Rotary antenna GNSS multi-path error correction method based on machine learning

Abstract

The invention discloses a rotating antenna GNSS multi-path error correction method based on machine learning, and the method comprises the steps: rotating an antenna at a target point, and collecting a GNSS signal; performing windowing processing on the Doppler frequency data, then performing Fourier transform, and extracting a phase spectrum as a signal feature; classifying the signals by using a support vector machine and a neural network; training the marked direct and non-direct signals by using a neural network, and adjusting weights of the two types of signals in positioning calculation through a back propagation algorithm, so that a calculated positioning value is close to an initial reference positioning value as much as possible, and a positioning error is smaller than a preset threshold value; according to the method, the positioning precision of the GNSS in a multi-path error environment is improved.

Description

Technical Field

[0001] The present invention belongs to the field of positioning applications in GNSS (Global Navigation Satellite System) denied environments, relates to GNSS signal processing and error correction methods, and specifically relates to a rotating antenna GNSS multipath error correction method based on machine learning. Background Art

[0002] GNSS multipath error is one of the main challenges to high-precision positioning in complex environments. In scenarios such as urban canyons and tall buildings, GNSS signals are susceptible to reflection and refraction, resulting in multipath errors. These errors can significantly reduce positioning accuracy and impact the reliability of applications such as autonomous driving and logistics delivery.

[0003] Existing multipath error correction methods primarily focus on reducing errors from non-line-of-sight (LOS) signals, such as narrow correlators and choke antennas. However, these methods typically have limited effectiveness against non-line-of-sight (NLOS) signals and are difficult to address simultaneously. Furthermore, some methods rely on additional sensors (e.g., LiDAR, vision cameras) or complex 3D maps, which increases system complexity and cost.

[0004] By rotating the antenna, the Doppler shift caused by antenna motion can be exploited to detect and correct multipath signals. This approach not only reduces errors in non-line-of-sight (LOS) signals but also effectively identifies and eliminates non-line-of-sight (NLOS) signals, thereby improving positioning accuracy in complex environments. However, existing methods based on antenna motion often require complex mechanical devices or multiple antennas, limiting their widespread adoption in practical applications. Summary of the Invention

[0005] To solve the above problems, the present invention discloses a rotating antenna GNSS multipath error correction method based on machine learning. GNSS signals are collected by a rotating antenna, and support vector machines (SVM) and neural networks are used to classify and weight the signals, thereby improving GNSS positioning accuracy.

[0006] To achieve the above object, the technical solution of the present invention is as follows:

[0007] A method for correcting multipath errors of a rotating antenna GNSS based on machine learning, comprising the following steps:

[0008] Step 1: Rotate the antenna at the target point to collect GNSS signals;

[0009] Step 2: Perform Fourier transform on the Doppler frequency data after windowing, and extract the phase spectrum as the signal feature;

[0010] Step 3: Use support vector machine (SVM) to classify the extracted phase spectrum and signal-to-noise ratio, distinguish direct signals from indirect signals, and determine the classification decision boundary;

[0011] Step 4: Use a neural network to train the labeled direct and indirect signals, and adjust the weights of the two types of signals in the positioning calculation through the back propagation algorithm, so that the calculated positioning value is as close as possible to the initial reference positioning value, and the positioning error is less than the preset threshold.

[0012] The specific steps are:

[0013] Step 1. Rotate the antenna at the target point to collect GNSS signals

[0014] At the target positioning point, the GNSS antenna is rotated horizontally at a preset radius and rotation speed. The rotation radius of the rotating device is 25 cm and the rotation speed is 27 rpm. It is placed in an environment blocked by high-rise buildings to receive GNSS signals.

[0015] Step 2. Window the collected Doppler frequency data and perform Fourier transform to extract the phase spectrum as the signal feature. The details are as follows:

[0016] The collected Doppler frequency data is recorded as Xn, where n is the sampling point index;

[0017] The present invention adopts Hanning window and is calculated as follows:

[0018]

[0019] X wn =Xn·W(n)

[0020] Where W(n) is the value of the window function, n is the index of the sampling point (0≤n≤N-1), and N is the length of the window (the number of sampling points).

[0021] Perform fast Fourier transform (FFT) on the windowed data:

[0022]

[0023] Where, X ik is the complex result after Fourier transform, k is the frequency index (0≤k≤N-1), and j is the imaginary unit.

[0024] Calculate the argument of the Fourier transform result and extract the phase spectrum:

[0025] φ i (k) = arg(X ik )

[0026] Where, φi (k) is the phase spectrum, X ik is the complex result after Fourier transform, arg(X ik ) represents the complex number X ik 's argument.

[0027] Step 3: Use support vector machine (SVM) to classify the extracted phase spectrum and signal-to-noise ratio, distinguish direct signals from indirect signals, and determine the classification decision boundary.

[0028] First, the extracted phase spectrum and signal-to-noise ratio are used as input feature vectors to construct a training sample set. The phase spectrum of the nth sample is denoted as Xn, and the signal-to-noise ratio is denoted as Yn. Each sample corresponds to a GNSS signal and is labeled as either direct or indirect.

[0029] Normalize the training sample set so that the value range of the input feature vector is unified to the interval [0,1], thereby improving the training efficiency and classification performance of the SVM. The normalization formula is as follows:

[0030]

[0031] Where Pn and Qn represent the normalized phase spectrum and signal-to-noise ratio of the nth input eigenvector; