A data-driven high-precision array direction finding method and system

Abstract

In the field of electronic countermeasures of the present invention, there is particularly involved a high-precision array direction finding method and system based on data driving. First, the present invention analyzes the characteristics of the covariance matrix of the array output under the condition of a single signal incident on a uniform linear array, extracts input features with small redundancy, and combines a convolutional neural network to construct an intelligent array direction finding framework; then, based on independent component analysis, the multi-signal direction finding problem is transformed into a single-signal direction finding problem, and the arrival direction of each signal is estimated one by one using the network trained with a single signal. This method reduces the dimension of the input feature vector and the number of samples required for network training, reduces the complexity of the direction finding model training, and at the same time effectively inherits the advantage of the machine learning technology in self-adapting to the model error.

Description

Technical Field

[0001] The present invention belongs to the field of electronic countermeasures and relates to a high-precision array direction finding method based on data driving. Background Art

[0002] Array direction finding has great application requirements in the field of electronic countermeasures. The current array direction finding systems mainly adopt two systems: spatial spectrum estimation and phase interferometer direction finding.

[0003] The spatial spectrum estimation system can realize direction estimation in the case of multiple signals with the same frequency and has high accuracy. However, it needs to perform spectral peak search, with a large amount of calculation and poor real-time performance in engineering applications. The phase interferometer direction finding system has a small amount of calculation and can perform real-time direction finding on a single radiation source, but it does not have the ability to find the direction of multiple signals with the same frequency. The above two direction finding systems are both based on accurate modeling. However, in the actual environment, the direction finding equipment is often affected by non-ideal factors such as array mutual coupling effect and multipath effect, resulting in the mismatch of the direction finding model and the inability of the direction finding accuracy to reach the performance index. Machine learning technology does not depend on a preset model and establishes a model through sample data learning, which can express the mapping relationship between signal reception data and the direction of the radiation source, so as to achieve high-precision direction finding under the condition of model error. The array direction finding method driven by data is expected to solve the bottleneck problem faced by traditional direction finding methods in model error, thereby effectively improving the application performance of the direction finding system in the actual environment.

[0004] Some literature uses unsupervised convolutional belief networks to extract angular features, and then inputs these extracted features into SVR for angle estimation. There is also literature that uses convolutional neural networks to solve the problem of array direction finding under error conditions. Simulations prove that it has extremely high direction estimation accuracy under the condition of microphone array position perturbation. Some literature proposes a deep learning direction finding framework that can adapt to various array errors. This framework adopts a structure combining autoencoders and deep neural networks. First, the autoencoder is used to perform spatial domain filtering on the received signal to roughly separate the signals in different regions, and then multiple deep neural networks are used to achieve high-precision direction finding for the signals in each relatively narrow region. Some literature uses the idea of hierarchical multi-granularity to divide the spatial domain and frequency domain, which alleviates the training pressure of the neural network to a certain extent. These methods are all designed for the case of single-signal incidence. When the number of signals increases, the number of training samples of the network will show an exponential explosion, resulting in difficulties in learning and the phenomenon of non-convergence. There is a publicly available literature that proposes a method for extracting array sparse covariance eigenvectors, which has strong adaptability under low signal-to-noise ratio conditions. This literature simulates and verifies that the direction finding network has a certain adaptability to the case of multi-signal simultaneous incidence, but does not conduct theoretical analysis on this, and the direction finding accuracy drops significantly under multi-signal conditions. To address the above problems, this patent proposes a data-driven high-precision array direction finding and estimation method. First, based on a uniform linear array, a method for extracting input features with low redundancy under single-signal conditions is studied, a convolutional neural network with good convergence performance is introduced, and a single-signal intelligent deep direction finding network framework is constructed. Then, using the independence of signals, the multi-signal direction finding problem is transformed into a single-signal direction finding problem, and the direction of arrival estimation of multi-signals is realized based on the single-signal training network.

[0005] Existing intelligent array direction finding principle: Assume that there are K narrowband signals, which are independent of each other, and are incident from the far field onto an M-element uniform linear array. Then the array output at time t can be expressed in matrix form as:

[0006]

[0007] where the vector x(t) = [x1(t), x2(t), …, x M (t)] T , x m (t) represents the output data of the m-th array element, m = 1, 2, …, M; the vector s(t) = [s1(t), s2(t), …, s K (t)] T , s k (t) represents the k-th incident signal, k = 1, 2, …, K; the array manifold matrix A = [a(θ1), a(θ2) …, a(θ K )], the steering vector θ kdenotes the direction of arrival of the k-th incident signal, c is the speed of light, and v(t) = [v1(t), v2(t), …, v M (t)] T , where T represents the transpose, and v i (t) represents Gaussian white noise with power σ 2 .

[0008] Spatial spectrum estimation methods usually build an objective function based on the array output model and achieve direction finding through spectrum peak search. However, in the actual environment, due to the influence of non-ideal factors such as wide-open signal frequencies, multiple signals incident simultaneously, and array mutual coupling effects, the array manifold matrix constructed in the model is not very accurate and is difficult to accurately calibrate, which makes the spatial spectrum estimation method perform poorly in practical applications. The direction finding method based on machine learning transforms the signal angle estimation problem into a classification problem, and the principle is as Figure 1 shown.

