A low-altitude target intelligent classification method based on multi-domain feature fusion

Abstract

This application provides a low-altitude target intelligent classification method based on multi-domain feature fusion, comprising: preprocessing echo data of low-altitude targets to obtain time-domain data and Doppler-domain data of the targets; extracting feature vectors of the low-altitude targets from at least two different feature domains based on the time-domain data and Doppler-domain data, and fusing the feature vectors from different feature domains to obtain multi-domain feature vectors; inputting the multi-domain feature vectors into a trained random forest classifier, and outputting the category of the low-altitude target. This application performs multi-domain feature extraction based on radar target measurement data, mining the essential differentiated features of targets from different dimensions; and introduces the random forest algorithm to construct an intelligent classification model, achieving effective fusion of multi-domain features and accurate target classification. This application has high operating efficiency, can realize the function of low-altitude target identification at the radar front end, and improves the practicality of feature extraction-based low-altitude target identification methods.

Description

Technical Field

[0001] This application belongs to the field of radar technology, specifically relating to an intelligent classification method for low-altitude targets based on multi-domain feature fusion. Background Technology

[0002] With the rapid development of the low-altitude economy, low-altitude airspace has been widely used in civilian aerial photography, logistics transportation, emergency rescue, and other fields. However, it also faces security threats such as illegal intrusion and malicious reconnaissance. Accurate classification of low-altitude targets has become a core technological support for ensuring low-altitude safety. Among numerous low-altitude targets, small drones, birds, and biomimetic wings (artificial biomimetic flapping-wing aircraft, abbreviated as biomimetic wings) share common characteristics such as small size, flexible flight attitude, low flight altitude, and small radar cross-section. They are also easily affected by complex environmental factors such as building obstruction, weather interference, and ground clutter, making it difficult for traditional identification methods to achieve accurate differentiation. Crucially, research on the classification of low-altitude targets such as birds, drones, and biomimetic wings has not yet developed a mature and effective technical solution, leaving a significant gap in related research.

[0003] Target classification is a crucial step in the low-altitude target identification process, and its performance directly determines the effectiveness of subsequent situational awareness and decision-making responses. Existing research on low-altitude target classification suffers from two major limitations: First, the focus is biased, often targeting large mid-to-high altitude targets (such as aircraft and helicopters) or single types of low-altitude targets (such as drones), neglecting specific research on groups of low-altitude targets with overlapping shapes and dynamic characteristics, such as birds, drones, and biomimetic wings. Second, data support is insufficient. More notably, most existing research relies on simulation data for experimental verification, lacking real-world experimental data support in actual low-altitude scenarios. Simulation data often simplifies complex factors in real-world environments, such as clutter interference and random changes in target attitude, resulting in significant differences from actual application scenarios. This leads to weak persuasiveness of research conclusions based on simulation data, making it difficult to directly transfer to engineering practice. Furthermore, existing research generally relies on single-domain features (such as time-domain, Doppler-domain, or time-frequency-domain features), resulting in incomplete feature representation. Therefore, for specific low-altitude targets such as birds, drones, and bionic wings, based on measured data, multi-domain feature fusion is used to fully explore the multi-dimensional differentiated information of the targets, improve the accuracy and robustness of classification in complex low-altitude environments, and fill the research gap in this field. This has become a key problem that urgently needs to be solved in the field of low-altitude target recognition. Summary of the Invention

[0004] To address the aforementioned problems in the existing technology, this application provides an intelligent classification method for low-altitude targets based on multi-domain feature fusion. The technical problem to be solved by this application is achieved through the following technical solution: A low-altitude target intelligent classification method based on multi-domain feature fusion includes: S100: Acquire measured echo data of the radar against low-altitude targets, and preprocess the echo data to obtain the target's time-domain data and Doppler-domain data. S200, based on the time-domain data and Doppler-domain data, extract feature vectors of the target from at least two different feature domains respectively, and fuse the feature vectors from different feature domains to obtain a multi-domain feature vector; wherein, the feature domains include at least two of the time domain, Doppler domain, time-frequency domain and wavelet domain; S300, the multi-domain feature vector is input into the trained random forest classifier, and the category of the low-altitude target is output.

