as ConvHRRP. The 1D-CNN model includes convolutional layer 1, pooling layer 1, convolutional layer 2, pooling layer 2, convolutional layer 3, pooling layer 3, fully connected layer 1, and fully connected layer 2 connected in sequence.

[0049] The process of training a 1D-CNN model and extracting ConvHRRP features is as follows:

[0050] S301: Modulo and normalize the HRRP data used for training to limit the values ​​to between 0 and 1;

[0051] S302: Input the normalized HRRP training data into the 1D-CNN network model, use 'sgdm' as the optimizer, set the maximum number of training epochs to 100, the size of each mini-batch to 16, and the initial learning rate to 0.001 to obtain the trained 1D-CNN network model.

[0052] S303: Use the feature vector output by fully connected layer 1 as ConvHRRP features;

[0053] Step 4: Fuse the normalized HRRP non-parametric features with the ConvHRRP features, and input the joint feature data into the trained Stacking ensemble model to obtain the HRRP target recognition result. The joint feature data construction process is as follows: The extracted HRRP non-parametric feature vectors, including power spectrum features, central moment features, normalized standard deviation features, and echo energy features, are serially concatenated with the extracted ConvHRRP features to obtain the joint feature data, which will be used for HRRP target recognition.

[0054] The training process of the Stacking ensemble model is as follows:

[0055] S401: Divide the joint feature data used for training into 5 parts, denoted as tr1, tr2, tr3, tr4 and tr5;

[0056] S402: Select the base model: SVM, KNN, fully connected neural network, and decision tree;

[0057] S403: For each base model, use one of tr1, tr2, tr3, tr4 and tr5 as the validation set in sequence, and the other four as the training set to perform 5-fold cross-validation to train the base model. Each base model can be trained to obtain 5 sub-models with different built-in parameters.

[0058] S404: Use the trained sub-model to classify and identify the validation set, and stack the results into a new feature; similarly, each base model can obtain a new feature by following this step, and finally obtain 4 sets of new features. This is because 4 base models are selected, and the number of samples for the new features is the same as the number of samples for the total training data.

[0059] S405: Use a logistic regression (LR) model to train the new features to obtain a trained meta-classifier model.

[0060] In this invention, the process of using a trained Stacking ensemble model to identify HRRP for unknown labeled targets is as follows:

[0061] S1: Normalize the HRRP of unknown labels, and then extract the joint features obtained by fusing non-parametric features and ConvHRRP features;

[0062] S2: Use the five sub-models of each base model in the trained Stacking model to classify and identify the joint features respectively, and average the five identification results under each base model. The four base models will yield four new features.

[0063] S3: Input the new features into the trained logistic regression model (meta-classifier) ​​to obtain the final recognition result.

[0064] In this invention, deep abstract features extracted by 1D-CNN and non-parametric features extracted directly from HRRP data are fused at the feature level, and the joint features are used for HRRP target recognition. These two types of features are complementary in HRRP target recognition, and by combining parametric and non-parametric features, a more comprehensive and accurate HRRP feature representation can be obtained.

[0065] In this invention, the Stacking method is used to integrate multiple classifier models. Decision-level fusion is used to reduce the false recognition rate of a single classifier and improve the robustness and recognition accuracy of the classification model.

[0066] Example 2:

[0067] This embodiment includes the following parts: computer simulation to establish a spatial model of vehicle target scattering points, radar echo signal simulation, joint feature extraction, and Stacking model classification and recognition.

[0068] 1. Computer simulation to establish a spatial model of vehicle target scattering points.

[0069] 1.1 Establishing a spatial model of tank scattering points

[0070] A spatial model of the tank's scattering points is established based on the tank's dimensions: length 9.77m, width 3.75m, height 2.44m, turret height 2.92m, turret length 3.95m, turret width 2.50m, and track width 0.61m.

[0071] 1.2 Establishing a spatial model of truck scattering points

[0072] Based on the dimensional parameters of the truck body (length 6.38m, width 2.44m, height 2.74m, track 1.93m, front overhang 1.60m, rear overhang 1.20m, trailer height 0.87m, tire width 0.305m, and tire height 0.965m), a spatial model of the truck's scattering points is established.

