ADS-B signal DOA estimation method based on deep learning

Abstract

The invention belongs to the field of ADS-B (Automatic Dependent Surveillance-Broadcast) signal DOA (Direction of Arrival) estimation, and particularly relates to an ADS-B signal DOA estimation method based on deep learning. According to the method, an ADS-B signal DOA estimation network based on CNN-LSTM is established, a covariance matrix sequence of ADS-B array signals with continuous timestamps is used as input, spatial features and time features of the signal covariance sequence are captured through CNN and LSTM respectively, time dependence of the covariance matrix sequence of the ADS-B signals with the continuous timestamps is introduced to improve DOA estimation precision, and the DOA estimation accuracy is improved. And mapping from the signal covariance matrix sequence to the signal arrival angle sequence is realized. According to the method, the problem that the performance of an ADS-B signal DOA estimation algorithm only using the CNN is deteriorated under the low signal-to-noise condition can be effectively solved.

Description

Technical Field

[0001] The present invention belongs to the technical fields of Automatic Dependent Surveillance - Broadcast (ADS - B) and Direction of Arrival (DOA) estimation, and particularly relates to a method for estimating the DOA of ADS - B signals based on deep learning. Background Art

[0002] As a new type of air traffic control surveillance system, ADS - B contains rich surveillance information and has a low deployment cost, and has been widely used worldwide. However, due to its open - broadcast, unencrypted, unauthenticated, and economical design concept, ADS - B is vulnerable to spoofing interference. The interference source intercepts the ADS - B signal, tampers with the position information in the message, or directly constructs the ADS - B signal of a "ghost aircraft" to interfere with the ground station's judgment of the current airspace situation, thus causing chaos in air traffic order. Since the message position of the spoofed ADS - B signal is different from the position of the interference source, an array antenna can be used to receive the ADS - B signal, and whether there is consistency between the arrival angle of the ADS - B signal and the azimuth angle of the ADS - B message position can be compared to identify spoofing interference. The detection performance of this anti - spoofing interference technology depends on the angle measurement accuracy and has high requirements for real - time performance.

[0003] The DOA estimation algorithm determines the azimuth angle of the incoming signal arriving at the reference array element through an array antenna. Traditional DOA estimation algorithms include beamforming algorithms, subspace algorithms, and maximum likelihood algorithms, etc. These algorithms all rely on strict mathematical models, perform poorly in low - signal - to - noise ratio environments, and have problems such as large computational complexity and low computational efficiency. With the continuous improvement of deep learning theory and technology, the DOA estimation technology based on deep learning has gradually become a new research direction for DOA estimation, effectively solving the problems existing in traditional DOA estimation algorithms. Considering from the classification idea, the DOA estimation problem can be converted into a multi - label single - classification or multi - label multi - classification task. Taking the covariance matrix of a single signal or the complete signal as the input, features are extracted through a Convolutional Neural Network (CNN) or a Feedforward Neural Network (FNN) to achieve the mapping to the signal angle. Considering from the regression idea, a neural network can be directly used to map the signal features to the angle values of a single or multiple targets. This method requires knowing the number of targets in advance.

[0004] However, the ADS-B signal is broadcast periodically by the aircraft, and the movement of the aircraft is continuous. Therefore, the arrival angle of the continuously received ADS-B signal also changes continuously and has time dependence. The existing DOA estimation techniques based on deep learning do not consider the time dependence between signals with continuous timestamps, and the DOA estimation accuracy needs to be improved. Summary of the Invention

[0005] In view of the above problems or deficiencies, the present invention proposes a DOA estimation method for ADS-B signals based on CNN-LSTM. CNN is used to extract the spatial features of the covariance matrix of the ADS-B signal. On this basis, the Long Short-Term Memory (LSTM) network is used to learn the time features of the covariance matrix of the ADS-B signal with continuous timestamps. By capturing the time dependence of the covariance matrix of the ADS-B signal with continuous timestamps, the DOA estimation accuracy is improved, and the problem of performance degradation of the DOA estimation algorithm for ADS-B signals using only CNN in low signal-to-noise ratio situations is overcome.

