Single-face array low-altitude radar target identification method and system based on deep learning

Abstract

The invention discloses a single-face array low-altitude radar target recognition system based on deep learning, and the system comprises a data collection module which obtains the radar echo data of different targets in various environments, a receiver has a specific sampling frequency and a dynamic range, the collection time is set according to the characteristics of the targets, and the data is compressed and stored; the data preprocessing module is used for denoising, clutter suppression and normalization; the feature extraction module extracts features by using CNN; the target recognition module classifies and recognizes a target; the model training and optimizing module is used for training a model and optimizing parameters; and the visual display module displays the result. According to the invention, data can be accurately collected and efficiently preprocessed, target features are accurately extracted and recognized by using deep learning, the model is optimized to improve the accuracy, the result can be visually displayed, the accuracy, real-time performance and reliability of low-altitude radar target recognition are effectively improved, and the monitoring requirements are met.

Description

Technical Field

[0001] The present invention relates to the technical field of low-altitude radar target recognition, and in particular to a single-array low-altitude radar target recognition method and system based on deep learning. Background Art

[0002] With the continuous development of science and technology, activities in the low-altitude field are becoming more and more frequent, including civil drones, small aircraft flying at low altitudes, etc. The monitoring and identification of these low-altitude targets are of vital importance to national security, aviation safety, and urban management. However, traditional low-altitude radar target recognition methods face many challenges.

[0003] In the past, low-altitude radar target recognition mainly relied on manual analysis and simple signal processing technology. Operators observed the echo signal on the radar screen and judged the type and location of the target based on experience. This method is not only inefficient, but also easily affected by human factors, and its accuracy and reliability are difficult to guarantee. At the same time, traditional signal processing algorithms are usually based on fixed feature extraction and pattern recognition methods, which have poor adaptability to complex and changeable low-altitude environments and target types.

[0004] As the diversity and complexity of low-altitude targets continue to increase, traditional methods are unable to cope with the following problems. First, different types of targets have different radar reflection characteristics. For example, the echo signals of metal targets and non-metal targets are quite different. In addition, the target's motion state and posture changes will also have a significant impact on the echo signal. Traditional methods are difficult to effectively extract these complex feature information, resulting in low target recognition accuracy.

[0005] Secondly, there is a lot of clutter interference in the low-altitude environment. For example, the echo signals reflected by buildings, terrain, and flying birds are intertwined with the target echo signals, making it more difficult to extract and identify the target signal. Traditional clutter suppression methods often have limited effects and cannot completely eliminate the influence of clutter, thereby reducing the performance of target detection and recognition.

[0006] In addition, the number and density of low-altitude radar targets are increasing, which puts higher requirements on the real-time performance and processing capabilities of the target recognition system. Traditional methods are slow in processing large amounts of data and cannot meet the needs of real-time monitoring and rapid response.

[0007] In order to solve these problems, deep learning technology has gradually been introduced into the field of low-altitude radar target recognition. Deep learning has powerful automatic feature extraction and pattern recognition capabilities, and can learn the deep-level features of the target from a large amount of data, improving the accuracy and reliability of target recognition. At the same time, deep learning algorithms can adapt to different low-altitude environments and target types through continuous training and optimization, and have strong generalization capabilities.

[0008] However, the current low-altitude radar target recognition method based on deep learning still has some shortcomings. For example, there may be problems such as low data quality and sample imbalance during data collection; the training of deep learning models requires a large amount of labeled data and computing resources, and the training time is long; in practical applications, the stability and real-time performance of the system need to be further improved.

[0009] Therefore, it has become an urgent need in the current low-altitude monitoring field to develop a single-array low-altitude radar target recognition method and system based on deep learning to effectively solve the above problems and improve the accuracy, real-time and reliability of low-altitude radar target recognition. Summary of the invention

[0010] The present invention proposes a single-array low-altitude radar target recognition method and system based on deep learning to solve the problems mentioned in the above-mentioned prior art.

[0011] In order to achieve the above object, the present invention adopts the following technical solution: a single-array low-altitude radar target recognition system based on deep learning, comprising:

[0012] The data acquisition module is used to obtain single-array low-altitude radar echo signal data, including echo data of different types of targets in various environments. The acquisition equipment is a radar receiver with high-precision sampling rate and wide dynamic range. The sampling frequency range is 100MHz to 500MHz, and the dynamic range is 60dB to 100dB. The data acquisition time interval is set according to the target motion characteristics and monitoring requirements. For low-speed moving targets, such as small drones flying at a speed of less than 5m / s, the acquisition time interval can be appropriately extended to 2 seconds; for high-speed targets, such as jet aircraft flying at a speed of 500m / s, the acquisition time interval is shortened to 0.1 seconds. The storage format is a specific binary format, and the LZ77 compression algorithm is used for data compression to reduce storage space occupancy and improve data transmission efficiency.

[0013] The data preprocessing module preprocesses the collected radar echo signal data, including denoising, clutter suppression and data normalization. Adaptive filtering algorithm is used for denoising, and the filter parameters are automatically adjusted according to the power spectral density characteristics of the noise. For Gaussian noise, Wiener filtering is used to calculate the optimal filter coefficient by estimating the power spectral density of the signal and noise. The Wiener filtering formula is: Where H(f) is the filter frequency response, S s (f) is the signal power spectrum density, S n(f) is the noise power spectral density. For salt and pepper noise, median filtering is used to sort the pixel values ​​in the window and take the median as the output. For mixed noise, adaptive weighted median filtering is used to dynamically adjust the median filter window size and weight according to the noise type and intensity. Clutter suppression is performed based on the constant false alarm rate CFAR detection technology, and the number of reference units and the number of protection units are dynamically adjusted according to the clutter environment. In the marine environment, due to the reflection of waves, the clutter is strong. The number of reference units can be set to 16 to 24, and the number of protection units can be set to 4 to 6; in the urban environment, the reflection of buildings will produce clutter with different characteristics. The number of reference units is adjusted to 12 to 20, and the number of protection units is adjusted to 3 to 5. The data is normalized to the [0,1] interval by linear or standard deviation normalization method. If the data distribution is approximately normal, standard deviation normalization is used. The calculation formula is: Where x is the original data, μ is the data mean, and σ is the data standard deviation; if the data distribution range is narrow, linear normalization is used, and the calculation formula is: Where x is the original data, x min is the minimum value of the data, x max is the maximum value of the data.