[0009] The airspace is divided into several non-overlapping intervals according to a certain angular spacing. For example, [90°, 90°) is divided at intervals of 1°, and can be divided into [90°, 89°), [89°, 88°), …… [89°, 90°), a total of 180 angular intervals. The angles within the same interval are regarded as the same category, while the angles in different intervals are different categories, corresponding to a total of 180 categories. The array direction finding problem can be equivalent to the problem of judging which angular interval the array observation data comes from. Then, machine learning technology is introduced to construct a classification network to achieve direction finding. This type of method can make the direction finding network adapt to non-ideal factors such as array element position errors and mutual coupling effects through data driving. To reduce the dimensionality of training samples, machine learning-based direction finding methods usually use the upper triangular matrix of the covariance matrix as the feature input because the upper triangular matrix of the covariance eliminates the influence of the signal waveform and completely retains all information related to the direction of incident signals. However, the training samples need to traverse the combinations of signal frequencies, the number of signals, and the angular space. When the signal frequency is in a relatively wide frequency band range, the direction finding range is large, and multiple signals are incident simultaneously, the number of required training samples will increase exponentially, resulting in difficult network training convergence and difficulties in engineering practical applications. Summary of the Invention

[0010] Aiming at the deficiencies of the prior art, the present invention discloses a high-precision array direction finding method based on data driving. The present invention extracts the input features of single-signal input and trains the array direction finding network, extracts the input feature vectors of each signal in the multi-signal array through the separation matrix, and then inputs them into the trained single-signal network to obtain the direction finding results of each signal in turn. Thus, high-precision array direction finding of multiple signals is achieved.

[0011] The technical solution of the present invention is: a high-precision array direction finding method based on data driving, which is characterized by including the following steps:

[0012] Step 1: Single-signal intelligent array direction finding network, including the following sub-steps:

[0013] Step 1.1 Single-signal input feature extraction

[0014] In the case of single-signal incidence, the array covariance matrix R is:

[0015] R = E[x(t)x H (t)] = a(θ)a H (θ)η + σ 2 I M (2)

[0016] Where: θ is the incident direction of the signal, η is the power of the incident signal, σ 2 is the noise power, I M is the M×M identity matrix, H represents conjugate transpose, and under the condition of a uniform linear array, the (pp, qq) element of R is:

[0017]

[0018] pp and qq respectively represent the row number and column number value of the matrix R. The first column vector of R is defined as the vector

[0019]

[0020] In formula (4), d represents the distance between the array elements of the antenna array. Take out the real part and imaginary part of the vector and normalize it to be used as the input feature vector z:

[0021]

[0022] In formula (5), real represents the real part of the vector z, and imag represents the imaginary part of the vector z;

[0023] Step 1.2 Training of the array direction finding network based on a convolutional neural network

[0024] Using the input feature vector z as the input, four convolutional network layers are used in the middle to learn and understand the mapping relationship between the input feature vector and the DOA for training the array direction finding network; first, the angular region is divided into several non-overlapping regions according to the angular spacing, and each region corresponds to an output node; then, the output nodes are encoded according to 01, and the node with an output value of 1 corresponds to the incoming wave direction, and the output values of other nodes are 0. According to the rules of region division, multiple groups of array data are collected to form a training sample D train ={(z 1, β 1 ), (z 2 , β 2 ), …, (z D , β D )}, z n represents the nth group of input sample feature vectors of the collected data, and β n is the nth group of target output vectors of the collected data, and the element values correspond to the 01 coding rule of the output node;

[0025]

[0026] Step 3 Multi-signal intelligent array direction finding network

[0027] The initial value of the separation matrix W is randomly generated according to the Gaussian function generation mode according to the size of the input data volume, so that:

[0028] WA = PΛ (8)

[0029] In the formula, P is a permutation matrix, and Λ is the signal of a diagonal matrix; through blind separation, the estimated A = W -1 PΛ of the array manifold matrix is obtained; it is defined that Then Each column vector of contains the direction information of an independent incident signal. Assuming that the column vector corresponds to the direction of θ k , then should have the following characteristics:

[0030]

[0031] where ρ represents the diagonal element in the diagonal matrix Λ corresponding to the column vector ; take the real part and imaginary part of and perform normalization processing to obtain the vector h k :

[0032]

[0033] Extract the input feature vector h of each signal one by one k , and input it into the trained single-signal network to obtain the direction finding results of each signal in turn.

[0034] Another object of the present invention is to provide a confidence ranking positioning system based on multi-propagation mode recognition, including:

[0035] A signal receiving array, configured in a radar system, for receiving and capturing signals emitted by an air target;

[0036] A signal processing unit, connected to the signal receiving array, for realizing intelligent single-signal array direction finding, including the calculation of the signal covariance matrix and feature extraction;

[0037] A convolutional neural network processor, receiving the feature vectors output by the signal processing unit, for learning and predicting the direction of the signal;

[0038] A multi-signal separation module, for separating and independently processing each signal when multiple signals are received;

[0039] A user interface, for displaying the signal direction finding results, including the direction and position of the target.

[0040] Another object of the present invention is to provide a high-precision array direction finding system based on data driving, including:

[0041] A single-signal intelligent array direction finding network, for extracting the array covariance matrix features in the case of single-signal incidence and performing normalization processing to form input feature vectors;

[0042] A convolutional neural network training module, for learning and understanding the mapping relationship between the input feature vectors and the direction of arrival (DOA);

[0043] A multi-signal intelligent array direction finding network, for processing multi-signal situations, including signal separation, feature extraction, and obtaining the direction finding results of each signal using the single-signal network model.

