Radar adaptive detection method and system fusing model driving and data driving

Abstract

The application discloses a radar adaptive detection method and system fusing model driving and data driving, and belongs to the technical field of radar signal processing. In view of the defects that the existing model driving method has small sample performance loss and is prone to model mismatch, and the data driving method has calculation complexity and inefficient feature extraction, the application first acquires radar to-be-detected unit data and reference unit data, and constructs a 3D feature vector based on a generalized likelihood ratio test, an adaptive matched filter and a test statistic of an adaptive coherence estimator; then a shallow feedforward neural network is constructed, and a gradient descent type algorithm is used to train the network to replace a traditional detector decision structure; finally, the feature vector is input into the trained network, a threshold is determined by adjusting an output layer bias, and a detection result is obtained by comparing network output scalars. The application takes into account the physical meaning support of the model and the nonlinear fitting advantages of the data, effectively adapts to complex scenes such as small samples and non-Gaussian clutter, and significantly improves the comprehensive performance of radar adaptive detection.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing technology, specifically relating to a radar adaptive detection method and system that integrates model-driven and data-driven approaches. Background Technology

[0002] Target signal detection is a core function and research hotspot in radar signal processing, directly determining the radar system's ability to identify and track targets. It has crucial application value in fields such as national defense, aerospace, and more. Essentially, it is a binary hypothesis testing problem in hypothesis testing statistics. Its core objective is to accurately determine the presence of a target signal in complex scenarios where key parameters such as the background noise characteristics and target amplitude of the target unit are unknown. Traditional optimal detection methods based on the Neyman-Pearson criterion are difficult to apply directly due to their reliance on known parameters. Therefore, adaptive detection techniques have emerged as the main approach to solving this problem.

[0003] Existing radar adaptive detection technologies mainly fall into two core categories: model-driven and data-driven. However, both have significant drawbacks and cannot meet the high-performance detection requirements of complex real-world scenarios. The model-driven approach is the traditional mainstream technology. Its core idea is to establish a mathematical model of the observed data based on hypothesis testing statistical theory, and then construct a detector. Classic schemes include the Generalized Likelihood Ratio Test (GLRT) and Adaptive Matched Filter (AMF) for Gaussian noise scenarios. Both estimate the noise covariance matrix through reference cells, exhibiting asymptotic optimality, but they require a large amount of reference data. Under small sample conditions, the covariance matrix estimation accuracy is insufficient, resulting in a severe loss of detection performance. While the Adaptive Coherence Estimator (ACE) for non-Gaussian clutter scenarios can improve adaptability to non-uniform environments, its performance under Gaussian noise is inferior to the former two. The fundamental limitation of this type of method is that detection performance is highly dependent on the degree of matching between the preset mathematical model and the actual data. When the statistical characteristics of the detection environment are non-stationary, model mismatch is inevitable, leading to a significant decrease in detection performance.

[0004] Unlike model-driven methods, data-driven methods are data-centric, building models by learning implicit patterns in the data without requiring pre-set complex physical / mathematical models. With the development of artificial intelligence, they are increasingly being applied to radar detection. Currently, their applications mainly fall into three categories: First, preprocessing radar echoes into two-dimensional matrices such as range pulse maps and range-Doppler maps, then training recognition through deep neural networks. This method requires complex preprocessing processes, and the computational load of deep networks is enormous, resulting in high engineering application costs. Second, end-to-end processing is used, directly taking the one-dimensional sequence of radar echoes as input and relying on neural networks to extract signal features. However, in low signal-to-noise ratio scenarios, useful signals are hidden in noise, making feature extraction extremely difficult and resulting in insufficient detection reliability. Third, replacing local modules of existing detectors with neural networks, such as replacing the threshold calculation module or optimizing the covariance matrix estimation. These solutions only make local improvements and do not fundamentally solve the problem of adapting feature extraction and decision structures, failing to fully leverage the nonlinear fitting advantages of data-driven methods. Overall, the core challenge of data-driven methods lies in how to efficiently extract effective signal features, and existing solutions have not effectively overcome this bottleneck.

[0005] In summary, model-driven methods are limited by model mismatch and performance loss due to small sample sizes, while data-driven methods suffer from computational complexity, low feature extraction efficiency, or incomplete improvements. Neither type of method can simultaneously adapt to complex real-world scenarios such as small sample sizes and non-Gaussian clutter. Therefore, there is an urgent need for a technical solution that integrates the advantages of both, balancing the physical support of model-driven methods with the nonlinear fitting capabilities of data-driven methods, in order to overcome the shortcomings of existing technologies and improve the overall performance of radar adaptive detection. Summary of the Invention

[0006] In view of this, the purpose of this invention is to provide a radar adaptive detection method and system that integrates model-driven and data-driven approaches. By converting the test statistics of the generalized likelihood ratio test detector, the adaptive matched filter detector, and the adaptive coherence estimator detector into 3D feature vectors, and using a shallow feedforward neural network to replace the decision structure of the traditional detector, this method balances physical meaning support and nonlinear fitting capability. It overcomes the shortcomings of existing technologies such as model mismatch, performance loss in small samples, computational complexity, and low feature extraction efficiency, thereby improving the overall adaptive detection performance of radar in complex scenarios such as small samples and non-Gaussian clutter.