[0014] The feature extraction module uses a deep learning model to automatically extract features from preprocessed data. The deep learning model is built based on a convolutional neural network (CNN), which contains multiple convolutional layers, pooling layers, and fully connected layers. The convolution kernel size ranges from 3×3 to 5×5, and the convolution layer can use different activation functions, including ReLU, LeakyReLU, or ELU. The ReLU activation function formula is: f(x) = max(0, x), the LeakyReLU activation function formula is: f(x) = max(0.1x, x), and the ELU activation function formula is: Where α is an adjustable parameter, usually set to 1. The pooling layer uses maximum or average pooling operations with a step size of 2 to 3. The pooling window size is adjusted in the range of 2×2 to 4×4 according to data characteristics and network structure requirements. The number of neurons in the fully connected layer is determined by the target category and feature dimension. A Dropout layer can be added with a Dropout probability set to 0.2 to 0.5. The Dropout operation randomly sets the output of some neurons to 0 during training to prevent overfitting.

[0015] The target recognition module classifies and recognizes the target based on the extracted features. It uses support vector machine SVM, random forest RF or deep neural network DNN classification algorithm, and the algorithm parameters are optimized through cross-validation and grid search methods. For SVM, different kernel functions are used, such as the linear kernel function formula: K(x, y) = x T y, the radial basis kernel function formula is: K(x, y) = exp(-γ||xy|| 2), where γ is the kernel function parameter. The classification performance is optimized by adjusting the kernel function parameter and penalty coefficient. For random forest, the number of decision trees is set to 50 to 100, and the classification accuracy is improved by adjusting the depth and splitting criteria of the decision tree. For deep neural network DNN, the number of network layers, the number of neurons and the activation function parameters are adjusted. The recognition results include target type, position, speed and motion trajectory prediction. The target position is calculated using the formula: Where P(x, y, z) is the target position coordinate, ω i is the weight coefficient, R i (x, y, z) is the position information component provided by the i-th feature or sensor, and n is the number of features or sensors involved in the position calculation. The target speed is calculated using the formula: Where V is the target speed, f j is the influencing factor of speed-related characteristics, S j is the th velocity-related eigenvalue, and Δt is the time interval. The motion trajectory prediction uses the Kalman filter algorithm, which is based on historical data and the target motion model.

[0016] The model training and optimization module uses a large amount of labeled data to train the deep learning model and continuously optimizes the model parameters to improve the recognition accuracy. The training data is labeled by professionals based on the real attributes of the target, and the labeled information includes the target type, position, and speed. The back propagation algorithm is used to adjust the model weights, and the optimization methods include stochastic gradient descent SGD, Adagrad, and Adadelta. In SGD, the learning rate adjustment strategy is exponential decay, and the learning rate formula is: η t =η 0 β t , where η t is the current learning rate, η 0 is the initial learning rate, β is the decay coefficient, and t is the number of iterations. In Adagrad, the adaptive learning rate calculation formula is: where η t is the current learning rate, η 0 is the initial learning rate, g i is the gradient of the i-th iteration. In Adadelta, the update rule is: where Δx t is the parameter update amount, g i is the gradient of the i-th iteration, d i is the parameter update amount of the i-th iteration, and ρ is the new attenuation coefficient. At the same time, a regularization term is introduced to prevent overfitting. The regularization term uses L2 regularization, and the calculation formula is: Where F is the loss function after regularization, F 0 is the original loss function, k is the regularization coefficient, u i is the model parameter.

[0017] The visualization module displays the target position and trajectory of the target recognition results in two-dimensional or three-dimensional radar images, and uses OpenGL technology to achieve image rendering and interaction. The chart shows the target speed and type distribution, and provides interactive functions. Users can view detailed information such as the specific feature vector and historical trajectory of the target by clicking or touching the mouse.

[0018] Furthermore, the radar receiver antenna of the data acquisition module has a specific polarization mode, including horizontal polarization, vertical polarization or circular polarization, which can automatically switch or simultaneously receive echo signals of multiple polarization modes according to target characteristics and environmental conditions, thereby improving the ability to obtain target information. For example, for metal targets, horizontal polarization may be more conducive to obtaining clear echo signals; for non-metal targets, circular polarization may be more effective. In complex environments, echo signals of multiple polarization modes are received simultaneously, and the accuracy of target detection and recognition is improved through fusion algorithms.

[0019] Furthermore, in the denoising process, the data preprocessing module automatically selects a suitable filtering algorithm according to the noise type, Gaussian noise uses Wiener filtering, salt and pepper noise uses median filtering, and mixed noise uses adaptive weighted median filtering; in the clutter suppression process, the number of reference units and the number of protection units of the CFAR detection technology are dynamically adjusted according to the clutter environment, the number of reference units ranges from 8 to 32, and the number of protection units ranges from 2 to 8. The specific adjustment strategy is as described in claim 1.

[0020] Furthermore, in the deep learning model of the feature extraction module, the convolution layer can use different activation functions, including ReLU, LeakyReLU or ELU, the pooling window size of the pooling layer can be adjusted in the range of 2×2 to 4×4 according to the data characteristics and network structure requirements, and the full connection layer can add a Dropout layer, and the Dropout probability is set to 0.2 to 0.5. The specific technical details are as described in claim 1.