[0044] Wherein the single-signal intelligent array direction finding network includes:

[0045] A feature extraction unit, for processing the array covariance matrix and extracting signal features;

[0046] A normalization unit, for normalizing the extracted signal features to generate input feature vectors.

[0047] Including:

[0048] A neural network training unit, for training the convolutional neural network so as to be able to map the input feature vectors to the corresponding DOA outputs.

[0049] Wherein the multi-signal intelligent array direction finding network includes:

[0050] A signal separation unit, for separating independent signals from multi-signal data;

[0051] A signal feature extraction unit, for extracting the features of each independent signal and performing normalization processing.

[0052] Its features also include:

[0053] A test processing unit for extracting input feature vectors from the received array data and inputting them into a trained network to obtain a direction finding result;

[0054] A direction finding result processing unit for determining the direction finding result according to the network output.

[0055] Among them, the feature extraction process includes:

[0056] A data whitening processing unit for whitening the array received data;

[0057] An iterative update unit for updating the separation matrix through an iterative algorithm to optimize the independence of the signals;

[0058] A separation matrix construction unit for estimating and constructing a separation matrix, and then obtaining the feature vectors of the independent signals.

[0059] This solution comprehensively covers the key components and working processes of a high-precision array direction finding system, from the processing of single signals to multi-signals, and then to the training and application of neural networks, ensuring that the system can efficiently and accurately determine directions.

[0060] Combined with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by the present invention are:

[0061] First, first analyze the characteristics of the covariance matrix of the array output under the condition of a single signal incident on a uniform linear array, extract input features with small redundancy, and combine with a convolutional neural network to construct an intelligent array direction finding framework; then, based on independent component analysis, transform the multi-signal direction finding problem into a single-signal direction finding problem, and use the network trained with a single signal to estimate the arrival directions of each signal one by one. This method reduces the dimension of the input feature vector and the number of samples required for network training, reduces the complexity of the direction finding model training, and at the same time effectively inherits the advantage of the machine learning technology's self-adaptation to model errors. Experimental results show that the method of this patent has high direction finding accuracy in complex environments such as low signal-to-noise ratio, mutual coupling effect, and multi-signal incident at the same time.

[0062] Second, the present invention effectively solves the problems of accuracy, noise interference, and efficiency in traditional array direction finding technology by combining advanced signal processing technology and convolutional neural networks, and realizes high-precision direction finding in multi-signal and complex environments.

[0063] Improve accuracy: This method significantly improves the direction finding accuracy through advanced feature extraction and intelligent algorithms.

[0064] Anti-noise ability: Through convolutional neural networks and complex signal processing technology, the present invention can still maintain high accuracy in a noisy environment.

[0065] Processing efficiency: By leveraging data-driven and intelligent technologies, the present invention can efficiently process large amounts of data and reduce latency.

[0066] Third, the significant technological advancements brought about by the data-driven high-precision array direction-finding method include:

[0067] 1. High-precision signal direction estimation

[0068] By applying convolutional neural networks and complex mathematical models, this method can accurately estimate the incident directions of single signals or multiple signals. Such high-precision direction-finding capabilities are difficult to achieve with traditional methods, especially in a multi-signal environment.

[0069] 2. Enhanced signal processing capabilities

[0070] Traditional direction-finding methods are not efficient in processing complex or overlapping signals. The new method, through a data-driven approach, especially by using machine learning and convolutional neural networks, improves the ability to handle complex signal environments and can more effectively separate and identify overlapping signals.

[0071] 3. Automated and intelligent direction-finding process

[0072] This method automates the direction-finding process, reducing manual intervention and subjective errors. The use of neural network models makes the direction-finding process more intelligent and capable of learning and adapting to different signal characteristics.

[0073] 4. Improvement in real-time performance

[0074] Compared with traditional direction-finding methods that require long-time processing and analysis, the new method significantly improves the real-time performance of direction-finding through real-time data processing and fast neural network prediction, and is suitable for real-time or near-real-time signal direction-finding requirements.

[0075] 5. Wide range of applications

[0076] This high-precision and automated direction-finding method is not only applicable to traditional communication and signal processing fields, but can also be extended to multiple high-tech fields such as wireless positioning, navigation, and drone control.

[0077] 6. Improved reliability of signal detection

[0078] Using advanced data processing and feature extraction technologies, this method improves the reliability of signal detection, reduces the probability of false positives and omissions, especially in signal detection in a noisy environment.

[0079] In summary, this data-driven high-precision array direction-finding method has achieved significant technological advancements in the field of signal processing, improving the accuracy, efficiency, and reliability of direction-finding. Description of the Drawings

[0080] Figure 1 It is the direction finding principle diagram provided by the embodiments of the present invention.

[0081] Figure 2 It is the network structure diagram provided by the embodiments of the present invention.

[0082] Figure 3 It is the direction finding process provided by the embodiments of the present invention.

[0083] Figure 4 It is the direction finding error at different signal-to-noise ratios provided by the embodiments of the present invention.

[0084] Figure 5 It is the direction finding error at different numbers of snapshots provided by the embodiments of the present invention.

[0085] Figure 6 It is the direction finding error at different signal-to-noise ratios (when 3 signals are incident simultaneously) provided by the embodiments of the present invention.

[0086] Figure 7 It is the DOA estimation error of multiple signals at different numbers of snapshots (when 3 signals are incident simultaneously) provided by the embodiments of the present invention.

[0087] Figure 8 It is the direction finding estimation error at different signal-to-noise ratios (mutual coupling effect) provided by the embodiments of the present invention.

