A method for capturing weak signals from navigation satellites based on deep convolutional neural networks
Through the deep convolutional neural network framework of image-related peak classification and related peak positioning, the problems of system reliability and low capture efficiency in navigation satellite weak signal capture methods are solved, and more efficient signal capture performance is achieved.
Patent Information
- Application Number
- CN202110499069.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-05-08
- Publication Date
- 2024-11-08
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing navigation satellite weak signal capture method based on deep convolutional neural networks has poor reliability and low capture efficiency, and has failed to fully utilize the potential of deep convolutional neural networks.
A deep convolutional neural network framework that uses image-related peak classification and related peak positioning is used to determine and optimize network parameters through data training to improve signal capture performance. Specific steps include signal reception, related operations, absolute position detection and area encoding, and optimize the classification and positioning accuracy of related peaks.
It improves the capture performance of navigation satellite weak signals, enhances the reliability and capture efficiency of the system, and outputs the optimal value by optimizing network parameters.
Smart Images

Figure CN113406674B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method for capturing weak signals of navigation satellites based on a deep convolutional neural network. More specifically, the present invention relates to a deep convolutional neural network framework that uses correlation peak classification and correlation peak positioning, which is applied to the capture and detection of weak signals of navigation satellites. By performing data training on the deep convolutional neural network, optimizing network parameters and outputting optimal values, the method improves the performance of capturing weak signals of navigation satellites. Background Art
[0002] Convolutional Neural Networks (CNNs) technology has received extensive attention from industry and academia due to the high accuracy of deep convolutional neural networks in capturing weak navigation satellite signals. In a fading environment, satellite navigation signals are very weak, and the detection probability of the receiver using traditional acquisition algorithms is very low. In satellite navigation systems such as GPS, code division multiple access technology is used, and its spread spectrum signal power on the earth's surface is about -128.5dBm. The receiver acquires the signal by making the correlation peak greater than the noise power through the spread spectrum gain. In fading environments such as urban canyons, vegetation, or indoors, GPS signals are further attenuated. The typical GPS signal attenuation is about 5 to 15dB under vegetation, about 10 to 30dB in urban canyons, and more than 25dB underground, so the acquisition threshold of the available GPS receiver is at least better than -145dBm. Traditional navigation satellite signal acquisition algorithms mainly use additional hardware costs to increase the capture probability. In the literature "Research of Weak GPS Signal Acquisition Algorithm" (S. Tian, and Y. Pi, 2008 International Conference on Communications, Circuits and Systems Communications, Circuits and Systems, 2008. pp. 793-796 May, 2008.) and the literature "A Double Dwell High Sensitivity GPS Acquisition Scheme Using Binarized Convolution Neural Network" (Z. Wang, Y. Zhuang, J. Yang, H. Zhang, J. Yang, H. Zhang, W. Dong, M. Wang, L. Hua, B. Liu, B. Liu, and L. Shi, Sensors 18 1482, 2018), differential correlation algorithm and spatial antenna diversity are used to achieve the capture of weak GPS signals, but it brings more computing, memory or other hardware costs. The current navigation satellite weak signal capture method based on deep convolutional neural network usually involves a deep convolutional neural network framework of image correlation peak classification and correlation peak positioning. The deep convolutional neural network includes a peak classification (Peak Classification, PC) and a peak positioning (Peak Position, PP), which respectively classify the correlation peaks and identify the absolute positions of the relevant images, thereby improving the performance of obtaining weak satellite navigation signals.
[0003] However, among the current methods for capturing weak signals from navigation satellites based on deep convolutional neural networks, there is still a lack of successful precedents for methods that can fully utilize deep convolutional neural networks. Without fully utilizing deep convolutional neural networks to capture signals, the existing methods for capturing weak signals from navigation satellites have poor system reliability and low capture efficiency.
[0004] In the present invention, the inventor proposes a navigation satellite weak signal capture method based on a deep convolutional neural network for capturing weak signals from navigation satellites. In the method of the present invention, a deep convolutional neural network framework for image correlation peak classification and correlation peak positioning is established, which is applied to navigation satellite weak signal detection. By training the deep convolutional neural network with data, the optimized network parameters are determined and the optimal values are output, thereby improving the method for capturing the performance of weak signals from navigation satellites. Summary of the invention
[0005] In order to overcome the shortcomings of the above technical problems, the present invention proposes a method for capturing weak signals of navigation satellites based on a deep convolutional neural network.