[0021] Furthermore, a single-array low-altitude radar target recognition method based on deep learning includes the following steps:

[0022] The data collection step uses a single-array low-altitude radar receiver to collect echo signal data of different types of targets under various environmental conditions, and sets the sampling frequency and acquisition time according to the target movement speed and monitoring range. The sampling frequency for low-speed targets is 100Hz to 200Hz, and the sampling frequency for high-speed targets is 500Hz to 1000Hz. For targets with a distance of less than 10 kilometers, the sampling frequency can be appropriately reduced to 300Hz; for targets with a distance of more than 50 kilometers, the echo signal strength can be increased by increasing the radar transmission power, extending the integration time, or using multi-pulse accumulation technology. The transmission power is increased by 5dB on the original basis, the integration time is extended to 50ms, and the number of multi-pulse accumulations is 8. The acquisition time is adjusted within the range of 1 second to 10 seconds, and is dynamically adjusted according to the target movement speed and monitoring needs.

[0023] The data preprocessing step is to perform denoising, clutter suppression and normalization on the collected radar echo signal data. The specific processing method is as described in claim 1.

[0024] In the feature extraction step, the preprocessed data is input into the deep learning model to automatically extract the feature vector that represents the target characteristics. The dimension of the feature vector is determined by the model structure and the complexity of the target, and is between 128 and 1024 dimensions. The deep learning model selects the appropriate network structure according to the complexity of the target recognition task and the characteristics of the data. A 3- to 5-layer convolutional neural network structure is used for simple targets, and an 8- to 15-layer convolutional neural network architecture is used for complex targets or multi-class target recognition. At the same time, transfer learning technology is used to accelerate model training and improve recognition effects using the pre-trained model VGG16. The parameters of the pre-trained model are adjusted and new layers are added to adapt to specific low-altitude radar target recognition tasks.

[0025] In the target recognition step, based on the extracted feature vector, a classification algorithm is used to classify and identify the target, output the target type, position, and speed information, and predict the target motion trajectory based on historical data and the target motion model. The classification algorithm can use a weighted loss function during the training process to give different weights to the misclassification of different types of targets, with the focus on setting the target weight to 3. Post-processing techniques, including trajectory association and filter smoothing techniques, are combined to improve the stability and accuracy of target recognition. The nearest neighbor algorithm is used for trajectory association, and the Kalman filter algorithm is used for filter smoothing techniques.

[0026] In the model training and optimization steps, a large amount of labeled sample data is used to train the deep learning model. By adjusting the model parameters and optimizing the algorithm, the recognition accuracy and generalization ability of the model are improved. The training data can use data enhancement technology, including random rotation angles between -5° and 5°, flipping (horizontal or vertical), and scaling between 0.9 and 1.1. The data set size is expanded to improve the model's ability to recognize targets in different postures and environments. The optimization algorithm can dynamically adjust the learning rate according to the convergence of the model training process. The initial learning rate can be set to 0.005. When the training error no longer decreases for 10 consecutive epochs, the learning rate is reduced to 0.3 times the original.

[0027] The visualization step presents the target recognition results to the user in a visual manner, providing the user with intuitive and clear target information to assist in decision-making.

[0028] Furthermore, in the data collection step, for targets with a distance greater than 50 kilometers or a radar reflection cross-sectional area less than 1 square meter, the radar transmission power is increased, the integration time is extended, or the multi-pulse accumulation technology is used to improve the echo signal strength to ensure the effective collection of target information. The transmission power is increased by 5dB on the original basis, the integration time is extended to 50ms, and the number of multi-pulse accumulations is 8.

[0029] Furthermore, in the data preprocessing step, the normalization method can be selected according to the characteristics of the data distribution. If the data is approximately normally distributed, standard deviation normalization is used. If the data distribution range is narrow, linear normalization is used to make the data more stable and consistent in subsequent processing.

[0030] Furthermore, in the feature extraction step, the deep learning model selects a suitable network structure according to the complexity of the target recognition task and the data characteristics. A 3- to 5-layer convolutional neural network structure is used for simple targets, and an 8- to 15-layer convolutional neural network architecture is used for complex targets or multi-category target recognition. At the same time, transfer learning technology is used to utilize pre-trained models to accelerate model training and improve recognition effects.

[0031] Furthermore, in the target recognition step, the classification algorithm can adopt a weighted loss function during the training process, giving different weights to the misclassification of different types of targets, and setting the weight of the key target to 3 to improve the recognition accuracy of the key target. At the same time, post-processing techniques, including trajectory association and filtering and smoothing techniques, are combined to improve the stability and accuracy of target recognition.

[0032] Furthermore, in the model training and optimization steps, data enhancement technology can be used for training data, including random rotation angles between -5° and 5°, flipping (horizontal or vertical), and scaling ratios between 0.9 and 1.1, to expand the size of the data set and improve the model's ability to recognize targets in different postures and environments. The optimization algorithm can dynamically adjust the learning rate according to the convergence of the model training process. The initial learning rate can be set to 0.005. When the training error no longer decreases for 10 consecutive epochs, the learning rate is reduced to 0.3 times the original.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] First, in terms of data acquisition, the use of radar receivers with high-precision sampling rate and wide dynamic range can accurately obtain echo signals of different types of targets in various environments. By setting the acquisition time interval according to the target motion characteristics and monitoring requirements, it can not only ensure a rapid response to high-speed targets, but also fully capture the characteristics of low-speed targets. The storage format uses a specific binary format combined with a data compression algorithm to reduce storage space occupancy, improve data transmission efficiency, and lay the foundation for subsequent rapid processing.

[0035] Secondly, the data preprocessing module effectively removes noise and clutter interference. It automatically selects appropriate filtering algorithms for different types of noise to ensure the purity of the data. The constant false alarm rate CFAR detection technology dynamically adjusts the number of reference units and protection units to adapt to different clutter environments and improves the accuracy of target detection. Data normalization processing makes the data more stable and consistent in subsequent processing, providing high-quality input data for the deep learning model.