[0073] 1.3 Establishing a spatial model of the misleading scatterer

[0074] A spatial model of the misleading scatterer is established based on the dimensions of 5m in length, 2m in width, 2m in height, and 1m spacing between scattering points.

[0075] 2. Radar echo signal simulation

[0076] In the simulation, two coordinate systems were established: the radar detection coordinate system and the target scattering point coordinate system. In the radar detection coordinate system, the radar is located at the origin, transmitting a linear frequency modulated (LFM) signal that propagates through free space to the target, is reflected back by the target's scattering points, and is finally received by the receiving antenna for beamforming. In the target scattering point coordinate system, each target scattering point can rotate freely 360° around the origin of this coordinate system's rotation center. The radar system parameters in the simulation experiment are: radar system operating frequency 4.5 GHz, sampling frequency 800 MHz, pulse repetition frequency 10 kHz, pulse width 20e-9 s, and frequency modulation bandwidth 300 MHz. The simulation generated 1800 HRRP data points, including three categories, with 600 data points in each category. The simulation experiment used 1300 data points for training and 500 data points for testing.

[0077] 3. Joint Feature Extraction

[0078] 3.1 Extracting HRRP Nonparametric Features

[0079] The obtained HRRP data was normalized, and non-parametric features were extracted from both the training and test data according to the following steps. Specifically, the non-parametric feature extraction process is as follows:

[0080] (1) Extract the power spectrum features of the normalized HRRP. To avoid information redundancy from adversely affecting the classification model, principal component analysis (PCA) is used to extract dimensions with a contribution rate greater than 0.95. The formula for calculating the power spectrum features is as follows:

[0081]

[0082] Where P(f) is the power spectral density at frequency f, N is the number of sample points, and T s is the sampling interval, and x(n) is the amplitude (normalized) of the nth sampling point of HRRP.

[0083] (2) Extract the normalized HRRP central moments feature. The calculation formula is as follows:

[0084] The expression for the second-order central moment (variance):

[0085]

[0086] The expression for the third-order central moment (skewness):

[0087]

[0088] Fourth-order central moment (kurtosis) expression:

[0089]

[0090] The expression for the fifth-order central moment is:

[0091]

[0092] Where is the length of the HRRP sequence, is the amplitude (normalized) of the HRRP th sampling point, and is the mean of the HRRP sequence.

[0093] (3) Extract the normalized HRRP echo energy characteristics. The calculation formula is as follows:

[0094]

[0095] Where N is the length of the HRRP sequence, and x(n) is the amplitude of the nth sampling point of HRRP.

[0096] (4) Extract the normalized standard deviation feature of normalized HRRP, the calculation formula is:

[0097]

[0098] Where N is the length of the HRRP sequence, and x(n) is the amplitude of the nth sampling point of HRRP.

[0099] 3.2 1D-CNN for HRRP Feature Extraction (ConvHRRP)

[0100] The obtained HRRP data is normalized and then input into a 1D-CNN model. The output of fully connected layer 1 is used as the feature vector, i.e., 1D-CNN extracts HRRP features (ConvHRRP). Specifically: During training, the normalized HRRP used for training is input into the 1D-CNN model with network parameters shown in Table 1 for training. 'sgdm' is used as the optimizer, the maximum number of training epochs is set to 100, the size of each mini-batch is 16, and the initial learning rate is 0.001, resulting in a trained 1D-CNN network model. In the last training epoch, the output of fully connected layer 1 is used as the ConvHRRP training data for subsequent classification. During testing, the normalized HRRP used for testing is input into the trained 1D-CNN network model, and the output of fully connected layer 1 is used as the ConvHRRP test data for subsequent classification.

[0101] 3.3 Construction of Joint Feature Data

[0102] The extracted HRRP nonparametric feature vector, including power spectrum features, central moment features, normalized standard deviation features, and echo energy features, is expressed as follows:

[0103] f non =[f non (1), f non (2), ..., f non (L)]

[0104] Where L is the dimension of the non-parametric HRRP feature vector. The ConvHRRP features extracted by 1D-CNN, i.e., the feature vector obtained after the weight parameters of the convolution kernel pass through pooling layers and fully connected layers, are expressed as:

[0105] f conv =[f conv (1), f conv (2), ..., f conv (N)]

[0106] Where N is the dimension of the ConvHRRP feature vectors. Joint features are constructed using a serial concatenation method, expressed as:

[0107] f j =[f non f conv ]

[0108] The joint features extracted from the normalized HRRP will be fed into the subsequent classification model for classification and recognition.