[0005] Beneficial effects: This application provides a low-altitude target intelligent classification method based on multi-domain feature fusion, comprising: acquiring measured echo data of low-altitude targets from radar, and preprocessing the echo data to obtain time-domain data and Doppler-domain data of the target; based on the time-domain data and Doppler-domain data, extracting feature vectors of the target from at least two different feature domains respectively, and fusing the feature vectors from different feature domains to obtain a multi-domain feature vector; wherein the feature domains include at least two of the time domain, Doppler domain, time-frequency domain, and wavelet domain; inputting the multi-domain feature vector into a trained random forest classifier, and outputting the category of the low-altitude target. This application performs multi-domain feature extraction based on measured radar target data, mining the essential differentiated features of the target from different dimensions; and introduces the random forest algorithm to construct an intelligent classification model, achieving effective fusion of multi-domain features and accurate target classification. This application has high operating efficiency, can realize the function of low-altitude target identification at the radar front end, and improves the practicality of low-altitude target identification methods based on feature extraction.

[0006] The present application will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of a low-altitude target intelligent classification method based on multi-domain feature fusion provided in this application; Figure 2a This is a schematic diagram of the process of training a random forest classifier provided in this application; Figure 2b This is a schematic diagram of the radar measured data processing flow provided in this application; Figure 3 This is a diagram of the bird targets provided in this application; Figure 4 This is a diagram of the UAV target provided in this application; Figure 5 This is a diagram of the biomimetic wing target provided in this application; Figure 6 This is a schematic diagram of the decision tree provided in this application; Figure 7 This is a schematic diagram of the random forest model provided in this application; Figure 8 This is a schematic diagram of the confusion matrix statistical results of a single-domain feature provided in this application; Figure 9 This is a schematic diagram of the statistical results of the confusion matrix of the two-domain feature combination provided in this application; Figure 10 This is a schematic diagram of the statistical results of the confusion matrix of the three-domain feature combination provided in this application; Figure 11 This is a schematic diagram of the statistical results of the confusion matrix of the four-domain full feature combination provided in this application. Detailed Implementation

[0008] The present application will be described in further detail below with reference to specific embodiments, but the implementation of the present application is not limited thereto.

[0009] like Figure 1 As shown, this application provides a low-altitude target intelligent classification method based on multi-domain feature fusion, including: S100: Acquire measured echo data of the radar against low-altitude targets, and preprocess the echo data to obtain the target's time-domain data and Doppler-domain data. S200, based on the time-domain data and Doppler-domain data, extract feature vectors of the target from at least two different feature domains respectively, and fuse the feature vectors from different feature domains to obtain a multi-domain feature vector; wherein, the feature domains include at least two of the time domain, Doppler domain, time-frequency domain and wavelet domain; S300, the multi-domain feature vector is input into the trained random forest classifier, and the categories of the multi-class low-altitude targets are output.

[0010] In one specific embodiment of this application, S100 includes: S110: Acquires measured echo data of the radar against low-altitude targets; S120, the measured echo data is sequentially subjected to pulse compression, clutter suppression, moving target detection and constant false alarm rate detection to obtain the time domain data and Doppler domain data.

[0011] In one specific embodiment of this application, S200 includes: S210, based on the time-domain data and the Doppler-domain data, select feature vectors from at least two feature domains in the time domain, Doppler domain, time-frequency domain and wavelet domain; When the feature domain includes the time domain, the time domain feature extraction in S220 includes: extracting the mean, variance, envelope standard deviation, average peak value, peak standard deviation, and time domain waveform entropy from the time domain data.

[0012] Time-domain echoes can effectively capture the dynamic behavior of targets over time. Furthermore, compared to other feature domains, features extracted in the time domain are less sensitive to parameters such as dwell time and pulse repetition frequency, while other domains require more stringent parameter constraints. Assuming the time-domain data is... Then, the following features are extracted from it.

[0013] 1. Mean (0-1); The mean value characterizes the average amplitude level of the time-domain signal within the sampling time, reflecting the overall energy shift characteristics of the target echo.

[0014] 2. Variance (0-2); Variance reflects the degree of fluctuation of the time-domain signal amplitude around the mean, and embodies the instantaneous change characteristics of the target echo. The larger the variance, the more violent the fluctuation of the echo amplitude.

[0015] 3. Envelope Standard Deviation This feature requires first performing a Hilbert transform on the time-domain signal to extract the echo envelope, and then calculating the standard deviation of the envelope, which requires two steps of derivation: (1) Hilbert transform extracts analytic signal and echo envelope For echo signal Perform Hilbert transform to obtain its orthogonal components. Construct analytical signals : (0-3); The envelope of the signal is obtained by taking the modulus of the analytic signal. : (0-4); (2) Calculate the envelope standard deviation (0-5); in, This represents the mean of the signal envelope. The standard deviation of the envelope reflects the dispersion and fluctuation characteristics of the echo envelope amplitude; a larger standard deviation indicates more significant envelope fluctuations.