[0036] In terms of feature extraction and target recognition, the deep learning model is built on the convolutional neural network (CNN), which can automatically extract deep features from preprocessed data. The convolutional layer uses different activation functions, the pooling layer flexibly adjusts the window size and step size, and the fully connected layer adds a Dropout layer, which improves the performance and generalization ability of the model. Using a variety of classification algorithms and optimizing parameters through methods such as cross-validation and grid search, it can accurately classify and identify targets and output rich information such as target type, position, speed, and motion trajectory prediction.

[0037] Furthermore, the model training and optimization module continuously optimizes model parameters and improves recognition accuracy. Stochastic gradient descent SGD, Adagrad, Adadelta and other optimization methods and the introduction of L2 regularization terms prevent overfitting and make the model more stable and reliable. Data enhancement technology expands the size of the data set and improves the model's ability to recognize targets in different postures and environments.

[0038] Finally, the visualization module provides users with intuitive and clear target information. Two-dimensional or three-dimensional radar images show the target position and trajectory, and charts show the target speed and type distribution, which is convenient for users to make decisions and analysis. At the same time, the interactive function enables users to easily view detailed information, which improves the ease of use and practicality of the system. In short, the present invention provides an efficient, accurate and reliable solution for low-altitude radar target identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 A schematic block diagram of a single-array low-altitude radar target recognition system based on deep learning proposed by the present invention;

[0040] Figure 2 This is a schematic block diagram of the single-array low-altitude radar target recognition method based on deep learning proposed in the present invention. DETAILED DESCRIPTION

[0041] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0042] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise" and "counterclockwise" indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0043] In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined. In addition, the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, and it can be the internal connection of two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances. The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0044] Reference Figure 1-2 :A single-array low-altitude radar target recognition system based on deep learning, including:

[0045] The data acquisition module plays a vital role in the single-array low-altitude radar target recognition system based on deep learning. It is mainly responsible for obtaining the single-array low-altitude radar echo signal data, covering the echo information of different types of targets in a variety of complex environments.

[0046] In order to ensure the acquisition of high-quality data, this module uses a radar receiver with high-precision sampling rate and wide dynamic range. Its sampling frequency range is set between 100MHz and 500MHz. Such a wide frequency range can capture the subtle features of various targets. The dynamic range is 60dB to 100dB, which enables the receiver to accurately distinguish target signals from echo signals of different strengths, and can effectively collect both weak small target echoes and strong large target echoes.

[0047] The time interval for data collection is not fixed, but is flexibly set according to the motion characteristics of the target and the specific monitoring needs. For example, for low-speed moving targets, such as small drones flying at speeds below 5m / s, the collection time interval can be appropriately extended to 2 seconds because its movement is relatively slow and the changes are not too drastic. This ensures that the echo information of the target is fully obtained without generating too much data redundancy. For high-speed targets, such as jet aircraft flying at a speed of 500m / s, their states change rapidly, and more frequent data collection is required to accurately track the target. At this time, the collection time interval is shortened to 0.1 seconds to ensure that changes in key information such as the target's position and speed can be captured in a timely manner.

[0048] In terms of data storage, a specific binary format is used. This format is efficient and compact, and can minimize the space required for data storage. At the same time, in order to further reduce storage space and improve data transmission efficiency, this module uses the LZ77 compression algorithm. The LZ77 algorithm compresses data by identifying repeated patterns in the data, greatly reducing the amount of data storage and transmission time without losing important information. In this way, in practical applications, both data storage and transmission between different devices are more convenient and efficient, providing strong support for subsequent data processing and analysis.

[0049] The data preprocessing module plays a key role in the entire single-array low-altitude radar target recognition system based on deep learning. Its core function is to perform preprocessing operations on the collected radar echo signal data. These operations mainly cover three important aspects: denoising, clutter suppression, and data normalization. Each aspect plays an indispensable role in the subsequent target recognition process.

[0050] In the denoising phase, this module uses an adaptive filtering algorithm to remove noise from radar echo signal data. This adaptive filtering algorithm is highly intelligent and can automatically adjust the filter parameters based on the power spectral density characteristics of the noise. For different types of noise, whether it is colored noise with a specific frequency distribution or white noise uniformly distributed in the entire frequency domain, the algorithm can accurately analyze its power spectral density. For example, when encountering Gaussian white noise, the algorithm quickly adjusts the filter parameters by calculating and analyzing its power spectral density, so that the filter can effectively suppress this noise while retaining the effective information of the target echo signal as much as possible. This ability to dynamically adjust parameters according to noise characteristics greatly improves the effect and efficiency of denoising compared to traditional fixed parameter filtering methods, creating a purer data environment for subsequent signal processing and target recognition.

[0051] In terms of clutter suppression, the module uses advanced technical means to effectively identify and suppress clutter from various sources. These clutter may come from natural objects in the environment, such as terrain undulations, vegetation reflections, or reflections from other artificial facilities. Through complex algorithms and models, the data preprocessing module can distinguish between clutter signals and target echo signals, thereby reducing the interference of clutter on target identification.

[0052] Data normalization is also one of the important functions of this module. It can normalize the radar echo signal data after denoising and clutter suppression, so that the value range and distribution characteristics of the data meet the requirements of the subsequent deep learning model, thereby improving the training efficiency and recognition accuracy of the model. Through this series of preprocessing operations, the quality of radar echo signal data has been significantly improved, laying a solid foundation for the stable operation and accurate recognition of the entire target recognition system.