[0088] Figure 9 It is the direction finding estimation error at different numbers of snapshots (mutual coupling effect) provided by the embodiments of the present invention. Detailed implementation manners

[0089] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0090] For the data-driven high-precision array direction finding method, the following are two specific embodiments and their implementation schemes:

[0091] Application Embodiment 1: Wireless signal source positioning

[0092] 1) Data acquisition: Arrange multiple array receiving stations in a city or other environments. Each station is equipped with a high-performance signal receiving array for capturing wireless signals.

[0093] 2) Feature extraction and training: Extract features from the captured signals, including the covariance matrix, power spectrum, etc. of the signals, and then input these features into a pre-trained convolutional neural network for estimating the direction of the signal source.

[0094] 3) Real-time positioning and analysis: Utilize neural networks to rapidly predict the direction of the signal source, and combine data from multiple stations for triangulation positioning, thereby achieving precise determination of the signal source's location.

[0095] 4) Data feedback and optimization: Continuously adjust and optimize the neural network model based on the actual positioning results to improve the accuracy of positioning and the robustness of the system.

[0096] Application Example 2: Tracking and management of aviation communication signals

[0097] 1) Array deployment: Deploy a receiving array with multiple array elements at key positions at airports or on air routes to receive communication signals from aircraft.

[0098] 2) Signal processing and feature extraction: The received signals are processed to extract features including the array covariance matrix, etc., and are normalized.

[0099] 3) Application of neural network: Use a convolutional neural network to analyze the signal features and estimate the direction of the signal source (aircraft). The network model should be able to adapt to the signal features of different aircraft models.

[0100] 4) Flight track monitoring and analysis: Combine data from multiple array stations to track the flight track of aircraft in real time. Utilize the analysis results for flight management and scheduling, as well as airspace safety monitoring.

[0101] These two application examples demonstrate the flexibility and effectiveness of the data-driven high-precision array direction finding method in practical applications. Whether it is for wireless signal source positioning in an urban environment or for signal tracking and management in the field of aviation communication, this method can provide accurate direction finding results to support relevant operations and activities.

[0102] Two specific application examples of the embodiments of the present invention are:

[0103] Example 1: Radar signal positioning

[0104] This solution can be used in radar systems for precisely positioning the directions of objects such as aircraft and missiles.

[0105] Intelligent array direction finding for single signal: The single signal received by the radar undergoes feature extraction through the array covariance matrix. The extracted feature vector is input into a pre-trained convolutional neural network, and the network learns and identifies the direction of the signal (DOA).

[0106] Intelligent array direction finding for multiple signals: When the radar receives multiple signals simultaneously, the signals are separated and processed through blind signal separation technology. The trained CNN model is used for direction finding for each independent signal to obtain the accurate positions of each object.

[0107] Example 2: Optimization of Wireless Communication Networks

[0108] For wireless communication base stations, optimizing signal coverage and locating mobile devices.

[0109] Intelligent single-signal array direction finding: After the base station receives a signal from a specific mobile device, it uses the covariance matrix to extract signal features. These features are processed through a convolutional neural network to determine the direction of the mobile device.

[0110] Intelligent multi-signal array direction finding: When the base station processes signals from multiple devices simultaneously, it uses the blind signal separation method to process each signal separately. The direction finding results of each signal help the base station optimize the signal coverage range and improve communication quality and efficiency.

[0111] The present invention mainly improves the following problems and defects of the existing technologies and achieves significant technological progress:

[0112] Accuracy limitation: Traditional array direction finding technologies have limitations in signal processing accuracy, especially in multi-signal or complex signal environments.

[0113] Noise interference: In a noisy environment, traditional methods are difficult to accurately determine the direction of a signal.

[0114] Data processing efficiency: Existing technologies are not efficient in processing a large amount of or high-speed data, resulting in latency problems.

[0115] In view of the problems existing in the existing technologies, the technical solution adopted by the present invention is as follows:

[0116] Improvement in feature extraction: The present invention improves the extraction accuracy of single-signal features through advanced signal processing methods, such as using the array covariance matrix.

[0117] Training based on convolutional neural network: Using a convolutional neural network (CNN) to learn the mapping relationship between the input feature vector and the direction of arrival (DOA) of the signal improves the direction finding accuracy.

[0118] The present invention effectively processes multi-signal situations, solves the accuracy problem of traditional methods in multi-signal environments through blind signal separation and array manifold matrix estimation, and automatically adjusts and optimizes the signal processing flow by using data-driven methods, improving the processing efficiency and accuracy.

[0119] As Figures 1 to 9 shown, a data-driven high-precision array direction finding method of the present invention includes the following steps:

[0120] Step 1: Intelligent single-signal array direction finding network, including the following sub-steps:

[0121] Step 1.1 Input feature extraction

[0122] The covariance matrix R contains the azimuth information of all incident signals and eliminates the influence of signal waveform changes. It is the most widely used input feature at present. In the case of single-signal incidence, the array covariance matrix R can be expressed as:

[0123] R = E[x(t)x H (t)] = a(θ)a H (θ)η + σ 2 I M (2)

[0124] where: θ is the incident direction of the signal, η is the power of the incident signal, σ 2 is the noise power, I M is the M×M identity matrix, and H represents the conjugate transpose. Combining Equation (1) and Equation (2), it can be seen that under the condition of a uniform linear array, the (pp, qq)th element of R can be expressed as

[0125]