[0006] The method for capturing weak navigation satellite signals based on a deep convolutional neural network of the present invention is characterized in that it is implemented by the following steps: Step 1, a navigation satellite signal receiver obtains a weak navigation satellite signal, receives an intermediate frequency signal r(t) and generates a locally generated signal s(t) after being processed by a radio frequency circuit, and obtains a correlation signal R(t) by performing a correlation operation between the intermediate frequency signal r(t) and the locally generated signal s(t); Step 2, an absolute position convolutional neural network is used to detect the correlation value of the signal; Step 3, an area coding method is used to improve the absolute position accuracy of the correlation peak, a training data set is trained, and a PC-CNN is obtained. S and PP-CNN S The optimal model.
[0007] The method for capturing weak signals from navigation satellites based on a deep convolutional neural network of the present invention, the step 1 is implemented by the following sub-steps:
[0008] a) Obtain the intermediate frequency signal r(t); For a navigation satellite signal acquisition system with a deep convolutional neural network, after being processed by the RF circuit, the received intermediate frequency signal is:
[0009] r(t)=AD r (t)C r (t-τ)×cos[2π(f IF +f d )t+φ0]+n(t) (1)
[0010] In formula (1), A is the signal amplitude, D r(t) is the modulated data, f d is the Doppler shift, φ0 is the original phase, C r (t-τ) is a pseudo-random code with a delay time;
[0011] b) The local generated signal generated by the receiver is s(t); for a navigation satellite signal acquisition system with a deep convolutional neural network, the local generated signal generated by the receiver is:
[0012]
[0013] In formula (2), Is a delay time The local pseudo-random code, is the local frequency shift, φ1 is the initial phase;
[0014] c) Obtaining the correlation signal R(t); Correlation processing is performed on the intermediate frequency signal r(t) and the locally generated signal s(t), and the obtained correlation signal is:
[0015]
[0016] In formula (3), Δφ=φ0-φ1,n cor is the equivalent noise including Gaussian noise and cross-correlation output.
[0017] The method for capturing weak signals from navigation satellites based on a deep convolutional neural network of the present invention, the step 2 is implemented by the following sub-steps:
[0018] d) The correlator processes the signal to obtain a correlation image, and inputs the correlation image into the convolutional neural network; M correlators are used to process the signal in parallel to reduce the acquisition time, and the entire correlation image is divided into multiple correlation sub-images, which are input into the convolutional neural network;
[0019] e) PC-CNN S Identify the correlation peaks in the image; absolute position The correlation peak classification in the convolutional neural network has S1 convolutional layers, an A×A kernel, M1 maximum pooling layers, and E1 fully connected layers to identify the correlation peaks in the image. Since only one segmented image has a correlation peak in the entire correlation image, PC-CNN S The high recognition accuracy can reduce the number of segmented images;
[0020] f)PP-CNN S Identify the absolute position of the correlation peak of the segmented image; PP-CNN SThere are N2 maximum pooling layers, S2 convolutional layers and E2 fully connected layers in the CNN. Each layer has a rectified linear unit (ReLU) as a nonlinear activation function. After peak classification, PP-CNN S The segmented images with correlation peaks are used for absolute position identification, and the segmented images without correlation peaks are only used to assist calculation.
[0021] In the method for capturing weak signals from navigation satellites based on a deep convolutional neural network of the present invention, step three is implemented by the following sub-steps:
[0022] g) In order to improve the absolute position accuracy of the correlation peak, the regional coding method is used; the resolution of the absolute position of the pixel decreases with the increase of the image size, and is linearly related to the inverse of the image size. When the two-dimensional coordinate is the training target of the convolutional neural network, the positioning accuracy will be affected as the image size increases. For example, when the image size is X1×X1, according to the normalization result, the coordinate resolution ratio of any two pixels is 1 / X1. For an image of X2×X2, the coordinate resolution ratio of any two pixels is 1 / X2. In the same convolutional neural network, the larger the image size, the lower the coordinate recognition accuracy. The regional coding method is used to divide the input image into several regions with their own X1×X1 coordinates, and the coordinate range of all regions is limited to between X1 and X2;
[0023] h) Obtain PC-CNN S The optimal model of the correlation peak image is divided into multiple correlation sub-images, among which there is only one image with correlation peak and multiple images with noise, which can be used in PC-CNN S China-Israel Accuracy detection of images without correlation peaks, classification accuracy of images with correlation peaks Higher, PC-CNN S The optimal model of
[0024] i) Derive PP-CNN S The optimal model of PP-CNN S The number of convolutional layers in the framework is S2, which can find the best absolute position accuracy. According to the recognition rate and absolute positioning accuracy of different models, PP-CNN is obtained. S The model has the highest position accuracy, confirming that PP-CNN S Model, the absolute positioning accuracy of the entire correlation image is The absolute position accuracy of the final correlation peak is obtained.