[0109] 4. Stacking model for classification and recognition

[0110] The HRRP joint feature data is input into the Stacking model as follows: During training, the training set is divided into five parts, denoted as tr1, tr2, tr3, tr4, and tr5; Support Vector Machine (SVM), K-Nearest Neighbor (KNN), Artificial Neural Network (NN), and Decision Tree (Tree) are selected as four base models; For each base model, one of tr1, tr2, tr3, tr4, and tr5 is used as the validation set, and the remaining four parts are used as the training set, and 5-fold cross-validation is performed to train the base model; The trained sub-models classify and recognize the validation set, and the results are stacked into a new feature. Similarly, each base model can obtain a new feature by following this step, and finally, four sets of new features are obtained; The logistic regression model is used to train the new features to obtain a meta-classifier model; During testing, the test set is input into the five sub-models of each base model that were previously trained, and they are classified and recognized separately. The five results under each base model are averaged; The four base models will obtain four new features, which are input into the trained meta-classifier to obtain the final recognition result. The performance comparison of the experiments is shown in Table 2. It can be seen that the method of using the joint features of ConvHRRP features and non-parametric features, and using the Stacking ensemble model for HRRP target recognition, has higher recognition accuracy than the individual non-parametric features or ConvHRRP features, as well as the single classification model. The overall classification recognition effect is the best.

[0111] Table 2

[0112]

[0113] Specifically, the training and testing process of the base model in the Stacking model described above is as follows: Figure 4 As shown, the training and testing process of the meta-classifier can be as follows: Figure 5 As shown; the specific structural block diagram of the HRRP target recognition method based on joint feature extraction and Stacking model classification is as follows. Figure 7 As shown.

[0114] Example 2:

[0115] This embodiment includes the following parts: obtaining HRRP from the MSTAR public dataset, joint feature extraction, and Stacking model classification and recognition.

[0116] 1. Obtain HRRP from the MSTAR public dataset.

[0117] according to Figure 6The steps shown extract HRRP data from the MSTAR public SAR dataset. Specifically, for vehicle target images in the MSTAR dataset, target information is concentrated in the center of the image. First, a 48×48 mask is placed over the center of the image, and global CFAR detection is performed within the mask area. The core idea of ​​global CFAR is that the number of reference cells is always equal to the sliding window size, which can fully utilize the information within the mask and has very high computational efficiency because global CFAR only needs to perform a summation operation on the data within the area. Since the acquisition scenes of the MSTAR dataset are grass-covered and flat with relatively uniform clutter, a Rayleigh CA-CFAR detector is selected. After extracting the target from the background clutter, zero-padding is performed to fill the image to the original image size, thus obtaining a target segment. For the obtained target segment, an IFFT transform is performed along the cross-range to obtain the range image acquired by the synthetic aperture, where each column can be used as an HRRP for the angle. The experiment selected three types of vehicle targets at pitch angles of 15° and 17°: a T62 tank, a BTR60 armored personnel carrier, and a ZIL131 cargo truck. A total of 3016 HRRPs were collected, with 2016 used for training and 1000 used for testing.

[0118] 2. Joint Feature Extraction

[0119] 2.1 Extracting HRRP Nonparametric Features

[0120] The obtained HRRP data was normalized, and non-parametric features were extracted from both the training and test data according to the following steps. Specifically, the non-parametric feature extraction process is as follows:

[0121] (1) Extract the power spectrum features of the normalized HRRP. To avoid information redundancy from adversely affecting the classification model, principal component analysis (PCA) is used to extract dimensions with a contribution rate greater than 0.95. The formula for calculating the power spectrum features is as follows:

[0122]

[0123] Where P(f) is the power spectral density at frequency f, N is the number of sample points, and T s is the sampling interval, and x(n) is the amplitude (normalized) of the nth sampling point of HRRP.