[0007] To achieve the above-mentioned objectives, the technical solution adopted is as follows:

[0008] This invention discloses a radar adaptive detection method integrating model-driven and data-driven approaches, comprising the following steps: S1, acquiring radar target cell data and reference cell data; the target cell data contains noise and may contain target signals, used for target signal presence determination; the reference cell data does not contain target signals and its noise statistical characteristics are consistent with those of the target cell, used to estimate the noise covariance matrix of the target cell; S2, based on the feature quantities relied upon by the model-driven adaptive detector, combining the target cell data, reference cell data, and noise covariance matrix from step S1, constructing a 3D feature vector in the feature space; the model-driven adaptive detector includes a generalized likelihood ratio test detector, an adaptive matched filter detector, and an adaptive coherence estimator detector; S3, constructing... A shallow feedforward neural network consisting of an input layer, hidden layers, and an output layer is used. The number of units in the input layer matches the dimension of the 3D feature vector, and the output layer has a single output unit. Based on the original data of the preset training dataset, training feature vectors and labeled training data are generated according to the feature vector construction method in step S2. A gradient descent algorithm is used to train the shallow feedforward neural network by minimizing the loss function, updating the network parameters, so that the shallow feedforward neural network replaces the decision structure of the model-driven adaptive detector. In step S4, the feature vector constructed in step S2 is input into the shallow feedforward neural network trained in step S3 to obtain the output scalar. The output scalar is compared with a preset threshold to obtain the radar target detection signal decision result. The preset threshold meets the radar's preset target false alarm probability requirement.

[0009] This invention transforms the core test statistics of the generalized likelihood ratio test detector, adaptive matched filter detector, and adaptive coherence estimator detector into 3D feature vectors. This retains the physical meaning of the model-driven method and comprehensively covers effective detection information in different scenarios, fundamentally alleviating the model mismatch problem of traditional model-driven methods. At the same time, it replaces the decision structure of traditional detectors with a shallow feedforward neural network, making full use of the nonlinear fitting characteristics of neural networks. This avoids the complex preprocessing and huge computational load of deep networks in data-driven methods, and solves the pain points of difficult feature extraction under low signal-to-noise ratio and incomplete improvement of local replacement schemes in end-to-end methods.

[0010] Under conditions of small sample reference data, this method can accurately uncover the feature differences between the target and clutter, and its detection performance is significantly better than that of conventional adaptive detectors. In complex real-world scenarios such as non-Gaussian clutter and non-uniformity, its detection performance is also effectively improved, and it possesses superior clutter suppression capabilities and cross-dataset and cross-polarity generalization abilities. Furthermore, by optimizing training and threshold adjustment mechanisms through gradient descent algorithms, this method can flexibly adapt to the radar's preset false alarm probability requirements, balancing detection accuracy and engineering practicality. It comprehensively overcomes the shortcomings of existing technologies and significantly improves the overall performance of radar adaptive detection.

[0011] Furthermore, the three features of the 3D feature vector correspond to the test statistics of the generalized likelihood ratio test detector, the adaptive matched filter detector, and the adaptive coherence estimator detector, respectively.

[0012] In this invention, the feature quantities of the 3D feature vector directly correspond to the test statistics of the generalized likelihood ratio test detector, the adaptive matched filter detector, and the adaptive coherence estimator detector. This ensures that the feature space can fully cover the core effective information of the model-driven method, providing the neural network with inputs that have clear physical meaning and improving the reliability and pertinence of detection decisions.

[0013] Furthermore, in step S2, the 3D feature vector is constructed using the following formula:

[0014] ;

[0015] In the formula, It is a 3D feature vector. The target guiding vector, For the data of the unit to be detected, The noise covariance matrix estimated from the reference cell is shown in the superscript. This is the conjugate transpose operation. This is for the transpose operation.

[0016] In this invention, a 3D feature vector is constructed using explicit mathematical formulas, which can characterize the core test statistics of three classic detectors, ensuring the accuracy, consistency, and repeatability of feature extraction. This provides a stable and effective input basis for the training of shallow feedforward neural networks, further guaranteeing the stability of detection performance.

[0017] Furthermore, in step S3, the hidden layer of the shallow feedforward neural network uses the hyperbolic tangent function as the activation function, which maps the real number axis to the interval (-1, 1).

[0018] In this invention, the hidden layer of the shallow feedforward neural network uses the hyperbolic tangent function as the activation function, which can map the input features to the (-1,1) interval. This can not only fully explore the nonlinear relationship in the feature vector, but also ensure the stability of the hidden layer output, thus helping the neural network to converge quickly and improving the feature classification accuracy and detection efficiency.

[0019] Further, in step S3, the labels of the training data are binary labels, including label 1 and label -1; wherein, label 1 indicates that the target unit has a target signal, and label -1 indicates that the target unit does not have a target signal.

[0020] In this invention, training data is defined using binary labels where 1 represents the presence of a target signal and -1 represents the absence of a target signal. This clearly distinguishes between the two types of detection scenarios, ensuring the target-oriented nature of neural network training, improving the model's ability to recognize the existence of target signals, and guaranteeing the accuracy of detection decisions.

[0021] Furthermore, in step S3, the gradient descent algorithm is the Levenberg-Marquardt backpropagation algorithm.

[0022] In this invention, the Levenberg-Marquardt backpropagation algorithm is selected as the gradient descent training algorithm. It has fast convergence speed and high training accuracy, can efficiently optimize the parameters of shallow feedforward neural networks, shorten the model training time, and at the same time improve the network's learning effect on the mapping relationship between features and detection decisions, thus ensuring the high performance of the detector.

[0023] Furthermore, in step S3, the gradient descent algorithm updates the parameters of the shallow feedforward neural network using the following formula:

[0024] ;

[0025] In the formula, For the first The updated network parameter set after the next iteration For the first The set of network parameters at the next iteration To update the step size, For gradient operators, The gradient operator is applied to the loss function. The result.

[0026] In this invention, a clear parameter update formula is used to achieve precise iterative optimization of the parameters of the shallow feedforward neural network. This effectively minimizes the loss function value, quickly corrects network parameter deviations, ensures that the neural network stably learns the correspondence between features and detection results, and improves the accuracy of detection decisions.