[0016] 4. Average peak value This feature requires first obtaining all the peak values ​​of the time-domain signal magnitude sequence, and then calculating the mean of the peak values. It needs to be derived through the following two steps: (1) Peak Extraction For the magnitude sequence of discrete-time domain signals All local peaks are extracted using peak detection technology, and a peak sequence is constructed. ,in The number of peak values. For the first The amplitude of each peak.

[0017] (2) Calculate the average peak value (0-6); The average peak value reflects the average level of local maxima in the time domain echo, and embodies the degree of peak energy concentration of the target signal.

[0018] 5. Peak standard deviation (0-7); The peak standard deviation reflects the degree of dispersion of the peak amplitude relative to the average peak value, and embodies the stability and periodicity of the target echo peak value.

[0019] 6. Time-domain waveform entropy (0-8); The entropy of a time-domain waveform reflects the degree of dispersion of energy in a time-domain signal; the higher the entropy value, the more dispersed the energy, and the lower the entropy value, the more concentrated the energy.

[0020] When the feature domain includes the Doppler domain, the Doppler feature extraction in S220 includes: extracting the amplitude variance, first-order origin moment, second-order center distance, L1 norm of normalized amplitude, Doppler spectrum differential mode after normalization amplitude, and Doppler spectrum waveform entropy from the Doppler domain data.

[0021] The Doppler spectrum of a target can intuitively characterize its micro-Doppler properties. Since the Doppler component generated by the translational motion of the target body is usually much stronger than the micro-motion component, the latter is easily masked by the former. However, different target structures have different energy ratios for their translational and micro-motion components. This difference is generally reflected in the outline and energy distribution of the Doppler spectrum. Therefore, this application can extract features from it to effectively distinguish different types of targets. Assuming the Doppler domain data is... Then, the following features were extracted from it.

[0022] 1. Amplitude Variance (0-9); The amplitude variance in the Doppler domain reflects the degree of dispersion of the Doppler spectrum amplitude relative to its mean, and characterizes the uniformity of the energy distribution of the Doppler spectrum. The larger the value, the more uneven the energy distribution.

[0023] 2. First-order moment at the origin (0-10); The first-order origin moment of the Doppler spectrum represents the central position of the energy of the Doppler spectrum, and its magnitude indicates the position of the frequency point with the highest energy percentage.

[0024] 3. Second-order center distance (0-11); The second central moment of the Doppler spectrum characterizes the degree of dispersion of the Doppler spectrum energy relative to the dominant frequency, and characterizes the distribution range of the target's micro-motion frequency; the larger the value, the more dispersed the energy.

[0025] 4. Normalized amplitude norm (0-12); The norm characterizes the degree of accumulation of the overall energy of the Doppler spectrum. The larger the value, the higher the accumulation and the fuller the Doppler spectrum.

[0026] 5. Difference modes of the Doppler spectrum after normalization of amplitude (0-13); This feature reflects the intensity of local variations in the Doppler spectrum amplitude, characterizing its smoothness and fluctuations.

[0027] 6. Doppler spectrum waveform entropy (0-14); The entropy of the Doppler spectrum waveform reflects the degree of dispersion of signal energy in the Doppler domain. The higher the entropy value, the more dispersed the signal; the lower the entropy value, the more concentrated the signal.

[0028] When the feature domain includes the time-frequency domain, the time-frequency domain feature extraction in S200 includes: performing a short-time Fourier transform on the time-domain data to obtain a time-frequency power matrix; and extracting spectral entropy, micro-Doppler clustering degree, and singular value entropy based on the time-frequency power matrix.

[0029] To fully observe the energy distribution of the echo signal in both the frequency and time dimensions, this paper implements joint time-frequency analysis using the Short Time Fourier Transform (STFT). This method can simultaneously capture the joint changes of the signal in both the time and frequency dimensions, and is a classic time-frequency analysis method.

[0030] Suppose the radar time-domain signal after passing through STFT is... The number of sampling points in the frequency dimension is The number of sampling points in the time dimension is , Indicates the first The frequency point, the first The amplitude of the time-frequency signal at each moment. Let represent the time-frequency power matrix. The following section describes feature extraction from this matrix.

[0031] 1. Spectral entropy This feature is based on the Shannon entropy of the average power spectrum and requires the following three steps to derive: (1) Calculate the average spectrum For time-frequency power matrix Average the time series data at each frequency point column by column to obtain the average power spectrum at each frequency point. : (0-15); in, Indicates the first Average power at each frequency point.

[0032] (2) Constructing the normalized probability spectrum Normalizing the average power spectrum to satisfy the constraint that the sum of probabilities is 1 yields the probability spectrum: (0-16); in, Indicates the first The power ratio at each frequency point satisfies .