[0053] For Gaussian noise, Wiener filtering is used to calculate the optimal filter coefficient by estimating the power spectral density of the signal and noise. The Wiener filtering formula is: Where H(f) is the filter frequency response, S s (f) is the signal power spectrum density, S n (f) is the noise power spectral density. For salt and pepper noise, median filtering is used to sort the pixel values ​​in the window and take the median as the output. For mixed noise, adaptive weighted median filtering is used to dynamically adjust the median filter window size and weight according to the noise type and intensity. Clutter suppression is performed based on the constant false alarm rate CFAR detection technology, and the number of reference units and the number of protection units are dynamically adjusted according to the clutter environment. In the marine environment, due to the reflection of waves, the clutter is strong. The number of reference units can be set to 16 to 24, and the number of protection units can be set to 4 to 6; in the urban environment, the reflection of buildings will produce clutter with different characteristics. The number of reference units is adjusted to 12 to 20, and the number of protection units is adjusted to 3 to 5. The data is normalized to the [0,1] interval by linear or standard deviation normalization method. If the data distribution is approximately normal, standard deviation normalization is used. The calculation formula is: Where x is the original data, μ is the data mean, and σ is the data standard deviation; if the data distribution range is narrow, linear normalization is used, and the calculation formula is: Where x is the original data, x min is the minimum value of the data, x max is the maximum value of the data.

[0054] The feature extraction module plays a core role in the entire single-array low-altitude radar target recognition system based on deep learning. It is the key link to achieve accurate target recognition. This module mainly uses advanced deep learning models to automatically extract features that are valuable for target recognition from preprocessed data.

[0055] This deep learning model is built on the basis of convolutional neural network (CNN), which has shown excellent performance in image recognition, object detection and other fields. It contains multiple carefully designed convolutional layers, pooling layers and fully connected layers, each of which has its own unique functions and effects.

[0056] The convolution kernel size is 3×3 to 5×5, and the convolution layer can use different activation functions, including ReLU, LeakyReLU or ELU. The ReLU activation function formula is: f(x) = max(0, x), the LeakyReLU activation function formula is: f(x) = max(0.1x, x), and the ELU activation function formula is: Where α is an adjustable parameter, usually set to 1. The pooling layer uses maximum or average pooling operations with a step size of 2 to 3. The pooling window size is adjusted in the range of 2×2 to 4×4 according to data characteristics and network structure requirements. The number of neurons in the fully connected layer is determined by the target category and feature dimension. A Dropout layer can be added with a Dropout probability set to 0.2 to 0.5. The Dropout operation randomly sets the output of some neurons to 0 during training to prevent overfitting.

[0057] The target recognition module classifies and recognizes the target based on the extracted features. It uses support vector machine SVM, random forest RF or deep neural network DNN classification algorithm, and the algorithm parameters are optimized through cross-validation and grid search methods. For SVM, different kernel functions are used, such as the linear kernel function formula: K(x, y) = x T y, the radial basis kernel function formula is: K(x, y) = exp(-γ||xy|| 2 ), where γ is the kernel function parameter. The classification performance is optimized by adjusting the kernel function parameter and penalty coefficient. For random forest, the number of decision trees is set to 50 to 100, and the classification accuracy is improved by adjusting the depth and splitting criteria of the decision tree. For deep neural network DNN, the number of network layers, the number of neurons and the activation function parameters are adjusted. The recognition results include target type, position, speed and motion trajectory prediction. The target position is calculated using the formula: Where P(x, y, z) is the target position coordinate, ω i is the weight coefficient, R i (x, y, z) is the position information component provided by the i-th feature or sensor, and n is the number of features or sensors involved in the position calculation. The target speed is calculated using the formula: Where V is the target speed, f j is the influencing factor of speed-related characteristics, S j is the th velocity-related eigenvalue, and Δt is the time interval. The motion trajectory prediction uses the Kalman filter algorithm, which is based on historical data and the target motion model.

[0058] The model training and optimization module plays a vital role in the entire single-array low-altitude radar target recognition system based on deep learning. It is a key link in improving system performance and ensuring recognition accuracy.

[0059] The core task of this module is to train the deep learning model with a large amount of carefully labeled data. This amount of data is very large, covering a rich variety of scenes and target information, providing sufficient material for the model to learn target features. During the training process, the model parameters are in a continuous optimization state. By continuously adjusting the parameters, the model can gradually adapt to the radar echo signal characteristics of different types of targets in a variety of complex environments, thereby achieving a steady improvement in recognition accuracy.

[0060] It is worth noting that the labeling of training data is done by professionals. These professionals have deep professional knowledge and rich practical experience. They label the data according to the real attributes of the target. The labeling information contains multiple key dimensions, among which the target type is the primary labeling content, which clarifies whether the target is a small drone, jet, bird or other low-flying object. Position information is also indispensable. It records the coordinate position of the target in space in detail, including horizontal direction and vertical height information, which is of great significance for accurately tracking the trajectory of the target and judging the target's action path. Speed ​​information is also an important part of the labeling. It describes the speed of the target. Whether it is a slow-moving target at a low speed or an object flying at a high speed, its speed data provides a key basis for the model to understand the dynamic characteristics of the target.

[0061] This training method based on data labeled by professionals ensures the accuracy and reliability of the training data, laying a solid foundation for the training and optimization of deep learning models, enabling the model to better learn and identify the characteristics of low-altitude radar targets, thereby providing strong support for the efficient and accurate operation of the entire system. The back propagation algorithm is used to adjust the model weights, and the optimization methods include stochastic gradient descent SGD, Adagrad, and Adadelta. In SGD, the learning rate adjustment strategy is exponential decay, and the learning rate formula is: η t =η 0 β t , where η t is the current learning rate, η 0 is the initial learning rate, β is the decay coefficient, and t is the number of iterations. In Adagrad, the adaptive learning rate calculation formula is: where η t is the current learning rate, η 0 is the initial learning rate, g i is the gradient of the i-th iteration. In Adadelta, the update rule is: where Δx t is the parameter update amount, g i is the gradient of the i-th iteration, d iis the parameter update amount of the i-th iteration, and ρ is the new attenuation coefficient. At the same time, a regularization term is introduced to prevent overfitting. The regularization term uses L2 regularization, and the calculation formula is: Where F is the loss function after regularization, F 0 is the original loss function, k is the regularization coefficient, u i is the model parameter.