[0006] The technical solution adopted by the present invention is as follows:

[0007] S1. Construct a covariance matrix sequence of the ADS-B array received signals as the training set (R xx , α) i,k , where the covariance matrix of the i-th ADS-B signal of the k-th track received by the array antenna is R xx , and the incident angle of the signal is α. Assume that an ADS-B signal is received using a uniform linear array with M array elements, and the covariance matrix of the array signal X(n) is calculated as:

[0008]

[0009] N s is the number of snapshots. Since the covariance matrix of the array received signal is a complex matrix and the neural network cannot directly process complex numbers, the covariance matrix R xx is divided into two channels of real and imaginary parts as inputs. In addition, spectral norm normalization is used to normalize the real and imaginary parts of the covariance matrix respectively. Spectral norm normalization constrains the largest singular value of the covariance matrix, which helps the stability of eigenvalue decomposition. The covariance matrix after spectral norm normalization is:

[0010]

[0011] ‖R xx ‖2 = max(σ(R xx ))

[0012] σ(R xx) is the singular value of the covariance matrix. As Figure 1 shown, assuming that the array antenna receives N consecutive ADS-B signals from the same aircraft, the input of the DOA estimation network is a matrix of N×2×M×M dimensions, and the arrival angles of N ADS-B signals can be obtained.

[0013] S2. Build the DOA estimation network, as Figure 2 shown:

[0014] The DOA estimation problem of ADS-B signals based on deep learning can be regarded as a single-label multi-classification task. Assume that the angle measurement range of the array antenna is [φ min , φ max , and the angle measurement range is equally divided at an interval of △φ to obtain a set containing N φ discrete angles. Each discrete angle is numbered [0, N φ -1], then the problem is transformed into a mapping problem from the covariance matrix sequence of the signals received by the array antenna to the discrete angle number sequence.

[0015] The DOA estimation network of ADS-B signals based on CNN-LSTM mainly consists of two parts. The first part contains three convolutional blocks, and each convolutional block consists of a convolutional layer, a ReLU activation function, and a max-pooling layer. The convolutional layer is used to extract the spatial feature information in the covariance matrix, the ReLU activation function is used to introduce non-linearity so that the network can learn complex mapping relationships, and the max-pooling layer can reduce the model complexity while retaining important features.

[0016] The second part consists of an LSTM network, which is used to capture the temporal feature information between the covariance matrices at different times, and can alleviate the problem of gradient disappearance or gradient explosion existing in typical recurrent neural networks. LSTM introduces a gating system to control the flow of information, controls the information input of the unit through the input gate, controls the retention of unit information through the forget gate, and controls the information output of the unit through the output gate.

[0017] The fully connected layer before the input of the LSTM network adjusts the output dimension of the convolutional block to be consistent with the input dimension of the LSTM network. The fully connected layer after the output of the LSTM network expands the spatial and temporal feature dimensions learned previously to the dimension of the number of discrete angles, and obtains the weight vector corresponding to the signal covariance matrix at each discrete angle.

[0018] Assume that each element of the weight matrix output by the DOA estimation network is y i,j , representing the weight of the weight vector corresponding to the i-th signal covariance matrix sample at the j-th discrete angle. Using the Softmax function, the probability distribution of each signal arrival angle at each discrete angle can be obtained:

[0019]

[0020] p i,j represents the probability that the angle of arrival corresponding to the i-th signal covariance matrix sample is the j-th discrete angle. The cross-entropy loss function is selected as the loss function for the ADS-B signal DOA estimation problem. This loss function is commonly used in classification tasks and is used to measure the difference between the probability distribution predicted by the model and the true label distribution. The cross-entropy loss is defined as follows:

[0021]

[0022] S is the number of signal covariance matrix samples, and e i,j is the label of the angle of arrival corresponding to the i-th signal covariance matrix sample at the j-th discrete angle. When using the network for DOA estimation, find the subscript of the maximum value in the output weight vector, and the discrete angle corresponding to this subscript is the angle of arrival of the incoming wave signal.

[0023] S3. Use the training dataset to train the DOA estimation network for ADS-B signals based on CNN-LSTM, and perform K-fold cross-validation to obtain the best model.

[0024] S4. Use the trained DOA estimation network for ADS-B signals based on CNN-LSTM to estimate the DOA of aircraft ADS-B signals.

[0025] The beneficial effects of the present invention are as follows:

[0026] (1) Use the covariance of ADS-B array signals with multiple consecutive timestamps as the input of the DOA estimation network, introducing time features; use the spectral norm as the normalization method, which helps to stabilize the stability of the eigenvalue decomposition of the signal covariance matrix under different signal-to-noise ratio conditions.

[0027] (2) Use LSTM on the basis of CNN to capture the time dependence of the covariance matrix sequence of ADS-B array signals, overcoming the problem of performance deterioration of the ADS-B signal DOA estimation algorithm that only uses CNN under low signal-to-noise ratio conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is the input schematic diagram of the DOA estimation network for ADS-B signals based on CNN-LSTM;

[0029] Figure 2 is the structure of the DOA estimation network for ADS-B signals based on CNN-LSTM;

[0030] Figure 3 is the scatter plot of the true angle and the estimated angle of the method of the present invention under different signal-to-noise ratio conditions;

[0031] Figure 4 This is a comparison of the angle measurement accuracy between the method of the present invention and the ADS-B signal DOA estimation algorithm using only CNN under different signal-to-noise ratio conditions. Specific implementation mode

[0032] The technical solution of the present invention will be described in detail below in conjunction with the drawings and embodiments.