[0033] (3) Calculate the spectral entropy (0-17); Spectral entropy reflects the disorder and uniformity of the energy distribution of the micro-Doppler average power spectrum. The higher the value, the more dispersed the energy distribution and the more complex the micro-motion characteristics of the target.

[0034] 2. MicroDoppler aggregation (0-18); Micro-Doppler concentration primarily describes the spatial concentration of micro-Doppler time-frequency energy, reflecting whether the energy is concentrated on a few time-frequency points. The larger the value, the more concentrated the time-frequency energy is on a few high-energy time-frequency points, and the stronger the micro-motion regularity of the target.

[0035] 3. Singular value entropy This feature is based on Shannon entropy from singular value decomposition, and mainly describes the complexity of the time-frequency matrix through the singular values ​​of the matrix. It requires the following three derivation steps: (1) Singular value decomposition For time-frequency matrix Perform singular value decomposition to extract non-zero singular values: (0-19); in, , It is an orthogonal array. It is a diagonal matrix, and the elements on its diagonal are Non-zero singular values: .

[0036] (2) Normalized singular values Normalizing the singular values ​​yields the singular value weights: (0-20); in, Indicates the first The weights of the singular values, and satisfying .

[0037] (3) Calculate the singular value entropy (0-21); Singular value entropy reflects the structural complexity and inherent regularity of a time-frequency matrix. The higher the value, the more complex the matrix structure and the less obvious the inherent regularity.

[0038] When the feature domain includes the wavelet domain, wavelet domain feature extraction in S200 includes: performing discrete wavelet transform on the time domain data to obtain detail coefficients and approximation coefficients at multiple scales; and extracting relative wavelet energy, detail coefficient kurtosis, and multi-scale log-variance based on the detail coefficients and approximation coefficients.

[0039] Compared to traditional time-frequency analysis, wavelet transform can decouple signal components through scale decomposition, giving it a strong ability to capture transient singular signals from targets. The main advantage of Discrete Wavelet Transform (DWT) lies in its inherent multi-resolution analysis capability, which aligns perfectly with the physical characteristic of "multi-scale motion coexistence" in radar micro-Doppler signals. Furthermore, the orthogonality and low computational complexity of DWT ensure the compactness and real-time performance of the feature space, providing complementary dimensions for distinguishing different targets; therefore, wavelet domain features are introduced here.

[0040] Assume the number of decomposition levels of DWT is After decomposition, the time-domain signal is obtained as follows: The detail factor of the layer is and the The approximation coefficient of the layer is ,in , For translation parameters, the specific implementation of DWT can be found using... In Function implementation. The following section describes feature extraction based on the results of DWT.

[0041] 1. Relative wavelet energy This feature is calculated through the following four steps: (1) Calculate the energy of each decomposition level. (0-22); in, Indicates the first Energy of layer detail components, This indicates the number of coefficients in this layer.

[0042] (2) Calculate the first Energy of the layer approximate component (0-23); in, Indicates the first Layer approximation coefficient The length.

[0043] (3) Calculate the total energy of the signal (0-24); (4) Calculate the relative wavelet energy at each node scale. (0-25); Relative wavelet energy reflects the spatial distribution of energy at different scales. The low-frequency scale mainly contains the translational Doppler information of the target body, while the high-frequency scale mainly contains the instantaneous motion information of the target's small components.

[0044] 2. Detail coefficient kurtosis With the first Taking a layer as an example, its detail coefficient kurtosis is defined as: (0-26); in, This represents the mean value of the coefficients in this layer.

[0045] The detail coefficient kurtosis characterizes the transient impact characteristics of a signal at multiple resolutions, reflecting the sharpness of signal flicker.

[0046] 3. Multiscale log-variance With the first Taking a layer as an example, the log-variance of its detail coefficients is defined as: (0-27); Multiscale logarithmic variance measures the range of energy fluctuations of a signal at a specific scale, and can reflect the dynamic intensity and regularity of target motion fluctuations at different resolutions.

[0047] S220 fuses feature vectors from different feature domains to obtain multi-domain feature vectors.

[0048] Based on the preprocessed time-domain and Doppler-domain data, this application uses the above-mentioned extraction principle to extract features, obtaining a total of 18 features of the target in the time domain, Doppler domain, time-frequency domain, and wavelet domain, which are then combined to form a multi-domain feature vector.