[0027] Further, in step S3, the loss function is defined using mean squared error, which is used to quantify the deviation between the output of the shallow feedforward neural network and the training data labels. The expression is:

[0028] ;

[0029] In the formula, Let be the mean squared error loss function of the shallow feedforward neural network. For training data batch size, For the index of the training samples, For the first The output of the shallow feedforward neural network corresponding to each training sample. For the first The binary labels of each training sample. This is the weight matrix of the hidden layer. The bias vector of the hidden layer. For the weights of the output layer, The bias coefficients of the output layer. For the activation function of the hidden layer of a shallow feedforward neural network, For the first The 3D feature vector corresponding to each training sample.

[0030] In this invention, the loss function is defined by mean squared error, which can accurately quantify the deviation between the neural network output and the training data labels, providing a clear and explicit target for network parameter optimization, ensuring the effectiveness of neural network training, and ensuring that the model can stably output detection results that meet the requirements.

[0031] Furthermore, in step S4, the preset threshold is determined by adjusting the bias coefficient of the output layer of the shallow feedforward neural network, with the goal of ensuring that the false alarm probability of the radar meets the preset requirements.

[0032] In this invention, a preset threshold is determined by adjusting the bias coefficient of the output layer of the shallow feedforward neural network, which can flexibly adapt to the radar's preset false alarm probability requirements. While ensuring detection performance, it meets the requirements of engineering applications for false alarm control, thereby improving the practicality and engineering adaptability of the method.

[0033] This invention also discloses a system for implementing the radar adaptive detection method that integrates model-driven and data-driven approaches, comprising: a data acquisition module for acquiring radar target unit data and reference unit data; the target unit data contains noise and may contain target signals, and is used for target signal presence determination; the reference unit data does not contain target signals and its noise statistical characteristics are consistent with those of the target unit, and is used to estimate the noise covariance matrix of the target unit; a feature construction module for constructing a 3D feature vector in the feature space based on the feature quantities relied upon by the model-driven adaptive detector, combined with the target unit data, reference unit data, and noise covariance matrix output by the data acquisition module; the model-driven adaptive detector includes a generalized likelihood ratio test detector, an adaptive matched filter detector, and an adaptive coherence estimator detector; and a network training module for constructing a network consisting of an input layer, a hidden layer, and an output layer. The shallow feedforward neural network consists of layers, with the number of input layer units matching the dimension of the 3D feature vector, and the output layer having a single output unit. It is also used to generate training 3D feature vectors and training data with binary labels based on the original data of the preset training dataset, according to the 3D feature vector construction method of the feature construction module. A gradient descent-like algorithm is also used to train the shallow feedforward neural network by minimizing the loss function, updating the network parameters, and allowing the shallow feedforward neural network to replace the decision structure of the model-driven adaptive detector. The detection decision module is used to input the 3D feature vector constructed by the feature construction module into the shallow feedforward neural network trained by the network training module to obtain the output scalar. It is also used to compare the output scalar with a preset threshold to obtain the radar target detection decision result. The preset threshold is determined by adjusting the bias coefficients of the output layer of the shallow feedforward neural network, with the goal of ensuring that the false alarm probability of the radar meets preset requirements.

[0034] This system achieves a precise radar adaptive detection method that integrates model-driven and data-driven approaches through the division of labor and collaboration among modules for data acquisition, feature construction, network training, and detection decision. The modules have clear functions and coherent logic, ensuring the stable implementation of the technical solution. This not only improves the overall detection performance in complex scenarios but also has good engineering feasibility.

[0035] The advantages of this invention compared to the prior art are as follows:

[0036] This invention integrates the core advantages of model-driven and data-driven approaches. It constructs a 3D feature vector using the features of a generalized likelihood ratio test detector, an adaptive matched filter detector, and an adaptive coherence estimator detector. This retains the physical meaning of the model-driven method, comprehensively covers effective detection information in different scenarios, and fundamentally alleviates the model mismatch problem of traditional model-driven methods. Simultaneously, it replaces the decision structure of traditional detectors with a shallow feedforward neural network, fully leveraging the nonlinear fitting characteristics of neural networks. This avoids the complex preprocessing and huge computational load of deep networks in data-driven methods, and solves the pain points of end-to-end methods such as difficulty in feature extraction under low signal-to-noise ratio and incomplete improvement of local replacement schemes. In complex scenarios such as small sample reference data and non-Gaussian clutter, its detection performance is significantly better than conventional adaptive detectors. Furthermore, by adapting to preset false alarm probability requirements, it balances detection accuracy and engineering practicality, comprehensively overcoming the shortcomings of existing technologies and significantly improving the overall performance of radar adaptive detection.

[0037] The following describes in detail the radar adaptive detection method and system that integrates model-driven and data-driven approaches, with reference to the embodiments shown in the accompanying drawings. Attached Figure Description

[0038] Figure 1 This is a flowchart of the steps of the radar adaptive detection method that integrates model-driven and data-driven approaches of the present invention.

[0039] Figure 2 This is a schematic diagram of the fusion structure of the ANN-assisted detector and the traditional model-driven adaptive detector of the present invention;

[0040] Figure 3 This is the curve showing the effect of the number of reference units K=12 on the detection performance in this invention;

[0041] Figure 4 This is the effect curve of the number of reference units K=32 on the detection performance in this invention;

[0042] Figure 5 This is a Doppler frequency sensitivity analysis diagram of the target of this invention;

[0043] Figure 6 This is a clutter suppression performance analysis diagram of the present invention;

[0044] Figure 7 This is the detection performance curve under the condition of guide vector mismatch in this invention;

[0045] Figure 8 This is an amplitude distribution diagram of the measured sea clutter data (data #1) of this invention;

[0046] Figure 9 This is an amplitude distribution diagram of the measured sea clutter data (data #2) of this invention;

[0047] Figure 10 This is a graph showing the generalization ability of the detector of this invention on different datasets;