[0025] The beneficial effects of the present invention are as follows: the navigation satellite weak signal capture method based on deep convolutional neural network of the present invention establishes a deep convolutional neural network framework for image correlation peak classification and correlation peak positioning, which is applied to navigation satellite weak signal detection, and by performing data training on the deep convolutional neural network, the optimized network parameters are determined and the optimal values are output, thereby improving the performance of capturing navigation satellite weak signals. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] Figure 1 This is a schematic diagram of the convolutional neural network navigation satellite signal acquisition framework system;
[0027] Figure 2 This is a schematic diagram of the framework module structure of the absolute position convolutional neural network;
[0028] Figure 3 This is a schematic diagram of the regional coding coordinates;
[0029] Figure 4 Simulation diagram of the peak classification accuracy of different models in PC-CNNs;
[0030] Figure 5 Simulation diagram of absolute position accuracy of models with different number of layers in PP-CNNs. DETAILED DESCRIPTION
[0031] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.
[0032] Since deep convolutional neural networks are widely used in navigation satellite weak signal capture systems, the present invention proposes a specific implementation method for capturing weak signals from navigation satellites based on deep convolutional neural networks for capturing and transmitting weak signals from navigation satellites. In order to improve the performance of signal capture, the present invention establishes a deep convolutional neural network framework for image correlation peak classification and correlation peak positioning, and by training the deep convolutional neural network with data, determines the optimized network parameters and outputs the optimal values, thereby improving the method for obtaining the performance of weak signals from navigation satellites.
[0033] The present invention proposes a method for capturing weak signals of navigation satellites based on deep convolutional neural networks. By establishing a deep convolutional neural network framework for correlation peak classification and correlation peak absolute position identification, the influence of different numbers of convolutional layers and different training data sets on the detection accuracy of the convolutional neural network is analyzed, the optimized network parameters are determined and the optimal values are output, thereby improving the performance of the convolutional neural network in capturing weak signals of navigation satellites.
[0034] like Figure 1As shown in the figure, it is a schematic diagram of the convolutional neural network navigation satellite signal acquisition framework system. For a deep convolutional neural network navigation satellite signal acquisition system, there are M correlation channels to obtain weak navigation satellite signals with different initial phases. After being processed by the RF circuit, the intermediate frequency signal r(t) is received, and the receiver generates a locally generated signal s(t). By performing correlation operations between the intermediate frequency signal r(t) and the locally generated signal s(t), the correlation signal R(t) is obtained.
[0035] In the method proposed by the present invention, for a navigation satellite weak signal acquisition system based on a deep convolutional neural network, after being processed by a radio frequency circuit, the intermediate frequency signal received is
[0036] r(t)=AD r (t)C r (t-τ)×cos[2π(f IF +f d )t+φ0]+n(t) (1)
[0037] In formula (1), A is the signal amplitude, D r (t) is the modulated data, C r (t-τ) is a pseudo-random code with a delay time, f d is the Doppler frequency shift, and φ0 is the original phase. On the other hand, the receiver generates a local generated signal s(t). For a navigation satellite weak signal acquisition system based on a deep convolutional neural network, after signal acquisition and processing, the receiver generates a local generated signal
[0038]
[0039] In formula (2), Is a delay time The local pseudo-random code, is the local frequency shift, and φ1 is the initial phase. Through the operation, the intermediate frequency signal r(t) and the locally generated signal s(t) are correlated to obtain the correlation signal:
[0040]
[0041] In formula (3), Δφ=φ0-φ1,n cor is the noise equivalent to the output including Gaussian noise and cross-correlation.