[0062] The visualization module displays the target position and trajectory of the target recognition results in two-dimensional or three-dimensional radar images, and uses OpenGL technology to achieve image rendering and interaction. The chart shows the target speed and type distribution, and provides interactive functions. Users can view detailed information such as the specific feature vector and historical trajectory of the target by clicking or touching the mouse.

[0063] In the present invention, the radar receiver antenna of the data acquisition module has a specific polarization mode, including horizontal polarization, vertical polarization or circular polarization, and can automatically switch or simultaneously receive echo signals of multiple polarization modes according to target characteristics and environmental conditions, thereby improving the target information acquisition capability. For example, for metal targets, the horizontal polarization mode may be more conducive to obtaining clear echo signals; for non-metal targets, the circular polarization mode may be more effective. In a complex environment, echo signals of multiple polarization modes are received simultaneously, and the accuracy of target detection and recognition is improved through a fusion algorithm.

[0064] In the present invention, during the denoising process, the data preprocessing module automatically selects a suitable filtering algorithm according to the noise type, Gaussian noise uses Wiener filtering, salt and pepper noise uses median filtering, and mixed noise uses adaptive weighted median filtering; during the clutter suppression process, the number of reference units and the number of protection units of the CFAR detection technology are dynamically adjusted according to the clutter environment, the number of reference units ranges from 8 to 32, and the number of protection units ranges from 2 to 8. The specific adjustment strategy is as described in claim 1.

[0065] In the present invention, in the deep learning model of the feature extraction module, the convolution layer can use different activation functions, including ReLU, LeakyReLU or ELU, the pooling window size of the pooling layer is adjusted in the range of 2×2 to 4×4 according to the data characteristics and network structure requirements, and the full connection layer can add a Dropout layer, and the Dropout probability is set to 0.2 to 0.5. The specific technical details are as described in claim 1.

[0066] In the present invention, a single-array low-altitude radar target recognition method based on deep learning includes the following steps:

[0067] The data collection step uses a single-array low-altitude radar receiver to collect echo signal data of different types of targets under various environmental conditions, and sets the sampling frequency and acquisition time according to the target movement speed and monitoring range. The sampling frequency for low-speed targets is 100Hz to 200Hz, and the sampling frequency for high-speed targets is 500Hz to 1000Hz. For targets with a distance of less than 10 kilometers, the sampling frequency can be appropriately reduced to 300Hz; for targets with a distance of more than 50 kilometers, the echo signal strength can be increased by increasing the radar transmission power, extending the integration time, or using multi-pulse accumulation technology. The transmission power is increased by 5dB on the original basis, the integration time is extended to 50ms, and the number of multi-pulse accumulations is 8. The acquisition time is adjusted within the range of 1 second to 10 seconds, and is dynamically adjusted according to the target movement speed and monitoring needs.

[0068] The data preprocessing step is to perform denoising, clutter suppression and normalization on the collected radar echo signal data. The specific processing method is as described in claim 1.

[0069] In the feature extraction step, the preprocessed data is input into the deep learning model to automatically extract the feature vector that represents the target characteristics. The dimension of the feature vector is determined by the model structure and the complexity of the target, and is between 128 and 1024 dimensions. The deep learning model selects the appropriate network structure according to the complexity of the target recognition task and the characteristics of the data. A 3- to 5-layer convolutional neural network structure is used for simple targets, and an 8- to 15-layer convolutional neural network architecture is used for complex targets or multi-class target recognition. At the same time, transfer learning technology is used to accelerate model training and improve recognition effects using the pre-trained model VGG16. The parameters of the pre-trained model are adjusted and new layers are added to adapt to specific low-altitude radar target recognition tasks.

[0070] In the target recognition step, based on the extracted feature vector, a classification algorithm is used to classify and identify the target, output the target type, position, and speed information, and predict the target motion trajectory based on historical data and the target motion model. The classification algorithm can use a weighted loss function during the training process to give different weights to the misclassification of different types of targets, with the focus on setting the target weight to 3. Post-processing techniques, including trajectory association and filter smoothing techniques, are combined to improve the stability and accuracy of target recognition. The nearest neighbor algorithm is used for trajectory association, and the Kalman filter algorithm is used for filter smoothing techniques.

[0071] In the model training and optimization steps, a large amount of labeled sample data is used to train the deep learning model. By adjusting the model parameters and optimizing the algorithm, the recognition accuracy and generalization ability of the model are improved. The training data can use data enhancement technology, including random rotation angles between -5° and 5°, flipping (horizontal or vertical), and scaling between 0.9 and 1.1. The data set size is expanded to improve the model's ability to recognize targets in different postures and environments. The optimization algorithm can dynamically adjust the learning rate according to the convergence of the model training process. The initial learning rate can be set to 0.005. When the training error no longer decreases for 10 consecutive epochs, the learning rate is reduced to 0.3 times the original.

[0072] The visualization step presents the target recognition results to the user in a visual manner, providing the user with intuitive and clear target information to assist in decision-making.

[0073] In the present invention, in the data collection step, for targets with a distance greater than 50 kilometers or a radar reflection cross-sectional area less than 1 square meter, the radar transmission power is increased, the integration time is extended, or a multi-pulse accumulation technology is used to improve the echo signal strength to ensure effective collection of target information. The transmission power is increased by 5dB on the original basis, the integration time is extended to 50ms, and the number of multi-pulse accumulations is 8.

[0074] In the present invention, in the data preprocessing step, the normalization process can select a suitable method according to the data distribution characteristics. If the data is approximately normally distributed, standard deviation normalization is used. If the data distribution range is narrow, linear normalization is used to make the data have better stability and consistency in subsequent processing.

[0075] In the present invention, in the feature extraction step, the deep learning model selects a suitable network structure according to the complexity of the target recognition task and the data characteristics. A 3-5-layer convolutional neural network structure is used for simple targets, and an 8-15-layer convolutional neural network architecture is used for complex targets or multi-category target recognition. At the same time, transfer learning technology is used to utilize pre-trained models to accelerate model training and improve recognition effects.