[0048] Figure 11 This is a graph showing the generalization ability of the detector of the present invention under different polarization conditions. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with embodiments of this invention. Obviously, the described embodiments are one embodiment of this invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0050] like Figure 1 As shown, the core process of the radar adaptive detection method that integrates model-driven and data-driven approaches disclosed in this invention includes the following steps: S1, acquiring radar target unit data and reference unit data; the target unit data contains noise and may contain target signals, which is used to determine the existence of target signals; the reference unit data does not contain target signals and its noise statistical characteristics are consistent with those of the target unit, which is used to estimate the noise covariance matrix of the target unit; S2, based on the feature quantities on which the model-driven adaptive detector depends, and combining the target unit data, reference unit data, and noise covariance matrix from step S1, constructing a 3D feature vector in the feature space; the model-driven adaptive detector includes a generalized likelihood ratio test detector, an adaptive matched filter detector, and an adaptive coherence estimator detector; S3, A shallow feedforward neural network consisting of an input layer, hidden layers, and an output layer is constructed. The number of units in the input layer matches the dimension of the 3D feature vector, and the output layer has a single output unit. Based on the original data of the preset training dataset, training feature vectors and labeled training data are generated according to the feature vector construction method in step S2. A gradient descent algorithm is used to train the shallow feedforward neural network by minimizing the loss function, updating the network parameters, so that the shallow feedforward neural network replaces the decision structure of the model-driven adaptive detector. In step S4, the feature vector constructed in step S2 is input into the shallow feedforward neural network trained in step S3 to obtain the output scalar. The output scalar is compared with a preset threshold to obtain the radar target detection signal decision result. The preset threshold meets the radar's preset target false alarm probability requirement.

[0051] To clearly present the specific implementation details, technical principles, and performance verification process of each step of this invention, the following will elaborate in detail according to the logic of "basic model preparation, core solution implementation, and performance verification analysis": First, the hypothesis testing model of radar adaptive detection, the principle of classical detectors, and the core technical pain points are clarified to provide a foundation for subsequent innovative solutions; then, the key design of adaptive detection assisted by artificial neural networks (ANN) is described in detail, including the construction of 3D feature vectors, the structure and training process of shallow feedforward neural networks; finally, through computer simulation data and IPIX radar measured data, the detection performance of this invention is verified in multiple scenarios (small sample, non-Gaussian clutter, guide vector mismatch, etc.), and its advantages over conventional adaptive detectors are intuitively compared.

[0052] This embodiment considers the adaptive detection problem of point targets under complex Gaussian noise. The hypothesis testing problem can be expressed as:

[0053]

[0054] In this hypothesis testing problem, for the null hypothesis Data of the unit to be detected Includes only noise data vectors The noise data vector has a length of A circularly symmetric, zero-mean, complex Gaussian random vector has the following covariance matrix: And the vector length satisfies ( This represents the number of pulses within the radar coherent processing interval. (Number of antenna elements). For the alternative hypothesis... Data of the unit to be detected In addition to noise components, it also contains a useful signal (i.e., the target signal), whose steering vector is... The amplitude is Among them, the target complex amplitude Closely related to target reflection characteristics and electromagnetic wave propagation conditions, these are usually unknown parameters; simultaneously, the noise covariance matrix... Also assumed to be an unknown parameter. Due to the target complex amplitude... With noise covariance matrix Since the values ​​are all unknown, conventional detectors based on the Neyman-Pearson criterion are no longer applicable, meaning there is no consistent maximum potential test. To address this issue, an effective solution is to use the generalized likelihood ratio test (GLRT), which possesses asymptotically optimal properties. This is to achieve a consistent maximum potential test for the noise covariance matrix. For an accurate estimate, this scheme assumes the existence of Reference unit data The reference cells are selected from the region near the cell to be detected, and the data of each reference cell is... Includes only noise vectors Since there is no target signal component, the noise covariance matrix of the unit to be detected can be completed using the reference cell data. The estimate.

[0055] In radar detection, reference cells are typically selected from the vicinity of the cell to be detected to ensure that the noise statistical characteristics of the reference cells and the cell to be detected are consistent. To ensure that the noise covariance matrix estimated based on reference cell data has positive definiteness, a certain number of reference cells is usually required. ( (length of the noise data vector); only when the number of reference cells... When the value approaches infinity, the performance of the adaptive detector can approach that of the optimal detector.

[0056] The above analysis shows that adaptive detectors belong to a typical model-driven design approach, and their detection performance largely depends on the degree of matching between the preset mathematical model and the actual detection conditions. Specifically, on the one hand, these detectors typically assume that the noise statistics follow a complex Gaussian distribution. If the noise statistics in the actual environment deviate from this assumption, the detector's performance will significantly decrease. On the other hand, the detection performance is highly dependent on the number of reference units. Highly sensitive, when When the value is small, the performance loss of the detector will increase significantly.

[0057] In summary, the effective operation of the radar adaptive detection model relies on several key assumptions, including: the noise data of both the reference cells and the cells to be detected are random vectors with complex Gaussian distributions that follow the same statistical properties; and the number of reference cells is not less than the length of the noise data vector. The assumptions include the known target steering vector. Once these assumptions are violated, technical challenges such as non-uniform non-Gaussian clutter scenarios, small sample conditions, and steering vector mismatch will arise. These challenges are currently the research hotspots in the field of adaptive radar signal detection.

[0058] For adaptive detection problems with unknown covariance matrix and complex Gaussian noise, a landmark detector, Kelly's GLRT, can be obtained by using the so-called one-step method. Its detector structure is as follows:

[0059]

[0060] Where the sample covariance matrix is:

[0061]

[0062] in, The target guiding vector, For the data of the unit to be detected, The noise covariance matrix, estimated using reference cell data, is used to replace the true, unknown noise covariance matrix, which is the sample covariance matrix. , The number of reference cells, representing the number of reference data points used to estimate the covariance, must satisfy the following condition: To ensure Positive definiteness, For detector threshold, For the first The reference cell data contains only noise components (no target signal) and is of length [length missing]. A circularly symmetric complex Gaussian random vector To Performing the conjugate transpose is applicable to matrix operations on complex vectors. Detector threshold. It is related to the false alarm probability specified by radar detection.