[0049] The training process of the random forest classifier provided in this application includes four stages: radar target echo preprocessing, multi-domain feature extraction, classifier training, and classifier testing. The overall flowchart is as follows: Figure 2a As shown. This application conducts multi-domain feature extraction research based on radar target measurement data, mining the essential differentiated features of targets from different dimensions; and introduces the random forest algorithm to construct an intelligent classification model, realizing the effective fusion of multi-domain features and accurate target classification; at the same time, to verify the impact of different feature combinations on classification performance, the classification results of single-domain features and different combinations of features are statistically analyzed, ultimately verifying the effectiveness and superiority of the proposed method. This application has high operating efficiency and can realize the function of low-altitude target identification at the radar front end, improving the practicality of low-altitude target identification methods based on feature extraction.

[0050] In one specific embodiment of this application, the training process of the completed random forest classifier includes: a. Acquire radar echo data of three types of low-altitude targets: birds, drones, and biomimetic wings, and preprocess the echo data to obtain time-domain and Doppler-domain data of the low-altitude targets. The processing flow of radar measured data is as follows: Figure 2b As shown, the preprocessing of radar measured data mainly involves pulse compression, clutter suppression, moving target detection, and constant false alarm rate (CFAR) detection. This paper uses adaptive moving target detection to simultaneously achieve clutter suppression and moving target detection, and employs cell-averaged CFAR detection. Through this process, this application can ultimately obtain time-domain and Doppler-domain data for three types of low-altitude targets: birds, UAVs, and biomimetic wings. The time-domain and Doppler-domain results for these three types of targets are presented below. Figure 3 , 4 As shown in Figure 5.

[0051] Depend on Figures 3-5It can be observed that, in terms of time domain results, the fluctuations of bird targets exhibit relatively stable quasi-periodicity, while the fluctuations of UAV targets are less stable compared to bird targets, and the time domain results of bionic wing targets show relatively large amplitude fluctuations. In terms of Doppler domain results, all three types of targets have obvious main peaks, which are caused by the translation of the target body. At the same time, there are also varying degrees of spectral broadening around the main peaks of the three types of targets, which are caused by the micro-motions of the targets. For example, the bird targets and bionic wing targets may be caused by the flapping motion of their wings, while the UAV targets are caused by the rotation of their rotors.

[0052] b. Based on the time domain data and Doppler domain data of the low-altitude target, extract the feature vector of the low-altitude target from at least two different feature domains respectively, and fuse the feature vectors from different feature domains to construct a multi-domain feature dataset; c. Use the multi-domain feature dataset to train a preset random forest classifier to obtain the trained random forest classifier.

[0053] In one specific embodiment of this application, c includes: c1, the multi-domain feature dataset is divided into a training set and a test set according to a preset ratio; c2 sets the model parameters for the random forest classifier, including the number of decision trees, the number of candidate features randomly selected when splitting each tree node, and the maximum growth depth of each decision tree; Random forest, proposed by Breiman, is a nonlinear classification model based on ensemble learning theory. Its core idea is an ensemble strategy of "parallel construction of multiple decision trees (DTs) + voting fusion" to compensate for the weak generalization ability and overfitting of single decision trees, thereby achieving high accuracy and robustness in classification tasks. A schematic diagram of a decision tree is shown below. Figure 6 As shown.

[0054] The Random Forest algorithm uses decision trees as base learners. By introducing sample randomness and feature randomness, it maintains strong independence for each decision tree. Then, through ensemble fusion, it improves the overall performance of the model, making it suitable for real-world datasets with high-dimensional, redundant, and noisy multi-domain features of low-altitude targets. A schematic diagram of the Random Forest model structure is shown below. Figure 7 As shown.

[0055] c3 uses a bootstrap resampling technique to extract multiple sample subsets with replacement from the training set. c4, For each of the sample subsets, randomly select a feature subset from all its features, and use the sample subset and the corresponding feature subset to independently train and generate a decision tree; c5. Repeat steps c3 to c4 until a preset number of decision trees are generated, which together form the initial trained random forest classifier. c6. The initial trained random forest classifier is evaluated using the test set, and the random forest classifier that passes the evaluation is determined as the trained random forest classifier.

[0056] Specifically, the training steps for the random forest classifier model are as follows: (1) From the original training set In this process, sampling with replacement is used to generate... Different training subsets .

[0057] (2) For the first A decision tree, when splitting at each node, splits from the whole... Randomly selected from 1 feature The system identifies several characteristics and selects the optimal splitting method based on a certain splitting criterion.

[0058] (3) Repeat the above process until each decision tree grows to its maximum depth or meets the stopping condition, and finally obtain the result. A series of independent decision trees together form a random forest classifier.

[0059] Specifically, the testing steps for the random forest classifier are as follows: (1) For each sample to be classified, input it into each decision tree in the random forest classifier.

[0060] (2) Each decision tree in a random forest classifies the feature samples based on its own structure and outputs a classification result (vote).