[0076] In the present invention, in the target recognition step, the classification algorithm can adopt a weighted loss function during the training process, give different weights to the misclassification of different types of targets, and set the weight of the key target to 3 to improve the recognition accuracy of the key target. At the same time, post-processing technology, including trajectory association and filtering smoothing technology, is combined to improve the stability and accuracy of target recognition.

[0077] In the present invention, in the model training and optimization steps, the training data can use data enhancement technology, including random rotation angles between -5° and 5°, flipping (horizontally or vertically), and scaling ratios between 0.9 and 1.1, to expand the size of the data set and improve the model's ability to recognize targets in different postures and environments. The optimization algorithm can dynamically adjust the learning rate according to the convergence of the model training process. The initial learning rate can be set to 0.005. When the training error no longer decreases for 10 consecutive epochs, the learning rate is reduced to 0.3 times the original.

[0078] The above are only preferred specific implementation modes of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can replace or change the technical solution and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A single-array low-altitude radar target recognition system based on deep learning, characterized in that: include: The data acquisition module is used to obtain the single-array low-altitude radar echo signal data, including the echo data of different types of targets in various environments. The acquisition equipment is a radar receiver with a sampling frequency range of 100MHz to 500MHz and a dynamic range of 60dB to 100dB. The data acquisition time interval is set according to the target motion characteristics and monitoring requirements. For drones flying at a speed below 5m / s, the acquisition time interval is extended to 2 seconds; for jet aircraft flying at a speed of 500m / s, the acquisition time interval is shortened to 0.1 seconds. The storage format is binary format, and the LZ77 compression algorithm is used for data compression; The data preprocessing module preprocesses the collected radar echo signal data, including denoising, clutter suppression and data normalization. It uses an adaptive filtering algorithm to denoise and automatically adjusts the filter parameters according to the power spectral density characteristics of the noise. For Gaussian noise, Wiener filtering is used to calculate the filter coefficient by estimating the power spectral density of the signal and noise. The Wiener filtering formula is: Where H(f) is the filter frequency response, S s (f) is the signal power spectrum density, S n (f) is the noise power spectrum density. For salt and pepper noise, median filtering is used to sort the pixel values ​​in the window and take the median as the output. For mixed noise, adaptive weighted median filtering is used. The median filter window size and weight are dynamically adjusted according to the noise type and intensity. Clutter suppression is performed based on the constant false alarm rate CFAR detection technology. The number of reference units and the number of protection units are dynamically adjusted according to the clutter environment. In the marine environment, the number of reference units is set to 16 to 24, and the number of protection units is set to 4 to 6; in the urban environment, the number of reference units is adjusted to 12 to 20, and the number of protection units is adjusted to 3 to 5. The data is normalized to the [0,1] interval by linear or standard deviation normalization method. If the data distribution is approximately normal, standard deviation normalization is used. The calculation formula is: Where x is the original data, μ is the data mean, and σ is the data standard deviation; if the data distribution range is narrow, linear normalization is used, and the calculation formula is: Where x is the original data, x min is the minimum value of the data, x max is the maximum value of the data; The feature extraction module uses a deep learning model to automatically extract features from the preprocessed data. The deep learning model is built based on a convolutional neural network (CNN), which contains multiple convolutional layers, pooling layers, and fully connected layers. The convolution kernel size is 3×3 to 5×5. The convolutional layer uses different activation functions, including ReLU, LeakyReLU, or ELU. The ReLU activation function formula is: K(x)=max(0,x), the LeakyReLU activation function formula is: f(x)=max(0.1x,x), and the ELU activation function formula is: Among them, α is an adjustable parameter, usually 1. The pooling layer uses maximum or average pooling operation with a step size of 2 to 3. The pooling window size is adjusted in the range of 2×2 to 4×4 according to the data characteristics and network structure requirements. The number of neurons in the fully connected layer is determined according to the target category and feature dimension. A Dropout layer is added, and the Dropout probability is set to 0.2 to 0.

5. The Dropout operation randomly sets the output of some neurons to 0 during the training process to prevent overfitting. The target recognition module classifies and recognizes the target based on the extracted features, using support vector machine SVM, random forest RF or deep neural network DNN classification algorithm. The algorithm parameters are optimized through cross-validation and grid search methods. For SVM, different kernel functions are used, including the linear kernel function formula: K(x, y) = x T y, the radial basis kernel function formula is: K(x, y) = exp(-γ||xy|| 2 ), where γ is the kernel function parameter. The classification performance is optimized by adjusting the kernel function parameter and the penalty coefficient. For random forest, the number of decision trees is set to 50 to 100, and the classification accuracy is improved by adjusting the depth and splitting criteria of the decision tree. For deep neural network DNN, the number of network layers, the number of neurons and the activation function parameters are adjusted. The recognition results include target type, position, speed and motion trajectory prediction. The target position is calculated using the formula: Where P(x, y, z) is the target position coordinate, ω i is the weight coefficient, R i (x, y, z) is the position information component provided by the i-th feature or sensor, n is the number of features or sensors involved in the position calculation, and the target speed is calculated using the formula: Where V is the target speed, f j is the influencing factor of speed-related characteristics, S j is the th velocity-related characteristic value, Δt is the time interval, and the motion trajectory prediction adopts the Kalman filter algorithm to make predictions based on historical data and the target motion model; The model training and optimization module uses a large amount of labeled data to train the deep learning model, continuously optimizes the model parameters to improve the recognition accuracy. The training data is labeled by professionals according to the real attributes of the target. The labeled information includes the target type, position, and speed. The back propagation algorithm is used to adjust the model weight. The optimization methods include stochastic gradient descent SGD, Adagrad, and Adadelta. In SGD, the learning rate adjustment strategy is exponential decay. The learning rate formula is: η t =η0·β t , where η t is the current learning rate, η0 is the initial learning rate, β is the decay coefficient, and t is the number of iterations. In Adagrad, the adaptive learning rate calculation formula is: where η t is the current learning rate, η0 is the initial learning rate, g i is the gradient of the i-th iteration. In Adadelta, the update rule is: where Δx t is the parameter update amount, g i is the gradient of the i-th iteration, d i is the parameter update amount of the i-th iteration, ρ is the new attenuation coefficient, and the regularization term is introduced to prevent overfitting. The regularization term uses L2 regularization, and the calculation formula is: Where F is the regularized loss function, F0 is the original loss function, κ is the regularization coefficient, and u i are model parameters; The visualization display module displays the target position and trajectory of the target recognition results in a two-dimensional or three-dimensional radar image, uses OpenGL technology to achieve image rendering and interaction, and displays the target speed and type distribution in charts. It provides interactive functions, and users can view detailed information, including the specific feature vector and historical trajectory of the target, by clicking or touching the mouse.