[0063] Based on this detector, a two-step generalized likelihood ratio method can also be used. First, assuming the noise covariance matrix is ​​known, a generalized likelihood ratio detector is constructed. Then, the estimated noise covariance matrix obtained from the reference data is substituted into the generalized likelihood ratio. The detector obtained in this way is called an adaptive matched filter (AMF).

[0064]

[0065] Among them, detector threshold The false alarm probability is related to the radar detection specified. It can be seen that the AMF detector has a simpler structure and requires less computation. Since the adaptive detection problem discussed in this invention does not have a uniformly optimal solution (no uniform maximum potential test), the superiority of AMF over Kelly's GLRT is not absolute, but highly dependent on the actual conditions of the specific detection environment. It is worth noting that both exhibit strong robustness in steering vector mismatch scenarios and have higher tolerance to steering vector deviations.

[0066] Furthermore, if we consider that the noise follows a non-Gaussian distribution, or that the covariance matrices of the reference cell and the cell to be detected have the same structure but differ by an unknown scaling factor, we can also obtain an adaptive coherent estimator (ACE) using the generalized likelihood ratio method:

[0067]

[0068] Among them, detector threshold It is related to the false alarm probability specified by radar detection. The ACE detector can be regarded as the data to be detected. With guide vector The generalized cosine of the included angle is used, hence it is also called an adaptive cosine estimator. This detector has good adaptability to complex detection scenarios that are non-Gaussian and non-uniform, and is a commonly used adaptive detection method in such scenarios. At the same time, it has strong selectivity under the condition of guide vector mismatch, which can effectively reduce the adverse effects of guide vector deviation on detection performance.

[0069] The three detectors mentioned above are some common and classic detectors in adaptive detection problems, and are usually used as standards to measure the performance of other detectors.

[0070] This invention uses a shallow feedforward neural network to replace the decision structure of the traditional adaptive detector. This shallow feedforward neural network has the advantages of low computational cost and short training time, and can give full play to the nonlinear fitting characteristics of the neural network, effectively improving the overall performance of the detector.

[0071] Assuming the feature vector dimension is m, the specific structure of this shallow feedforward neural network is as follows: the input layer contains m units, matching the feature vector dimension; the hidden layer contains n units, used to extract deep nonlinear features of the feature vector; the output layer has 1 unit, used to output the decision criteria for object detection. The detector structure based on this feedforward neural network can be represented as:

[0072]

[0073] Where c is a 3D feature vector, U is the weight matrix from the input layer to the hidden layer, a is the hidden layer bias vector, f is the hidden layer activation function, w is the weight row vector from the hidden layer to the output layer, and b is the output layer bias coefficient.

[0074] The output of this shallow feedforward neural network is a scalar, which needs to be compared with a preset threshold of 0.5. If the scalar is ≥ 0.5, it is determined that a target signal exists; otherwise, it is determined that a target signal does not exist, thus completing the target detection decision.

[0075] The input to this neural network is of length [length missing]. eigenvectors The connection weights between the input layer and the hidden layer are determined by... dimensional weight matrix This indicates that the bias of the hidden layer is determined by... 3D bias vector This indicates that the connection weights between the hidden layer and the output layer are determined by... Dimensional weighted row vector This indicates that the bias coefficients of the output layer are determined by a scalar. The activation function of the hidden layer is defined as the hyperbolic tangent function, with the specific expression as follows:

[0076]

[0077] The hyperbolic tangent activation function can map the real number axis to the interval (-1, 1).

[0078] When applying artificial neural networks to radar signal detection tasks, the first step is to determine the input data and labels required for network training. For the radar adaptive detection problem focused on in this invention, a dimension can be defined by combining the three adaptive detector structures proposed in this invention. eigenvectors This is used as the input to the neural network:

[0079]

[0080] In the formula, It is a 3D feature vector. The target guiding vector, For the data of the unit to be detected, The noise covariance matrix estimated from the reference cell is shown in the superscript. This is the conjugate transpose operation. This is a transpose operation. This design can accurately quantify the core test statistics of the generalized likelihood ratio test detector, adaptive matched filter detector, and adaptive coherence estimator detector, effectively ensuring the accuracy, consistency, and repeatability of the feature extraction process. It provides a stable and effective input basis for the subsequent training of the shallow feedforward neural network, thereby further guaranteeing the performance stability of the entire radar adaptive detection method.

[0081] The above analysis shows that the test statistics of Kelly's GLRT, AMF, and ACE detectors can all be expressed as eigenvectors. The functions, and the specific correspondences are as follows: Figure 2 As shown.

[0082] Furthermore, the labeled data used during training. The values ​​are 1 and -1, corresponding to the presence and absence of useful signals (target signals) in the unit to be detected, respectively. Clearly labeled data can distinguish between the two detection scenarios, providing a clear target orientation for the training of the shallow feedforward neural network. This helps the model accurately learn the feature differences between the presence and absence of target signals, thereby improving the model's ability to identify the presence of target signals and ensuring the accuracy of subsequent radar target detection decisions.

[0083] The loss function for network training is defined using mean squared error:

[0084]

[0085] In the formula, Let be the mean squared error loss function of the shallow feedforward neural network. These are the training parameters in the neural network. For training data batch size, For the index of the training samples, For the first The output of the shallow feedforward neural network corresponding to each training sample. For the first The binary labels of each training sample. This is the weight matrix of the hidden layer. The bias vector of the hidden layer. For the weights of the output layer, The bias coefficients of the output layer. For the activation function of the hidden layer of the neural network, For the first Each training sample corresponds to a 3D feature vector. This loss function design based on mean squared error can accurately quantify the degree of deviation between the neural network output and the true label, providing a clear and explicit target guidance for gradient descent algorithms to optimize network parameters. This effectively ensures the effectiveness and stability of the neural network training process, ensuring that the model can stably output results that meet the requirements of radar target detection after training.