[0126] pp and qq represent the row and column numbers of matrix R respectively. By analyzing the above formula, it can be seen that for a uniform linear array, R has the conjugate symmetry property, and the elements on the same diagonal are equal. Therefore, the first column vector of R contains all the information of the covariance matrix. To further reduce data redundancy and the input dimension of the network, this patent considers using the first column vector of the covariance matrix as the input feature. The first column vector of R is defined as the vector

[0127]

[0128] In Equation (4), d represents the distance between the antenna array elements. Take out the real and imaginary parts of the vector and normalize them as the input feature vector z:

[0129]

[0130] In Equation (5), real represents the real part of vector z, and imag represents the imaginary part of vector z. It can be seen from the above formula that the dimension of the input feature vector z is 2M, while the covariance matrix has M 2 elements. After splitting into real and imaginary parts and used as the input feature vector, the dimension becomes 2M 2 , and some literature uses the upper triangular matrix of the covariance matrix as the input, with a dimension of 2M(M - 1). Obviously, the dimension of the input feature vector proposed in this patent is greatly reduced, and at the same time, the effective information contained in the covariance matrix is not lost.

[0131] Step 1.2 Training of the array direction-finding network based on the convolutional neural network

[0132] The main difficulty faced in using a neural network to learn and train the spatial direction of incoming waves is that the amount of training sample data is extremely large, resulting in the inability of the network to converge. Therefore, when selecting a neural network, it is necessary to focus on networks with fewer parameters and faster convergence speeds. The neural network based on convolutional operations has the characteristics of sparse connections and weight sharing. Each convolutional kernel only focuses on a local area, and the convolutional kernel parameter weights used in different local areas are the same. This greatly reduces the network parameters and significantly improves the convergence performance. This patent proposes to construct an array direction-finding network based on a convolutional neural network, as Figure 2 shown. Taking the input feature vector z as the input, four convolutional network layers are used in the middle to learn and understand the mapping relationship between the input feature vector and the DOA. Assume that the sizes of the convolutional kernels of the four convolutional layers are 25*12, 15*6, 5*3, and 3*1 respectively, the convolutional step size is 1, and finally a fully connected layer is added to output the direction-finding result. Using the network for direction finding includes two parts: training and testing. In the training part, first, the angle region of interest needs to be divided into several non-overlapping regions according to a certain angle interval, and each region corresponds to an output node. Then, the output nodes are encoded according to 01. The node with an output value of 1 corresponds to the direction of the incoming wave, and the output values of other nodes are 0. According to the region division rule, multiple groups of array data are collected to form the training sample D train ={(z 1 ,β 1 ),(z 2 ,β 2 ),…,(z D ,β D )}, (Should D be changed to n?) z n represents the nth group of input sample feature vectors of the collected data, and β n is the nth group of target output vectors of the collected data (n represents the sequence number of the collected data), that is, the sample label, and its element values correspond to the 01 encoding rule of the output node. In this patent, the training sample can be trained using the existing collected data.

[0133] The convolutional neural network is trained by minimizing the error between the actual network output azimuth angle and the target output azimuth angle β to adjust the network parameters, Ω = {κ l ,b l} l＝1,…,4 , κ l , b l represent all the convolutional kernels and biases of the lth layer. The optimized objective function is:

[0134]

[0135] Usually, when training the network, it is necessary to first solve the gradient of L(Ω) with respect to Ω Then update the parameters in the opposite direction of the gradient to minimize the loss function, thereby achieving the purpose of training the network model.

[0136] During the testing process, input feature vectors are extracted from the array received data to obtain z, which is then input into the trained network, and the output vector The position corresponding to the peak is the direction finding result. Since according to the encoding rule of the target output, the position of the output vector is obtained through spatial division and is a discrete angle set, a certain quantization error will be introduced. The magnitude of the quantization error is related to the size of the spatial division angle interval. To reduce the influence of the quantization error, linear interpolation is used to improve the estimation result. After obtaining the output vector The direction finding result θ is obtained through the following formula:

[0137]

[0138] In formula (7): φ, respectively represent the direction values corresponding to the left and right sides of the maximum peak in, ξ1 and ξ2 are the vector element values at the corresponding positions of φ and respectively.

[0139] Step 2 Multi-signal intelligent array direction finding network

[0140] When multiple signals are incident simultaneously, the training samples need to traverse multiple combinations of incident angles in the spatial domain, and the amount of training data increases explosively, making it impossible for the network to converge during training. To solve the above problems, in the testing stage, this paper proposes a feature extraction method based on independent component analysis. Using the independence of signals, multiple single-signal feature vectors similar to formula (5) are decomposed from the array received data, and the multi-signal network direction finding problem is transformed into a single-signal network direction finding problem. The decomposed feature vectors are respectively input into the single-signal network trained in Section 2, and the direction finding results of each signal can be obtained. For the array output model, according to the blind source separation theory, using the independence of signals, a separation matrix W can be found. The initial value of the separation matrix W is randomly generated according to the Gaussian function generation mode according to the size of the input data volume, so that:

[0141] WA = PΛ (8)

[0142] In the formula, P is a permutation matrix, and Λ is the signal of a diagonal matrix. It can be seen from the above formula that through blind source separation, an estimate of the array manifold matrix A = W-1PΛ can be obtained. Define Then Each column vector of contains the direction information of an independent incident signal, and must be in a proportional relationship with a certain column vector of A). Assume that the column vector corresponds to the direction of θk , then should have the following characteristics:

[0143]

[0144] where ρ represents the diagonal element in the diagonal matrix Λ corresponding to the column vector . Take the real and imaginary parts of the vector and perform normalization to obtain the vector h k :

[0145]

[0146] Comparing and analyzing formulas (4), (5) with formulas (9), (10), it can be seen that when the incident angles are equal, h k is equal to z. Therefore, inputting the vector h k to the trained single-signal network can obtain accurate direction-finding results. For the direction-finding problem of multiple signals incident simultaneously, the input feature vectors h k of each signal can be extracted one by one based on the blind separation algorithm and input to the trained single-signal network to obtain the direction-finding results of each signal in turn. As Figure 3 shown.