2. According to the single-sided array low-altitude radar target recognition system based on deep learning in claim 1, it is characterized in that: The radar receiver antenna of the data acquisition module has a specific polarization mode, including horizontal polarization, vertical polarization or circular polarization. It automatically switches or simultaneously receives echo signals of multiple polarization modes according to target characteristics and environmental conditions, thereby improving the ability to obtain target information. For metal targets, horizontal polarization is used to obtain echo signals; for non-metallic targets, circular polarization is used to obtain echo signals. In complex environments, echo signals of multiple polarization modes can be received simultaneously.

3. According to the single-sided array low-altitude radar target recognition system based on deep learning in claim 1, it is characterized in that: During the denoising process, the data preprocessing module automatically selects the appropriate filtering algorithm according to the noise type. Gaussian noise uses Wiener filtering, salt and pepper noise uses median filtering, and mixed noise uses adaptive weighted median filtering. During the clutter suppression process, the number of reference units and protection units of the CFAR detection technology are dynamically adjusted according to the clutter environment. The number of reference units ranges from 8 to 32, and the number of protection units ranges from 2 to 8.

4. The single-sided array low-altitude radar target recognition system based on deep learning according to claim 1 is characterized in that: In the deep learning model of the feature extraction module, the convolution layer can use different activation functions, including ReLU, LeakyReLU or ELU. The pooling window size of the pooling layer can be adjusted in the range of 2×2 to 4×4 according to the data characteristics and network structure requirements. The Dropout layer can be added to the fully connected layer, and the Dropout probability is set to 0.2 to 0.

5.

5. A single-array low-altitude radar target recognition method based on deep learning, characterized in that: The following steps are involved: Data collection steps: Use a single-array low-altitude radar receiver to collect echo signal data of different types of targets under various environmental conditions. Set the sampling frequency and acquisition time according to the target movement speed and monitoring range. For targets with a distance of less than 10 kilometers, the sampling frequency is reduced to 300Hz; for targets with a distance of more than 50 kilometers, increase the radar transmission power, extend the integration time, or use multi-pulse accumulation technology to improve the echo signal strength. The transmission power is increased by 5dB on the original basis, the integration time is extended to 50ms, the number of multi-pulse accumulation is 8, and the acquisition time is adjusted within the range of 1 second to 10 seconds; Data preprocessing step, denoising, clutter suppression and normalization of the collected radar echo signal data; In the feature extraction step, the preprocessed data is input into the deep learning model to automatically extract the feature vector that represents the target characteristics. The dimension of the feature vector is determined by the model structure and the complexity of the target, and is between 128 and 1024 dimensions. The deep learning model selects a suitable network structure according to the complexity of the target recognition task and the characteristics of the data. It uses transfer learning technology and the pre-trained model VGG16 to accelerate model training and improve recognition effects. It adapts to the low-altitude radar target recognition task by adjusting the parameters of the pre-trained model and adding new layers. In the target recognition step, based on the extracted feature vector, a classification algorithm is used to classify and identify the target, output the target type, position, and speed information, and predict the target motion trajectory based on historical data and the target motion model. The classification algorithm uses a weighted loss function during the training process, and gives different weights to the misclassification of different types of targets. The focus is on setting the target weight to 3. Post-processing techniques, including trajectory association and filtering and smoothing techniques, are combined to improve the stability and accuracy of target recognition. The nearest neighbor algorithm is used for trajectory association, and the Kalman filtering algorithm is used for filtering and smoothing techniques. Model training and optimization steps: Use labeled sample data to train the deep learning model. By adjusting model parameters and optimizing algorithms, the recognition accuracy and generalization ability of the model can be improved. Data enhancement technology can be used for training data, including random rotation angles between -5° and 5°, flipping, and scaling between 0.9 and 1.

1. Expand the data set size to improve the model's ability to recognize targets in different postures and environments. The optimization algorithm dynamically adjusts the learning rate according to the convergence of the model training process. The initial learning rate can be set to 0.

005. When the training error no longer decreases for 10 consecutive epochs, the learning rate is reduced to 0.3 times the original value. The visualization step presents the target recognition results to the user in a visual manner, providing the user with intuitive and clear target information to assist in decision-making.

6. The single-sided array low-altitude radar target recognition method based on deep learning according to claim 5 is characterized in that: In the data collection step, for targets with a distance greater than 50 kilometers or a radar reflection cross-sectional area less than 1 square meter, the radar transmission power is increased, the integration time is extended, or the multi-pulse accumulation technology is used to improve the echo signal strength to ensure the effective collection of target information. The transmission power is increased by 5dB on the original basis, the integration time is extended to 50ms, and the number of multi-pulse accumulations is 8.

7. The single-sided array low-altitude radar target recognition method based on deep learning as claimed in claim 5 is characterized in that: In the data preprocessing step, the normalization method can be selected according to the characteristics of the data distribution. If the data is approximately normally distributed, standard deviation normalization is used. If the data distribution range is narrow, linear normalization is used to make the data stable and consistent in subsequent processing.

Images