[0061] (3) The classification results of all decision trees are obtained by using the "majority voting rule". The category with the most votes is the predicted label of the sample.

[0062] The above process reveals the following advantages of the random forest classifier: (1) Strong resistance to overfitting: Random forests reduce the variance of the training model by introducing sample randomness and feature randomness, effectively reducing the over-dependence of individual trees on local samples.

[0063] (2) Excellent ability to process high-dimensional data: Through random feature selection, random forest can process high-dimensional features without dimensionality reduction of features.

[0064] (3) Easy to parallelize: The training process of each decision tree is independent, which is suitable for parallel computing using multi-core CPUs, thus greatly shortening the training time.

[0065] Based on the above model construction and testing steps, the obtained feature dataset is randomly divided into training and testing sets in a 7:3 ratio. Next, the parameters of the random forest model are initialized, determining the number of decision trees, the number of candidate features for each decision tree node split, and the maximum tree depth. Then, the model is trained using the training set, generating several training subsets through bootstrap resampling. Each tree is assigned an independent training subset and a random candidate feature set, thus constructing multiple independent decision trees, each independently completing node splitting and growth. Finally, after all decision trees have been trained, the ensemble fusion strategy is determined through majority voting to form the final random forest classification model.

[0066] In one specific embodiment of this application, c6 includes: c61, input the multi-domain feature vector of the test set into the initially trained random forest classifier, so that each decision tree in the random forest classifier independently classifies each test sample and outputs a predicted category; c62, for each test sample, count the prediction results of all decision trees, and use the majority voting principle to determine the final prediction category of the sample; c63, calculate the classification performance evaluation index based on the true category of all test samples and the final predicted category; During the testing phase, the reserved test set feature vectors are input into the trained random forest model. Each tree in the model independently classifies the test set and outputs its own prediction results. Then, the prediction results of all decision trees are fused according to the majority voting rule to obtain the final classification label of each test sample, and the confusion matrix of the classification results is calculated. Finally, the model's classification accuracy, recall, F1 score and other quantitative evaluation indicators are calculated to comprehensively evaluate the model's performance.

[0067] Since this experimental dataset contains three types of targets, the final classification can be obtained as follows: confusion matrix Before providing the classification performance evaluation metrics, the following definitions are made for some symbols: Row index Real category ( As a biomimetic wing, For drones, (For birds) Column index Predicted category ( (The corresponding category should be consistent with the actual category) · The truth is the first Class, prediction is the first Number of samples per class · : No. The true total number of samples of the target class · Predicted as the first Total number of samples of the target class · Total number of experimental samples The following are three classification performance metrics: accuracy, recall, and F1 score: Classification accuracy (0-28) Classification accuracy characterizes the model's ability to correctly classify all samples as a whole, and is the most intuitive global performance indicator.

[0068] Recall (1) Single-class target recall rate (based on the first class) (Taking class targets as an example) (0-29) Single-class recall rate represents the ability of a class of targets to be correctly classified, and its essence is the probability of "not missing any targets".

[0069] (2) Overall macro average recall (0-30) The overall macro average recall treats all three target classes equally, reflecting the model's balanced recognition ability across all types, and avoiding the situation where superior performance in one class masks the shortcomings of other classes.

[0070] F1 value The F1 score is the harmonic mean of precision and recall. Precision for each target class can be calculated first. (0-31) The precision of a single-class target is defined as the proportion of samples predicted to be of a certain class that are actually of that class.

[0071] (1) F1 score of a single target (in order of magnitude) (Taking class targets as an example) (0-32) Single-class target The value represents the ability to "not miss or misjudge" a certain type of target.

[0072] (2) Overall macro average F1 value (0-33) Overall Macro Average The value is obtained by comprehensively evaluating three types of objectives. Performance reflects the model's balanced optimization effect across all categories.

[0073] c64 identifies random forest classifiers that meet the classification performance evaluation criteria as successfully trained random forest classifiers.

[0074] The low-altitude target dataset collected in this chapter is based on Ku-band pulse-Doppler radar. The data includes three types of low-altitude targets: birds, bionic wings, and drones (DJI Mini 4 Pro), all flying within a 2.5km range. Data acquisition was conducted under the same radar parameters, as shown in Table 1. There are 823 sets of bird target data, 824 sets of bionic wing target data, and 878 sets of drone target data. The training and test sets for each target type are divided in a 7:3 ratio. When performing STFT on the target time-domain signals, the window size is set to 128, the Hanning window function is used, and the number of overlap points is 125.