[0086] Furthermore, when using the gradient descent method, the first... The network parameter update expression for this step is:

[0087]

[0088] In the formula, For the first The updated network parameter set after the next iteration For the first The set of network parameters at the next iteration To update the step size, For gradient operators, The gradient operator is applied to the loss function. The results show that this update formula provides a clear and quantifiable basis for network parameter optimization, enabling precise iterative adjustment of parameters, effectively minimizing the loss function value, quickly correcting network parameter deviations that occur during the iteration process, ensuring that the shallow feedforward neural network stably learns the mapping relationship between the 3D feature vector and the radar target detection results, thereby improving the accuracy and reliability of the final detection decision.

[0089] After completing network training, the detector threshold needs to be adjusted according to the preset false alarm probability requirements. This can be achieved by adjusting the bias coefficients in the network. This ensures that the false alarm probability of the detector meets the specified conditions.

[0090] To verify the comprehensive performance of the radar adaptive detection method assisted by the shallow feedforward neural network of this invention, a comparative experiment was conducted using computer simulation data and measured sea clutter data from the IPIX radar. The analysis focused on the detection performance under small sample conditions, target Doppler frequency shift, clutter suppression capability, steering vector mismatch, and non-Gaussian and non-uniform scenarios, quantitatively demonstrating the advantages of this invention compared to traditional adaptive detectors.

[0091] Analysis was performed using computer simulation data. The simulation parameters were N=8 and K=20. The noise covariance matrix was defined as follows: ,in , guide vector ,in , In the simulation, to ensure a balance between positive and negative examples in the training data, it is assumed that... and The training data below are all 10 4 One, of which Below, the signal-to-noise ratio (SNR) in the data is uniformly distributed from -10dB to 30dB. The definition of SNR is:

[0092]

[0093] In the neural network, the number of hidden layer neurons is n=10. The Levenberg-Marquardt backpropagation algorithm is used for training. This algorithm has the inherent advantages of fast convergence speed and high training accuracy, and can efficiently optimize the core parameters such as the weight matrix and bias vector of the shallow feedforward neural network. It can effectively shorten the model training cycle and enhance the network's learning effect on the mapping relationship between the 3D feature vector and the detection decision result, providing a reliable guarantee for the detector to maintain high performance in complex scenarios such as small samples and non-Gaussian clutter. The training data is divided into training set, validation set, and test set according to 70%, 15%, and 15% respectively. The probability of false alarm (PFA) is set to 10. -3 .

[0094] First, we analyze the impact of the number of reference data on detection performance. The detection results are as follows: Figure 3 and Figure 4 As shown. Figure 3 For a small sample scenario with K=12 reference data, Figure 4This corresponds to a sufficient sample scenario with K=32 reference data points. The horizontal axis of both graphs represents the signal-to-noise ratio (SNR) in dB; the vertical axis represents the detection probability (Pd). The leftmost curve in the graph represents the theoretically optimal matched filter detector (MF); the other curves represent the neural network detector (NN) of this invention, as well as conventional adaptive detectors, including the generalized likelihood ratio test detector (GLRT), the adaptive matched filter detector (AMF), and the adaptive coherence estimator detector (ACE).

[0095] from Figure 3 As can be seen, under small sample conditions, the curve of the NN detector of this invention is significantly more to the left than that of GLRT, AMF, and ACE, with the leftward shift of the curve indicating superior performance. That is, at the same signal-to-noise ratio (SNR), the detection probability of the NN detector is significantly higher than these conventional adaptive detectors, especially in the SNR range of 10dB-15dB, where the performance improvement is even more pronounced. And... Figure 4 In scenarios with sufficient samples, the curves of NN, GLRT, and AMF tend to overlap, and their detection probabilities are quite similar.

[0096] This result verifies the core advantage of the present invention: with the assistance of artificial neural networks, significant performance gains are achieved under small sample conditions with limited reference data. This is because the neural network, trained on a 3D feature vector driven by a fusion model, can more accurately extract and distinguish target and clutter features in small sample scenarios, thereby improving detection performance. Furthermore, when reference data is sufficient, conventional adaptive detectors can already achieve superior performance through statistical estimation; therefore, except for the theoretically optimal MF, the performance differences among detectors are reduced.

[0097] Consider the detector's sensitivity to the target's Doppler frequency shift. If the target's Doppler frequency differs between the training and testing data, the detection performance will inevitably degrade. Here, the signal-to-noise ratio (SNR) is 15 dB, and the simulation results are as follows... Figure 5 As shown, the sensitivity of the NN detector in this invention is close to that of GLRT, while the AMF detector is insensitive to Doppler frequencies.

[0098] The detector's ability to suppress clutter mainly considers the impact of the deviation of the clutter Doppler center in the training and test data on the detection performance. Simulation results are as follows: Figure 6 As shown, the NN detector of this invention has a wider notch for clutter suppression, and its clutter suppression capability is superior to that of conventional adaptive detectors.

[0099] Guide vector mismatch refers to the misalignment of the guide vector assumed during the detector structure design phase. and the steering vector in the simulation data There is a difference between them, which can be represented by the cosine of the generalized angle between the two vectors:

[0100]

[0101] When the guide vector matches, i.e. You can get When two vectors are perfectly orthogonal, Therefore, the detection performance of the detector can be analyzed under different steering vector mismatch degrees and signal-to-noise ratio conditions. The analysis results are as follows: Figure 7 As shown, the performance of the NN detector of this invention under steering vector mismatch conditions is between that of GLRT and AMF.