[0042] Figure 2 The schematic diagram of the framework module structure of the absolute position convolutional neural network is given. The absolute position convolutional neural network has two deep networks, one is the correlation peak classification and the other is the correlation peak location. PC-CNN SThere are 2 convolutional layers, a 3×3 kernel, 2 max pooling layers, and 2 fully connected layers to identify the correlation peaks in the image. Since only one correlation image has a correlation peak among the entire correlation images, PC-CNN S The high recognition accuracy of PP-CNN can reduce the number of segmented images. S The segmented images with correlation peaks are used to identify the absolute position, and the segmented images without correlation peaks are only used to assist calculation. S There are N max pooling layers, S convolutional layers, and 2 fully connected layers. For absolute position, it is very important to have a zero padding process in each layer. There is a ReLU as a nonlinear activation function behind each layer.
[0043] Figure 3 A schematic diagram of region coding coordinates is given. In order to improve the absolute position accuracy of the correlation peak, the region coding method is used. The resolution of the absolute position of the pixel decreases as the image size increases, and is linearly related to the inverse of the image size. When two-dimensional coordinates are the training target of the convolutional neural network, the positioning accuracy will be affected as the image size increases. For example, when the image size is 8×8, according to the normalized result, the coordinate resolution ratio of any two pixels is 1 / 8=0.125. For a 32×32 image, the coordinate resolution ratio of any two pixels is 1 / 32=0.03125. In the same convolutional neural network, the larger the image size, the lower the coordinate recognition accuracy. The region coding method is used to divide the input image into several regions with their own 8×8 coordinates, and the coordinate range of all regions is limited to between 0 and 7.
[0044] Figure 4 Different models are given in PC-CNN S The correlation peak classification accuracy simulation graph in . The correlation peak image is divided into 832 32×32 pixel images, of which there is only one image with correlation peak and 831 noisy images. S China-Israel Accuracy detection of images without correlation peaks, classification accuracy of images with correlation peaks Higher. The correlation peak classification model can be obtained by changing the number of maximum pooling layers and convolution layers. Model 1 is a correlation peak classification model with 2 maximum pooling layers and 2 convolution layers, Model 2 is a correlation peak classification model with 4 maximum pooling layers and 4 convolution layers, and Model 3 is a correlation peak classification model with 3 maximum pooling layers and 3 convolution layers. When the signal power is weak, the classification accuracy of Model 1 and Model 2 will be significantly reduced, and the classification accuracy of Model 3 is higher than that of Model 1 and Model 2, which confirms that PC-CNN S The optimal model is Model 3.
[0045] Figure 5 Models with different number of layers are given in PP-CNN S The absolute position accuracy simulation diagram in . Analyze the absolute position model with different maximum pooling layers and convolutional layers, and change the PP-CNN S The number of convolutional layers S of the framework is used to obtain four different absolute position models to find the best absolute position accuracy, and the PP-CNN with the highest position accuracy is obtained. S Model. Model 1 is an absolute position model with 4 maximum pooling layers and 4 convolutional layers, Model 2 is an absolute position model with 10 maximum pooling layers and 4 convolutional layers, Model 3 is an absolute position model with 13 maximum pooling layers and 4 convolutional layers, and Model 4 is an absolute position model with 14 maximum pooling layers and 4 convolutional layers. When the signal power is weak, the absolute position accuracy of Model 1, Model 2, and Model 4 will be significantly reduced. The absolute position accuracy of Model 3 is higher than that of Model 1, Model 2, and Model 4. According to the recognition rate and absolute positioning accuracy of different models, Model 3 is determined to be PP-CNN. S The absolute positioning accuracy of the entire correlation image is The final correlation peak absolute position accuracy is obtained.
[0046] In summary, the present invention provides a method for capturing weak signals of navigation satellites based on a deep convolutional neural network, by establishing a deep convolutional neural network framework for image correlation peak classification and correlation peak positioning, by performing data training on the deep convolutional neural network, determining the optimized network parameters and outputting the optimal values, thereby improving the method for capturing weak signals of navigation satellites.
[0047] The above technical solution is only one implementation mode of the present invention. For those skilled in the art, it is easy to make various types of improvements or modifications based on the application methods and principles disclosed in the present invention, and it is not limited to the method described in the above specific implementation mode of the present invention. Therefore, the method described above is only preferred and does not have a restrictive meaning.