[0124] (2) Extract the normalized HRRP central moments feature. The calculation formula is as follows:

[0125] The expression for the second-order central moment (variance):

[0126]

[0127] The expression for the third-order central moment (skewness):

[0128]

[0129] Fourth-order central moment (kurtosis) expression:

[0130]

[0131] The expression for the fifth-order central moment is:

[0132]

[0133] Where is the length of the HRRP sequence, is the amplitude (normalized) of the HRRP th sampling point, and is the mean of the HRRP sequence.

[0134] (3) Extract the normalized HRRP echo energy characteristics. The calculation formula is as follows:

[0135]

[0136] Where N is the length of the HRRP sequence, and x(n) is the amplitude of the nth sampling point of HRRP.

[0137] (4) Extract the normalized standard deviation feature of normalized HRRP, the calculation formula is:

[0138]

[0139] Where N is the length of the HRRP sequence, and x(n) is the amplitude of the nth sampling point of HRRP.

[0140] 2.2 1D-CNN for HRRP Feature Extraction (ConvHRRP)

[0141] The obtained HRRP data is normalized and then input into a 1D-CNN model. The output of fully connected layer 1 is used as the feature vector, i.e., 1D-CNN extracts HRRP features (ConvHRRP). Specifically: During training, the normalized HRRP used for training is input into the 1D-CNN model with network parameters shown in Table 1 for training. 'sgdm' is used as the optimizer, the maximum number of training epochs is set to 100, the size of each mini-batch is 16, and the initial learning rate is 0.001, resulting in a trained 1D-CNN network model. In the last training epoch, the output of fully connected layer 1 is used as the ConvHRRP training data for subsequent classification. During testing, the normalized HRRP used for testing is input into the trained 1D-CNN network model, and the output of fully connected layer 1 is used as the ConvHRRP test data for subsequent classification.

[0142] Table 1

[0143]

[0144]

[0145] 2.3 Construction of Joint Feature Data

[0146] The extracted HRRP nonparametric feature vector, including power spectrum features, central moment features, normalized standard deviation features, and echo energy features, is expressed as follows:

[0147] f non =[f non (1), f non (2), …f non (L)]

[0148] Where L is the dimension of the HRRP nonparametric feature vector. The ConvHRRP features extracted by 1D-CNN, i.e., the feature vector obtained after the weight parameters of the convolution kernel pass through pooling layers and fully connected layers, are expressed as:

[0149] f conv =[f conv (1), f conv (2), ..., f conv (N)]

[0150] Where N is the dimension of the ConvHRRP feature vectors. Joint features are constructed using a serial concatenation method, expressed as:

[0151] f j =[f non f conv ]

[0152] The joint features extracted from the normalized HRRP will be fed into the subsequent classification model for classification and recognition.

[0153] 3. Stacking model for classification and recognition

[0154] The HRRP joint feature data is input into the Stacking model as follows: During training, the training set is divided into five parts, denoted as tr1, tr2, tr3, tr4, and tr5; Support Vector Machine (SVM), K-Nearest Neighbor (KNN), Artificial Neural Network (NN), and Decision Tree (Tree) are selected as four base models; For each base model, one part of tr1, tr2, tr3, tr4, and tr5 is used as the validation set, and the remaining four parts are used as the training set, and 5-fold cross-validation is performed to train the base model; The trained sub-models classify and recognize the validation set, and the results are stacked into a new feature. Similarly, each base model can obtain a new feature by following this step, and finally, four sets of new features are obtained; The logistic regression model is used to train the new features to obtain a meta-classifier model; During testing, the test set is input into the five sub-models of each base model that were previously trained, and they are classified and recognized separately. The five results of each base model are averaged; The four base models will obtain four new features, which are input into the trained meta-classifier to obtain the final recognition result. The performance comparison of the experiments is shown in Table 3. It can be seen that the method of using the joint features of ConvHRRP features and non-parametric features, and using the Stacking ensemble model for HRRP target recognition, has higher recognition accuracy than the individual non-parametric features or ConvHRRP features, as well as the single classification model. The overall classification recognition effect is the best.