[0102] Next, we will analyze the measured data, which comes from publicly available IPIX radar sea clutter data. The IPIX radar is an X-band multi-polarization radar. This invention considers two sets of data for testing detection performance: data 19980205_171203_ANTSTEP (1#) and 19980205_170935_ANTSTEP (2#). This data corresponds to a range resolution of 30 meters, a pulse repetition frequency of 1000 Hz, an acquisition time of 1 minute, and a total of 60,000 pulses, divided into 28 range cells. The amplitude distributions of the two sets of data are as follows: Figure 8 and Figure 9 As shown, where Figure 8 Corresponding to data #1, Figure 9 Corresponding to data #2, both graphs include HH horizontal polarization and VV vertical polarization channels. The horizontal axis represents time (in seconds), and the vertical axis represents distance units. The color intensity represents the magnitude of clutter amplitude. As can be seen from the graphs, both sets of measured data exhibit significant non-uniformity, and there are obvious echo fluctuations near the 7th distance unit (corresponding to a physical distance of approximately 210 meters). Since it is impossible to determine whether this fluctuation is caused by a real target, this invention considers it as an inherent non-uniformity of the scene.

[0103] Based on the analysis of measured data, we assume N=4, K=20, and PFA=10. -2 The neural network has 18 hidden layer units (n=18). The 14th distance unit is selected as the unit to be detected. Along the fast time direction, 20 distance units (4-13 and 15-24) are used as reference units. Along the slow time direction, a sliding window of length N is used to select data with a 50% overlap ratio, resulting in 29999 sets of training data. Since the clutter covariance matrix of the measured data is unknown, this invention considers the relationship between the detection probability and the useful signal amplitude α. The results are as follows: Figure 10 and Figure 11 As shown, in Figure 10In this study, polarization data of 1#HH was used as training data, and polarization data of 2#HH was used as testing data to examine the detector's generalization ability across different datasets. Figure 11 In this study, 1#HH polarization data was used as training data and 1#VV polarization data was used as testing data to examine the detector's generalization ability under different polarization conditions. It can be seen that due to the significant non-Gaussianity and non-uniformity of the measured data, the detection performance of conventional adaptive detectors based on uniform Gaussian clutter scenarios, namely Kelly's GLRT and AMF, significantly decreases, while the ACE detector, designed for non-uniform non-Gaussian clutter scenarios, performs better. The detection performance proposed in this invention is optimal when PD > 0.4. While its performance is slightly worse than ACE at lower detection probabilities, it is still superior to the other two detectors. This is because the NN detector can capture the nonlinearity in the data to the greatest extent possible, but the construction of the feature quantities does not consider the non-Gaussianity that may exist in the clutter.

[0104] In the field of radar signal detection, adaptive detection is a crucial core technology. However, from the perspective of hypothesis testing statistics, the performance of existing conventional adaptive detectors is not optimal, and there is still significant room for improvement. This invention addresses the problem of adaptive radar signal detection by innovatively integrating model-driven and data-driven approaches, taking advantage of both. It employs a neural network-assisted adaptive detection method, using a shallow feedforward neural network composed of a small number of neural units to classify features. Specifically, this invention proposes a targeted improvement scheme: extracting the three core statistics relied upon in the structure of conventional adaptive detectors to construct a dedicated feature space, enabling conventional adaptive detectors to be represented as functions of these feature vectors; simultaneously, replacing the original structure of traditional detectors with the aforementioned shallow feedforward neural network, fully leveraging the highly nonlinear fitting advantages of artificial neural networks to deeply explore the complex nonlinear correlation between feature vectors and detection results. Verified by both computer simulation data and measured sea clutter data, this fusion-based adaptive detector demonstrates excellent adaptability and detection performance. Especially under conditions of small sample size with limited reference data, its detection performance is significantly better than that of conventional adaptive detectors. It can consistently achieve better detection results in both uniform Gaussian noise and non-uniform non-Gaussian noise scenarios, and has better clutter suppression capabilities and generalization capabilities across datasets and polarizations, providing a more reliable technical solution for radar target detection in complex environments.

[0105] This invention also discloses a system for implementing the radar adaptive detection method that integrates model-driven and data-driven approaches, comprising: a data acquisition module for acquiring radar target unit data and reference unit data; the target unit data contains noise and may contain target signals, and is used for target signal presence determination; the reference unit data does not contain target signals and its noise statistical characteristics are consistent with those of the target unit, and is used to estimate the noise covariance matrix of the target unit; a feature construction module for constructing a 3D feature vector in the feature space based on the feature quantities relied upon by the model-driven adaptive detector, combined with the target unit data, reference unit data, and noise covariance matrix output by the data acquisition module; the model-driven adaptive detector includes a generalized likelihood ratio test detector, an adaptive matched filter detector, and an adaptive coherence estimator detector; and a network training module for constructing a network consisting of an input layer, a hidden layer, and an output layer. The shallow feedforward neural network consists of layers, with the number of input layer units matching the dimension of the 3D feature vector, and the output layer having a single output unit. It is also used to generate training 3D feature vectors and training data with binary labels based on the original data of the preset training dataset, according to the 3D feature vector construction method of the feature construction module. A gradient descent-like algorithm is also used to train the shallow feedforward neural network by minimizing the loss function, updating the network parameters, and allowing the shallow feedforward neural network to replace the decision structure of the model-driven adaptive detector. The detection decision module is used to input the 3D feature vector constructed by the feature construction module into the shallow feedforward neural network trained by the network training module to obtain the output scalar. It is also used to compare the output scalar with a preset threshold to obtain the radar target detection decision result. The preset threshold is determined by adjusting the bias coefficients of the output layer of the shallow feedforward neural network, with the goal of ensuring that the false alarm probability of the radar meets preset requirements.