[0033] Step 1: Generate a dataset for the ADS-B signal DOA estimation network based on CNN-LSTM through a simulated ADS-B array antenna system. The ADS-B array antenna system includes the modeling of the aircraft flight trajectory, the modeling of the ADS-B baseband signal, and the modeling of the array antenna system.

[0034] Establish a northeast celestial coordinate system with the array antenna as the origin, and simulate the flight track of the aircraft. Convert the simulated northeast celestial coordinates of the aircraft into longitude, latitude, and altitude. After simple position report coding, use pulse position modulation to obtain an 112 μs ADS-B data segment signal, and add an 8 μs leading pulse before the data segment signal to obtain a complete 120 μs ADS-B baseband signal.

[0035] Set the array antenna as an 8-element uniform linear array, the element spacing is half of the wavelength of the ADS-B signal, and the sampling frequency is 10 MHz. Simulate the flight trajectories of 20 aircraft through the ADS-B array antenna system, and each trajectory is within the angle measurement range of the array antenna. Each trajectory contains 500 trajectory sampling points, so 500 ADS-B signals can be obtained for each trajectory. Solve the covariance matrix for the 500 ADS-B signals, and use spectral normalization to process the real and imaginary parts of the covariance matrix respectively, and the input of the DOA estimation network is a sequence of 500×2×8×8 continuous timestamp ADS-B signal covariance matrices. After passing through the DOA estimation network, the arrival angles of 500 ADS-B signals can be obtained.

[0036] To enable the model to perform DOA estimation under different signal-to-noise ratios, simulate the above-mentioned ADS-B array received signals for each signal-to-noise ratio in the interval [-10 dB, 20 dB]. The angle measurement range of the array antenna system is [45°, 135°], and the angle measurement interval is divided with an interval of 0.1° to obtain 901 angle numbers. The signal covariance matrix of each ADS-B array received signal and the corresponding angle number form a sample, and a total of 310,000 samples can be obtained.

[0037] Step 2: According to Figure 2 Build an ADS-B signal DOA estimation network based on CNN-LSTM. The DOA estimation network consists of two parts: CNN and LSTM.

[0038] The CNN part consists of three convolutional blocks in sequence. Each convolutional block contains a convolutional layer, an activation function, and a max pooling layer. For the first convolutional block, the number of channels in the convolutional layer is 32, the size of the convolutional kernel is 3×3, the padding is 1, the size of the pooling matrix in the max pooling layer is 3, the padding is 1, and the stride is 1. This layer maintains the original size of the input matrix and preserves the original features of the signal as much as possible. For the second convolutional block, the number of channels in the convolutional layer is 32, the size of the convolutional kernel is 3×3, the padding is 1; the size of the pooling matrix in the max pooling layer is 3, the padding is 0, and the stride is 1. This layer reduces the model complexity while preserving important features as much as possible. For the third convolutional block, the number of channels in the convolutional layer is 64, the size of the convolutional kernel is 3×3, the padding is 1; the size of the pooling matrix in the max pooling layer is 3, the padding is 0, and the stride is 1. This layer further extracts the matrix features while reducing the model complexity.

[0039] LSTM part The dimension of the hidden layer of the LSTM is 256 and the number of layers is 2. The linear layer before the LSTM adjusts the dimension of the output of the CNN part from 1024 to 128, which is the input dimension of the LSTM. Given that the angular measurement range is [45°, 135°] and the angular resolution is 0.1°, it can be known that the output dimension of the network is 901. After the LSTM layer, a linear layer is used to map the output dimension from 256 to 901 to obtain the weight vector corresponding to the signal covariance of the continuous time stamp. Each output weight vector passes through the Softmax activation function to obtain a probability vector. The subscript corresponding to the maximum value of the obtained probability vector is found, and the angle value corresponding to this subscript is the DOA estimation angle of the incoming wave signal.

[0040] Step 3: Use the 5-fold cross-validation method to divide the dataset to train the model, and select the model with the best comprehensive performance of the angular measurement error under different signal-to-noise ratio conditions as the best model.

[0041] Step 4: Use the trained best DOA estimation network for ADS-B signals to estimate the DOA of aircraft ADS-B signals.