[0155] Table 3

[0156]

[0157] Specifically, the training and testing process of the base model in the Stacking model described above is as follows: Figure 4 As shown, the training and testing process of the meta-classifier can be as follows: Figure 5 As shown; the specific structural block diagram of the HRRP target recognition method based on joint feature extraction and Stacking model classification is as follows. Figure 7 As shown.

[0158] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A HRRP target recognition method based on joint feature extraction and stacking model classification, characterized in that, The method comprises the following steps: Step 1: obtaining HRRP data, taking the HRRP data modulo and normalizing the HRRP data to limit the value to 0 to 1, so as to eliminate the amplitude sensitivity of the HRRP data; Step 2: extracting non-parametric features of the normalized HRRP data, which include power spectrum, central moment, echo energy and normalized standard deviation feature; Step 3: inputting the normalized HRRP data into a trained 1D-CNN model, taking the output of the first fully connected layer in the network as feature data, and obtaining deep convolution features of the HRRP data, which are referred to as ConvHRRP hereinafter; Step 4: fusing the non-parametric features of the normalized HRRP and the ConvHRRP features, inputting the joint feature data into a trained Stacking integrated model, and obtaining an HRRP target recognition result. 2.The HRRP target recognition method based on joint feature extraction and stacking model classification according to claim 1, characterized in that, The process of training the 1D-CNN model in step 3 and extracting the ConvHRRP features is as follows: S301: taking the HRRP data used for training modulo and normalizing the HRRP data to limit the value to 0 to 1; S302: inputting the normalized HRRP training data into the 1D-CNN network model, using'sgdm' as the optimizer, setting the maximum number of training rounds to 100, the size of each small batch to 16, and the initial learning rate to 0.001, and obtaining a trained 1D-CNN network model; S303: taking the feature vector output by the first fully connected layer as the ConvHRRP feature. 3.The HRRP target recognition method based on joint feature extraction and stacking model classification according to claim 1, characterized in that, The 1D-CNN model comprises, in sequence, a convolution layer 1, a pooling layer 1, a convolution layer 2, a pooling layer 2, a convolution layer 3, a pooling layer 3, a first fully connected layer and a second fully connected layer.

4. The HRRP target recognition method based on joint feature extraction and stacking model classification according to claim 1, characterized in that, The joint feature data construction process in step 4 is as follows: the extracted HRRP non-parametric feature vectors, including power spectrum features, central moment features, normalized standard deviation features and echo energy features, and the extracted ConvHRRP features are serially spliced to obtain joint feature data, which will be used for HRRP target recognition.

5. The HRRP target recognition method based on joint feature extraction and stacking model classification according to claim 1, characterized in that, The training process of the Stacking integrated model in step 4 is as follows: S401: dividing the joint feature data used for training into five parts, denoted as tr1, tr2, tr3, tr4 and tr5; S402: selecting base models: SVM, KNN, fully connected neural network and decision tree; S403: for each base model, one of tr1, tr2, tr3, tr4 and tr5 is used as a validation set, and the remaining four are used as training sets, and the base model is trained by 5-fold cross-validation, and each base model can be trained to obtain five sub-models with different built-in parameters; S404: using the trained sub-models to classify and identify the validation set, and stacking the results as a new feature; similarly, each base model can obtain a new feature according to this step, and finally four new features are obtained, which is because four base models are selected, and the sample number of the new feature and the sample number of the total training data are the same; S405: Train the new features using a logistic regression (LR) model to obtain a trained meta-classifier model.

6. The HRRP target recognition method based on joint feature extraction and stacking model classification according to claim 1, characterized in that, The process of using the trained Stacking integrated model to identify the HRRP of unknown label targets is as follows: S1: Normalize the HRRP of unknown label, then extract the non-parametric feature and the joint feature after ConvHRRP feature fusion; S2: Use the 5 sub-models of each base model in the trained Stacking model to classify and identify the joint feature respectively, take the average of the 5 identification results under each base model, and 4 new features will be obtained for the 4 base models; S3: Input the new features into the trained logistic regression model (meta-classifier) to obtain the final identification result.

Images