[0075] Table 1 Radar Operating Parameters

[0076] Multi-domain feature extraction was performed on the training and test sets of three target classes: birds, biomimetic wings, and drones, according to the aforementioned feature extraction principles. This yielded 18 features across the time, Doppler, time-frequency, and wavelet domains. A random forest classifier was used for both training and testing, with 50 decision trees in the random forest. This paper uses classification accuracy, overall recall, and overall F1 score as the core evaluation metrics. To verify the impact of multi-domain features, different feature combinations, and different classifiers on classification performance, the classification results of single-domain features, multi-domain feature combinations, and all features under different classifiers were statistically analyzed. To ensure the reliability of the experimental results, each result was obtained by averaging 50 Monte Carlo experiments.

[0077] This application presents statistics on classification performance from two aspects: single-domain feature classification performance and multi-domain feature combination classification performance. First, experiments were conducted on four types of single-domain features: time domain, Doppler domain, time-frequency domain, and wavelet domain. The overall classification index corresponding to each domain feature was statistically obtained, and the results are shown in Table 2.

[0078] Table 2. Statistical results of classification performance for single-domain features

[0079] The overall recall and overall retrieval rates in Table 2 are both calculated using macro averaging. The statistical results in Table 2 show that, among single-domain features, Doppler domain features exhibit the best overall classification performance, with accuracy, recall, and F1 score all exceeding 90%. Wavelet domain features are second best, while time-frequency domain features and time-domain features are relatively poor. Figure 8 The statistical results of the confusion matrix for single-domain features are presented in probabilistic form, revealing significant confusion between drones and birds. Overall, single-domain features are heavily limited by feature dimensionality, resulting in a clear bottleneck in classification performance and making it impossible to accurately classify all three target classes simultaneously.

[0080] To overcome the limitations of single-domain features, this paper designs multi-domain feature combination schemes based on the "optimal complementarity" criterion, encompassing pairwise combinations, three-class combinations, and full four-domain fusion. Through systematic classification comparison experiments, the classification effectiveness of different combination strategies is quantitatively evaluated, thereby verifying the effectiveness of feature fusion strategies in improving model performance.

[0081] (1) Random combination of two-domain features The classification performance statistics of the two-domain feature combination are shown in Table 3. Compared with single-domain features, it can be found that the classification accuracy of the two-domain feature combination is above 90%, and the overall classification accuracy of the three feature combinations of "time domain + wavelet domain", "Doppler domain + time-frequency domain" and "Doppler domain + wavelet domain" can reach above 95%. Figure 9 The statistical results of the confusion matrix of the two-domain feature combination are shown. It can be found that the numerical proportion of its diagonal elements is significantly improved compared with the single domain, that is, the classification performance of the three types of targets is significantly improved.

[0082] Table 3. Statistical results of classification performance of two-domain feature combinations

[0083] The classification performance statistics of random combination of three-domain features are shown in Table 4. Compared with the experimental results of combination of two-domain features, the classification accuracy of combining three-domain features can reach more than 95%. Among them, the overall classification accuracy of the combination of the three feature domains "time domain + Doppler domain + time-frequency domain" can reach 96.50%. Figure 10 The statistical results of the confusion matrix of random combination of three domain features are presented. The results show that the three types of targets have achieved good classification results, but there is still some confusion between the two types of targets: bionic wings and drones.

[0084] Table 4. Statistical results of classification performance of three-domain feature combinations

[0085] The experiment combines all domain features and the classification performance statistics are shown in Table 5. As can be seen from the table, the overall accuracy of the four-domain feature fusion reaches 98.28%, the recall reaches 98.26%, and the F1 score reaches 98.25%, which are all better than the previous results of single-domain features, random combination of two-domain features, and random combination of three-domain features, thus verifying the effectiveness of multi-domain feature fusion. Figure 11 The statistical results of the confusion matrix of the four-domain full feature combination are presented. The results show that multi-domain feature fusion can effectively reduce the cross-misclassification rate and significantly improve the classification performance of three types of low-altitude targets, basically realizing the accurate classification requirements of birds, drones and bionic wings.

[0086] Table 5. Statistical results of classification performance of four-domain full feature combinations

[0087] It is worth noting that the terms "first" and "second" in this application are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0088] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of this application and should not be construed as limiting the specific implementation of this application to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of this application, and all such modifications or substitutions should be considered within the scope of protection of this application.

Claims

1. A low-altitude target intelligent classification method based on multi-domain feature fusion, characterized in that, include: S100: Acquire measured echo data of the radar against low-altitude targets, and preprocess the echo data to obtain the target's time-domain data and Doppler-domain data. S200, based on the time-domain data and Doppler-domain data, extract feature vectors of the target from at least two different feature domains respectively, and fuse the feature vectors from different feature domains to obtain a multi-domain feature vector; wherein, the feature domains include at least two of the time domain, Doppler domain, time-frequency domain and wavelet domain; S300, the multi-domain feature vector is input into the trained random forest classifier, and the category of the low-altitude target is output.