[0147] Since the array reception model is a complex model, the classical complex independent component analysis algorithm FastICA is used in this paper to obtain the separation matrix W, and on this basis, the feature input vector is further extracted. The specific feature extraction process is as follows:

[0148] 1) Whiten the array reception data x(t). First, calculate the covariance matrix R x of x(t), and then perform orthogonal decomposition on R x :

[0149] R x = E[x(t)x(t) H = QΛQ H (11)

[0150] In the formula: Λ is a diagonal matrix, and Q is an orthogonal matrix. The data v(t) is obtained by whitening the data through the following formula:

[0151] v(t) = Λ -1 / 2 Q H x(t) (12)

[0152] 2) Assume that w q is the q-th column vector of the separation matrix W, initialize the vector w q , and set the initial value as w q (0). Calculate the covariance matrix P vv = E[v(t)v(t)H . Assume that N is the number of array snapshots, n = 1, 2, …, N, let i = 1, and y(n) = w q (i - 1)v(n). Update the vector w using the following formula q .

[0153]

[0154] 3) Assume that the estimated vectors are w1, w2, …, w q-1 , then:

[0155]

[0156]

[0157] 4) If then obtain w q , otherwise, let i ← i + 1 and return to step 3).

[0158] 5) Let q ← q + 1. If q < M, return to step 2) until all column vectors are estimated to obtain the separation matrix W.

[0159] 6) Calculate Assume is the k-th column vector, take the real and imaginary parts of it, and after normalizing them, use them as the input feature vector h k , and sequentially obtain h 1, h 2, …, h K .

[0160]

[0161] The present invention compares and analyzes the direction-finding performance of different methods through simulation experiments. The antenna array is set as an 8-element uniform linear array, and the element spacing is set as half of the wavelength of the incident signal. In the simulation, the direction-finding error is used to evaluate the direction-finding performance of the algorithm. The direction-finding error E θ is defined as follows:

[0162]

[0163] In the formula: P is the number of test samples, K is the number of source signals, and correspond to the estimated value and the true value of the k-th signal direction at the p-th test respectively.

[0164] Single-signal incoming wave direction estimation: In this experiment, the direction-finding performances of the MUSIC method, the deep neural network method (AEDNN), and the method of this patent are compared and analyzed under the condition of single-signal incidence. The training samples are generated by simulating the array reception model. Considering single-signal incidence, the number of snapshots is 256, the angle range is [60°, 60°], the step is 0.1°, the signal-to-noise ratio range is [10 dB 10 dB], and the step is 1 dB. A total of 1201 * 21 = 25221 training samples are generated. Subsequently, two groups of test samples are generated by simulating the array reception model. Also considering single-signal incidence, 100 angle values are randomly selected within the range of [60°, 60°] for the incident angle. For the first group of test samples, the direction-finding accuracies of the three different methods under different signal-to-noise ratio conditions are tested. Assuming the number of snapshots is 256 and the signal-to-noise ratio varies in the range of [10 dB 10 dB] with a step of 5 dB, a total of 100 * 5 = 500 test samples are generated; for the second group of test samples, the estimation performances of the three different methods under different numbers of snapshots are tested. The signal-to-noise ratio is set to 2 dB, and the number of snapshots varies in the range of [100 1000] at intervals of 100, and a total of 100 * 10 = 1000 test samples are generated.

[0165] The MUSIC method, the AEDNN method, and the method of this paper are respectively used to estimate the direction of the above two groups of test samples, and the direction-finding error curves are obtained, as Figure 4 , Figure 5 shown. It can be seen from the figure that under different signal-to-noise ratio and different number of snapshots conditions, the direction-finding accuracies of the method of this patent and the AEDNN method are relatively close, and both are better than the MUSIC method. This shows that the data-driven methods achieve higher estimation accuracies than the traditional methods, and the method of this paper can still achieve an estimation performance similar to that of the AEDNN method when the input feature dimension is greatly reduced.

[0166] Multi-signal incoming wave direction estimation: Analyze the direction-finding performance of the MUSIC algorithm, the DCNNDOA method, and the method proposed in this paper when multiple signals arrive simultaneously. The AEDNN method is difficult to adapt to the multi-signal situation. In this paper, the DCNNDOA method with multi-signal adaptability is selected for comparison. The training set of the DCNNDOA method is the same as that given in the literature. The array direction-finding network of the method proposed in this paper uses the network trained in Experiment 1. The test samples are generated by simulating the array reception model. Considering three signals arriving simultaneously, 100 angle values are randomly selected within the range of [60°, 60°], and the signal angle interval is 2°. For the first group of test samples, the estimation performance of the algorithm at different signal-to-noise ratios is tested. The number of snapshots is set to 256, and the signal-to-noise ratio varies in the range of [10dB, 10dB] at intervals of 5dB, generating a total of 100 * 5 = 500 test samples; for the second group of test samples, the estimation performance of the algorithm at different numbers of snapshots is tested. The signal-to-noise ratio is set to 2dB, and the number of snapshots varies in the range of [100, 1000] at intervals of 100, generating a total of 100 * 10 = 1000 test samples.