Claims
1. A method for capturing weak signals from navigation satellites based on deep convolutional neural networks, characterized in that: This is achieved through the following steps: Step 1: The navigation satellite signal receiver obtains a weak navigation satellite signal, and after being processed by the radio frequency circuit, receives the intermediate frequency signal r(t) and generates a locally generated signal s(t). By performing a correlation operation between the intermediate frequency signal r(t) and the locally generated signal s(t), the correlation signal R(t) is obtained; Step 2: The absolute position convolutional neural network is used to detect the correlation value of the signal; Step 3: The area coding method is used to improve the absolute position accuracy of the correlation peak, and the training data set is used to obtain the PC-CNN. S and PP-CNN S The optimal model of step 1 is achieved by the following steps: a) Obtain the intermediate frequency signal r(t); For a navigation satellite signal acquisition system with a deep convolutional neural network, after being processed by the RF circuit, the received intermediate frequency signal is: r(t)=AD r (t)C r (t-τ)×cos[2π(f IF +f d )t+φ0]+n(t) (1) In formula (1), A is the signal amplitude, D r (t) is the modulated data, f d is the Doppler shift, φ0 is the original phase, C r (t-τ) is a pseudo-random code with a delay time; b) The local generated signal generated by the receiver is s(t); for a navigation satellite signal acquisition system with a deep convolutional neural network, the local generated signal generated by the receiver is: In formula (2), Is a delay time The local pseudo-random code, is the local frequency shift, φ1 is the initial phase; c) Obtaining the correlation signal R(t); Correlation processing is performed on the intermediate frequency signal r(t) and the locally generated signal s(t), and the obtained correlation signal is: In formula (3), Δφ=φ0-φ1,n cor is the equivalent noise including Gaussian noise and cross-correlation output; The step 2 is implemented by the following steps: d) The correlator processes the signal to obtain a correlation image, and inputs the correlation image into the convolutional neural network; M correlators are used to process the signal in parallel to reduce the acquisition time, and the entire correlation image is divided into multiple correlation sub-images, which are input into the convolutional neural network; e) PC-CNN S Identify the correlation peaks in the image; absolute position The correlation peak classification in the convolutional neural network has S1 convolutional layers, an A×A kernel, M1 maximum pooling layers, and E1 fully connected layers to identify the correlation peaks in the image. Since only one segmented image has a correlation peak in the entire correlation image, PC-CNN S The high recognition accuracy can reduce the number of segmented images; f)PP-CNN S Identify the absolute position of the correlation peak of the segmented image; PP-CNN S There are N2 maximum pooling layers, S2 convolutional layers and E2 fully connected layers in the CNN. Each layer has a ReLU as a nonlinear activation function. After peak classification, PP-CNN S The segmented images with correlation peaks are used for absolute position identification, and the segmented images without correlation peaks are only used to assist calculation; The step three is implemented by the following sub-steps: g) In order to improve the absolute position accuracy of the correlation peak, the regional coding method is used; the resolution of the absolute position of the pixel decreases with the increase of the image size, and is linearly related to the inverse of the image size. When the two-dimensional coordinate is the training target of the convolutional neural network, the positioning accuracy will be affected as the image size increases. When the image size is X1×X1, according to the normalization result, the coordinate resolution ratio of any two pixels is 1 / X1. For an X2×X2 image, the coordinate resolution ratio of any two pixels is 1 / X2. In the same convolutional neural network, the larger the image size, the lower the coordinate recognition accuracy. The regional coding method is used to divide the input image into several regions with their own X1×X1 coordinates, and the coordinate range of all regions is limited to between X1 and X2; h) Obtain PC-CNN S The optimal model of the correlation peak image is divided into multiple correlation sub-images, among which there is only one image with correlation peak and multiple images with noise, which can be used in PC-CNN S China-Israel Accuracy detection of images without correlation peaks, classification accuracy of images with correlation peaks Higher, PC-CNN S The optimal model of i) derive PP-CNN S The optimal model of PP-CNN S The number of convolutional layers in the framework is S2, which can find the best absolute position accuracy. According to the recognition rate and absolute positioning accuracy of different models, PP-CNN is obtained. S The model has the highest position accuracy, confirming that PP-CNN S Model, the absolute positioning accuracy of the entire correlation image is The final correlation peak absolute position accuracy is obtained.