2. The intelligent classification method for low-altitude targets based on multi-domain feature fusion according to claim 1, characterized in that, S100 includes: S110: Acquires measured echo data of the radar against low-altitude targets; S120, the measured echo data is sequentially subjected to pulse compression, clutter suppression, moving target detection and constant false alarm rate detection to obtain the time domain data and Doppler domain data.

3. The intelligent classification method for low-altitude targets based on multi-domain feature fusion according to claim 1, characterized in that, S200 includes: S210, based on the time-domain data and the Doppler-domain data, select feature vectors from at least two feature domains in the time domain, Doppler domain, time-frequency domain and wavelet domain; S220 fuses feature vectors from different feature domains to obtain multi-domain feature vectors.

4. The intelligent classification method for low-altitude targets based on multi-domain feature fusion according to claim 3, characterized in that, When the feature domain includes the time domain, the time domain feature extraction in S220 includes: extracting the mean, variance, envelope standard deviation, average peak value, peak standard deviation, and time domain waveform entropy from the time domain data.

5. The intelligent classification method for low-altitude targets based on multi-domain feature fusion according to claim 4, characterized in that, When the feature domain includes the Doppler domain, the Doppler feature extraction in S220 includes: extracting the amplitude variance, first-order origin moment, second-order center distance, L1 norm of normalized amplitude, Doppler spectrum differential mode after normalization amplitude, and Doppler spectrum waveform entropy from the Doppler domain data.

6. The intelligent classification method for low-altitude targets based on multi-domain feature fusion according to claim 4, characterized in that, When the feature domain includes the time-frequency domain, the time-frequency domain feature extraction in S200 includes: performing a short-time Fourier transform on the time-domain data to obtain a time-frequency power matrix; and extracting spectral entropy, micro-Doppler clustering degree, and singular value entropy based on the time-frequency power matrix.

7. The intelligent classification method for low-altitude targets based on multi-domain feature fusion according to claim 4, characterized in that, When the feature domain includes the wavelet domain, wavelet domain feature extraction in S200 includes: performing discrete wavelet transform on the time domain data to obtain detail coefficients and approximation coefficients at multiple scales; and extracting relative wavelet energy, detail coefficient kurtosis, and multi-scale log-variance based on the detail coefficients and approximation coefficients.

8. The intelligent classification method for low-altitude targets based on multi-domain feature fusion according to claim 4, characterized in that, The training process of the completed random forest classifier includes: a. Acquire radar echo data of three types of low-altitude targets: birds, drones, and biomimetic wings, and preprocess the echo data to obtain time-domain and Doppler-domain data of the low-altitude targets. b. Based on the time domain data and Doppler domain data of the low-altitude target, extract the feature vector of the low-altitude target from at least two different feature domains respectively, and fuse the feature vectors from different feature domains to construct a multi-domain feature dataset; c. Use the multi-domain feature dataset to train a preset random forest classifier to obtain the trained random forest classifier.

9. The intelligent classification method for low-altitude targets based on multi-domain feature fusion according to claim 1, characterized in that, c includes: c1, the multi-domain feature dataset is divided into a training set and a test set according to a preset ratio; c2 sets the model parameters for the random forest classifier, including the number of decision trees, the number of candidate features randomly selected when splitting each tree node, and the maximum growth depth of each decision tree; c3 uses a bootstrap resampling technique to extract multiple sample subsets with replacement from the training set. c4, For each of the sample subsets, randomly select a feature subset from all its features, and use the sample subset and the corresponding feature subset to independently train and generate a decision tree; c5. Repeat steps c3 to c4 until a preset number of decision trees are generated, which together form the initial trained random forest classifier. c6. The initial trained random forest classifier is evaluated using the test set, and the random forest classifier that passes the evaluation is determined as the trained random forest classifier.

10. The intelligent classification method for low-altitude targets based on multi-domain feature fusion according to claim 9, characterized in that, c6 includes: c61, input the multi-domain feature vector of the test set into the initially trained random forest classifier, so that each decision tree in the random forest classifier independently classifies each test sample and outputs a predicted category; c62, for each test sample, count the prediction results of all decision trees, and use the majority voting principle to determine the final prediction category of the sample; c63, calculate the classification performance evaluation index based on the true category of all test samples and the final predicted category; c64 identifies random forest classifiers that meet the classification performance evaluation criteria as successfully trained random forest classifiers.

Images