[0167] The MUSIC method, the DCNNDOA method, and the method proposed in this paper are respectively used to estimate the directions of the above two groups of test samples, and the direction-finding error curves are obtained, as Figure 6 、 Figure 7 shown. It can be seen from the figure that the direction-finding performance of the method proposed in this paper is the best, and the estimation performance of the DCNNDOA method is the second. The DCNNDOA method has a certain generalization ability for the number of signals. However, since it is not trained specifically for the multi-signal situation, the obtained direction-finding accuracy is not high. The method proposed in this paper has a strong generalization ability for the number of signals. The simulation results show that when the signal-to-noise ratio is 2dB, the number of snapshots is 500, and three signals arrive simultaneously, the average root mean square error of direction-finding can reach about 0.35°. It should be noted that the number of signals that can be separated by the blind separation method is not greater than the number of array elements. Therefore, the number of signals that the method proposed in this paper can adapt to is not greater than the number of array elements. At the same time, affected by the fact that the blind separation method cannot separate multipath signals, the method of this patent cannot adapt to the multipath environment.

[0168] Adaptability to mutual coupling effect: This method is data-driven. Through learning and training a large amount of sample data, it can adapt to various inherent errors of the array. Here, a relatively typical mutual coupling error situation is selected for simulation experiment analysis. The direction-finding performances of two typical mutual coupling correction direction-finding methods, the SS Method and the YL Method, and the method of this patent are compared. To add the influence of the mutual coupling effect to the simulation data, the steering vector in the simulation model is corrected when generating the array reception data by simulation, where C represents the mutual coupling error matrix.

[0169]

[0170] where \(L\) represents the degrees of freedom, and \(\mathbf{b}=[b 1, b_2,\cdots,b L-1 \) T represents the mutual coupling coefficient vector. In the simulation experiment, it is assumed that the degrees of freedom is 3 and the mutual coupling coefficient vector is \(\mathbf{b} = [0.4 + 0.3j, 0.3 + 0.1j]\) T . Both the training samples and the test samples are generated by simulating the corrected array output model. For the training samples, single-signal incidence is considered, the number of snapshots is 256, the angle range is \([60^{\circ}, 60^{\circ}]\), the step is \(0.1^{\circ}\), and the signal-to-noise ratio range is \([10\mathrm{dB}, 10\mathrm{dB}]\), with a step of \(1\mathrm{dB}\). A total of \(1201\times21 = 25221\) training samples are generated. For the test samples, three signals are considered to be incident simultaneously, and 100 angle values are randomly selected within the range \([60^{\circ}, 60^{\circ}]\) for the incident angles, with a signal angle interval of \(2^{\circ}\). For the first group of test samples, the estimation performance of the algorithm at different signal-to-noise ratios is tested. Assuming the number of snapshots is 256 and the signal-to-noise ratio varies in the range \([10\mathrm{dB}, 10\mathrm{dB}]\) with a step of \(5\mathrm{dB}\), a total of \(100\times5 = 500\) test samples are generated; for the second group of test samples, the estimation performance of the algorithm at different numbers of snapshots is tested. Assuming the signal-to-noise ratio is \(2\mathrm{dB}\) and the number of snapshots varies in the range \([100, 1000]\) with an interval of 100, a total of \(100\times10 = 1000\) test samples are generated.

[0171] The SS Method, the YL Method, and the method proposed in this paper are respectively used to estimate the directions of arrival of the above two groups of test samples, and the direction-of-arrival error curves are obtained, as shown in Figure 8 and Figure 9 . It can be seen from the figure that under different signal-to-noise ratios and different numbers of snapshots, the direction-of-arrival errors of the method proposed in this paper are lower than those of the other two methods, which fully demonstrates that the method proposed in this paper has better error adaptability compared with the traditional model calibration methods.

[0172] The present invention uses a data-driven method for array direction finding, which has stronger adaptability to complex environments compared with traditional direction finding methods based on mathematical models. It should be noted that the implementation mode of the present invention can be realized through hardware, software, or a combination of software and hardware. The hardware part can be realized by using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. Those of ordinary skill in the art can understand that the above-mentioned devices and methods can be realized using computer-executable instructions and / or included in processor control code, for example, such code is provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be realized by hardware circuits of programmable hardware devices such as very large scale integrated circuits or gate arrays, semiconductors such as logic chips and transistors, or field programmable gate arrays and programmable logic devices, can also be realized by software executed by various types of processors, and can also be realized by a combination of the above hardware circuits and software, such as firmware.