[0042] For the method of the present invention, the experimental results obtained by the above implementation are as follows:

[0043] Figure 3 It shows the angular measurement performance of the method of the present invention under different signal-to-noise ratios. Randomly select a flight track from the training set, and use the DOA estimation network to measure the angle of the ADS-B signal emitted by the aircraft under the conditions of signal-to-noise ratios of -10, 0, 10, and 20 dB. Taking the true angle of the signal as the horizontal axis and the estimated angle as the vertical axis, we get as Figure 3The scatter plot shown. Under different signal-to-noise ratios, the distribution of the scatter points is close to the diagonal, indicating that the DOA estimation network has good DOA estimation accuracy and stability under different signal-to-noise ratio conditions. Moreover, as the signal-to-noise ratio increases, the distribution of the scatter points gets closer and closer to the diagonal, indicating that the angle measurement error becomes smaller and smaller.

[0044] Figure 4 It is a comparison of the DOA estimation accuracy between the CNN-based ADS-B signal DOA estimation algorithm and the method of the present invention under different signal-to-noise ratios. It can be seen that the CNN-based ADS-B signal DOA estimation algorithm has a large angle measurement error under low signal-to-noise ratio conditions, and its performance is close to that of the method of the present invention under high signal-to-noise ratio conditions. Under high signal-to-noise ratio conditions, the signal characteristics are more obvious, and the mapping from the signal covariance matrix to the angle can be completed only by relying on CNN. However, under low signal-to-noise ratio conditions, the signal characteristics are submerged by noise, and it is impossible to complete the high-precision DOA estimation task only by relying on CNN. Since the aircraft is moving during flight, the arrival angle of the emitted ADS-B signal has dynamic changes in continuous timestamps. Only using CNN to process the ADS-B signal will ignore the temporal correlation of the ADS-B signals in continuous timestamps. After introducing LSTM on the basis of CNN, the DOA estimation network can capture the temporal dependence between the covariance matrices of the ADS-B signals in continuous timestamps, further reducing the deviation of the estimated angle from the true angle under low signal-to-noise ratio conditions, thereby improving the angle measurement accuracy of the model under low signal-to-noise ratio conditions and effectively suppressing the problem of performance deterioration of the CNN-based ADS-B signal DOA estimation algorithm under low signal-to-noise ratio conditions.

Claims

1. A DOA estimation method for ADS-B signals based on deep learning, characterized in that, Including the following steps: S1. Define N consecutive ADS - B signals received by the array antenna from the same aircraft, and construct a covariance matrix sequence of the ADS - B array received signals as the training set (R xx , α) i,k , indicating that the covariance matrix of the i - th ADS - B signal of the k - th track received by the array antenna is R xx , and the angle of arrival of the signal is α; divide the covariance matrix R xx into real and imaginary parts as two inputs, and use spectral norm normalization to normalize the real and imaginary parts of the covariance matrix respectively. If the array antenna has M array elements, the obtained training data is a matrix of N×2×M×M dimensions; S2. Build a DOA estimation network and define the angle measurement range of the array antenna as [φ min , φ max . Divide the angle measurement range equally with a spacing of △φ to obtain a set containing N φ discrete angles, numbered sequentially as [0, N φ - 1]. Then, the DOA estimation problem is transformed into a mapping problem from the covariance matrix sequence of the signals received by the array antenna to the discrete angle number sequence. The built DOA estimation network consists of two parts. The first part contains three convolutional blocks, and each convolutional block consists of a convolutional layer, a ReLU activation function, and a max-pooling layer. The second part consists of an LSTM network. The output of the first part adjusts the dimension to be consistent with the input dimension of the LSTM network through a fully connected layer. The output of the LSTM network expands the number of spatial features and temporal features learned previously to the number of discrete angles through a fully connected layer to obtain the weight vector corresponding to the signal covariance matrix at each discrete angle. Define each element of the weight matrix output by the DOA estimation network as y i,j , which represents the weight of the weight vector corresponding to the i-th signal covariance matrix sample at the j-th discrete angle. The Softmax function is used to obtain the probability distribution of each signal arrival angle at each discrete angle: where p i,j represents the probability that the arrival angle corresponding to the i-th signal covariance matrix sample is the j-th discrete angle; The cross-entropy loss function is selected as the loss function for the ADS-B signal DOA estimation problem and is defined as follows: where S is the number of samples of the signal covariance matrix, and e i,j is the label of the arrival angle corresponding to the i-th sample of the signal covariance matrix at the j-th discrete angle; when using the network for DOA estimation, find the subscript of the maximum value in the output weight vector, and the discrete angle corresponding to this subscript is the arrival angle of the incoming wave signal; S3. Use the training dataset to train the DOA estimation network and perform K-fold cross-validation to obtain the best model; S4. Use the trained DOA estimation network to estimate the DOA of the ADS-B signal.

Images