[0106] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0107] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment includes only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A radar adaptive detection method integrating model-driven and data-driven approaches, characterized in that, Includes the following steps: S1, acquire the radar's target unit data and reference unit data; the target unit data contains noise and may contain target signals, and is used to determine the existence of target signals; the reference unit data does not contain target signals and its noise statistical characteristics are consistent with those of the target unit, and is used to estimate the noise covariance matrix of the target unit. S2, based on the feature quantities on which the model-driven adaptive detector depends, and combined with the data of the unit to be detected, the reference unit data and the noise covariance matrix in step S1, a 3D feature vector in the feature space is constructed. The model-driven adaptive detector includes a generalized likelihood ratio test detector, an adaptive matched filter detector and an adaptive coherence estimator detector. S3, construct a shallow feedforward neural network consisting of an input layer, hidden layers, and an output layer. The number of units in the input layer matches the dimension of the 3D feature vector, and the output layer is a single output unit. Based on the original data of the preset training dataset, generate training feature vectors and labeled training data according to the feature vector construction method in step S2. Use a gradient descent algorithm to train the shallow feedforward neural network by minimizing the loss function, update the network parameters, and make the shallow feedforward neural network replace the decision structure of the model-driven adaptive detector. S4. Input the feature vector constructed in step S2 into the shallow feedforward neural network trained in step S3 to obtain the output scalar; compare the output scalar with a preset threshold to obtain the radar target detection signal decision result; the preset threshold meets the radar's preset target false alarm probability requirement.

2. The radar adaptive detection method integrating model-driven and data-driven approaches according to claim 1, characterized in that, The three features of the 3D feature vector correspond to the test statistics of the generalized likelihood ratio test detector, the adaptive matched filter detector, and the adaptive coherence estimator detector, respectively.

3. The radar adaptive detection method integrating model-driven and data-driven approaches according to claim 2, characterized in that, In step S2, the 3D feature vector is constructed using the following formula: ； In the formula, It is a 3D feature vector. The target guiding vector, For the data of the unit to be detected, The noise covariance matrix estimated from the reference cell is shown in the superscript. This is the conjugate transpose operation. This is for the transpose operation.

4. The radar adaptive detection method integrating model-driven and data-driven approaches according to claim 3, characterized in that, In step S3, the hidden layer of the shallow feedforward neural network uses the hyperbolic tangent function as the activation function, which maps the real number axis to the interval (-1, 1).

5. The radar adaptive detection method integrating model-driven and data-driven approaches according to claim 4, characterized in that, In step S3, the labels of the training data are binary labels, including label 1 and label -1; wherein, label 1 indicates that the target unit has a target signal, and label -1 indicates that the target unit does not have a target signal.

6. The radar adaptive detection method integrating model-driven and data-driven approaches according to claim 1, characterized in that, In step S3, the gradient descent algorithm is the Levenberg-Marquardt backpropagation algorithm.

7. The radar adaptive detection method integrating model-driven and data-driven approaches according to claim 6, characterized in that, In step S3, the gradient descent algorithm updates the parameters of the shallow feedforward neural network using the following formula: ； In the formula, For the first The updated network parameter set after the next iteration For the first The set of network parameters at the next iteration To update the step size, For gradient operators, The gradient operator is applied to the loss function. The result.

8. The radar adaptive detection method integrating model-driven and data-driven approaches according to claim 7, characterized in that, In step S3, the loss function is defined using mean squared error and is used to quantify the deviation between the output of the shallow feedforward neural network and the training data labels. The expression is: ； In the formula, Let be the mean squared error loss function of the shallow feedforward neural network. For training data batch size, For the index of the training samples, For the first The output of the shallow feedforward neural network corresponding to each training sample. For the first The binary labels of each training sample. This is the weight matrix of the hidden layer. The bias vector of the hidden layer. For the weights of the output layer, The bias coefficients of the output layer. For the activation function of the hidden layer of a shallow feedforward neural network, For the first The 3D feature vector corresponding to each training sample.

9. The radar adaptive detection method integrating model-driven and data-driven approaches according to claim 8, characterized in that, In step S4, the preset threshold is determined by adjusting the bias coefficient of the output layer of the shallow feedforward neural network, with the goal of ensuring that the false alarm probability of the radar meets the preset requirements.

10. A system for implementing the radar adaptive detection method that integrates model-driven and data-driven approaches as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire the radar's data of the target unit and the reference unit. The data of the unit to be detected contains noise and may contain target signals, and is used to determine the existence of target signals; the data of the reference unit does not contain target signals and its noise statistical characteristics are consistent with those of the unit to be detected, and is used to estimate the noise covariance matrix of the unit to be detected. The feature construction module is used to construct a 3D feature vector in the feature space based on the feature quantities on which the model-driven adaptive detector depends, combined with the target cell data, reference cell data, and noise covariance matrix output by the data acquisition module; the model-driven adaptive detector includes a generalized likelihood ratio test detector, an adaptive matched filter detector, and an adaptive coherence estimator detector. The network training module is used to construct a shallow feedforward neural network consisting of an input layer, hidden layers, and an output layer. The number of units in the input layer matches the dimension of the 3D feature vector, and the output layer is a single output unit. It is also used to generate a 3D feature vector for training and training data with binary labels based on the original data of a preset training dataset, according to the 3D feature vector construction method of the feature construction module. Furthermore, it employs a gradient descent algorithm to train the shallow feedforward neural network by minimizing the loss function, updating the network parameters, and enabling the shallow feedforward neural network to replace the decision structure of the model-driven adaptive detector. The detection and decision module is used to input the 3D feature vector constructed by the feature construction module into the shallow feedforward neural network trained by the network training module to obtain the output scalar; it is also used to compare the output scalar with a preset threshold to obtain the radar target detection decision result; the preset threshold is determined by adjusting the bias coefficient of the output layer of the shallow feedforward neural network to ensure that the false alarm probability of the radar meets the preset requirements.

Images