[0173] As described above, the above are only specific implementation modes of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A data-driven high-precision array direction finding method, characterized in that Including the following steps: Step 1: Single-signal intelligent array direction-finding network, including the following sub-steps: Step 1.1 Single-signal input feature extraction In the case of single-signal incidence, the array covariance matrix R is: R = E[x(t)x H (t)] = a(θ)a H (θ)η + σ 2 I M (2) Where: θ is the incident direction of the signal, η is the power of the incident signal, and σ 2 is the noise power, I M is the M×M identity matrix, H represents the conjugate transpose, and the (pp, qq)-th element of R under the condition of a uniform linear array is: pp and qq represent the number of rows and columns of matrix R respectively, and the first column vector of R is defined as the vector In formula (4), d represents the distance between the elements of the antenna array. Take out the real part and the imaginary part of the vector and, after normalization, use them as the input feature vector z: In Equation (5), real represents the real part of vector z, and imag represents the imaginary part of vector z; Step 1.2 Training of the array direction-finding network based on a convolutional neural network; Taking the input feature vector z as the input, four convolutional network layers are used in the middle to learn and understand the mapping relationship between the input feature vector and the DOA for the training of the array direction finding network; first, the angular region is divided into several non - overlapping regions according to the angular interval, and each region corresponds to an output node; then, the output nodes are encoded according to 01 coding, and the node with the output value of 1 corresponds to the incoming wave direction, and the output values of other nodes are 0. According to the rules of region division, multiple groups of array data are collected to form the training sample D train ={(z 1 ,β 1 ),(z 2 ,β 2 ),…,(z D ,β D )}, z n represents the n - th group of input sample feature vectors of the collected data, and β n is the n - th group of target output vectors of the collected data, and the element values correspond to the 01 coding rule of the output nodes; Step 2 Multi-signal intelligent array direction-finding network The initial value of the separation matrix W is randomly generated according to the Gaussian function generation mode according to the size of the input data volume, so that: WA = PΛ (7) where \(P\) is a permutation matrix and \(\Lambda\) is the signal of a diagonal matrix; through blind separation, the estimated array manifold matrix \(A = W\) -1 \(P\Lambda\); define then each column vector of contains the direction information of an independent incident signal. Assume that the column vector corresponds to the direction \(\theta\) k then should have the following properties: where ρ represents the diagonal element in the diagonal matrix Λ corresponding to the column vector ; take the real and imaginary parts of the vector , and perform normalization processing to obtain the vector h k : Extract the input feature vector h of each signal one by one k , and input it into the trained single-signal network to obtain the direction-finding results of each signal in turn; When performing network training, solve for the gradient of L(Ω) with respect to Ω Then update the parameters in the opposite direction of the gradient to minimize the loss function; The specific feature extraction process in step 2 is as follows: 1) Whiten the array received data x(t). First, calculate the covariance matrix R of x(t) x , and then perform orthogonal decomposition on R x : R x = E[x(t)x(t) H = QΛQ H (12) Where: Λ is a diagonal matrix, and Q is an orthogonal matrix; the data is whitened to obtain data v(t) through the following formula: v(t) = Λ -1 / 2 Q H x(t) (13) 2) Assume w q Separate the q-th column vector of the separation matrix W and initialize the vector w q , and set the initial value as w q (0), and calculate the covariance matrix P vv = E[v(t)v(t) H , assume N is the number of snapshots of the array, n = 1, 2,..., N, let i = 1, y(n) = w q (i - 1)v(n), and update the vector w using the following formula q : 3) Assume that the estimated vectors are \(w_1, w_2, \ldots, w\) q-1 , then: 4) If then w is obtained q , otherwise, let i←i + 1, and return to step 3); 5) Let q←q + 1. If q < M, return to step 2) until all column vectors are estimated, thereby obtaining the separation matrix W; 6) Calculate Assume is the k-th column vector, take the real and imaginary parts of, and after normalizing them, use them as the input feature vector h k .

2. The high-precision array direction finding method based on data driving according to claim 1, characterized in that, Step 1.2 also includes a testing process: extracting the input feature vector from the array received data to obtain z, and inputting it into the trained network to output a vector The position corresponding to the peak is the direction finding result. After obtaining the output vector the direction finding result θ is obtained through the following formula: In formula (7): φ, respectively represent the direction values corresponding to the left and right sides of the largest peak in and ξ1, ξ2 are respectively the vector element values at the corresponding positions of φ and 3. A confidence ranking positioning system based on multi - propagation mode recognition for the method according to claim 1, characterized in that, Including: Single-signal intelligent array direction-finding network, used to extract the array covariance matrix features in the case of single-signal incidence and perform normalization processing to form an input feature vector; Convolutional neural network training module, used to learn and understand the mapping relationship between the input feature vector and the direction of arrival (DOA); Multi-signal intelligent array direction-finding network, used to process multi-signal situations, including signal separation, feature extraction, and obtaining the direction-finding results of each signal using the single-signal network model.

4. The system according to claim 3, wherein The single-signal intelligent array direction-finding network includes: Feature extraction unit, used to process the array covariance matrix and extract signal features; Normalization unit, used to normalize the extracted signal features to generate an input feature vector.

5. The system according to claim 3, characterized in that, Also included: Neural network training unit, used to train the convolutional neural network so that the input feature vector can be mapped to the corresponding DOA output.

6. The system according to claim 3, characterized in that The multi-signal intelligent array direction-finding network includes: Signal separation unit, used to separate independent signals from multi-signal data; Signal feature extraction unit, used to extract the features of each independent signal and perform normalization processing.

7. The system according to claim 3, characterized in that Also included: Test processing unit, used to extract the input feature vector from the received array data and input it into the trained network to obtain the direction-finding result; Direction-finding result processing unit, used to determine the direction-finding result according to the network output.

8. The system according to claim 3, characterized in that, The feature extraction process includes: Data whitening processing unit, used to whiten the array received data; Iterative update unit, used to update the separation matrix through an iterative algorithm to optimize the independence of the signal; Separation matrix construction unit, used to estimate and construct the separation matrix, and then obtain the feature vector of the independent signal.

Images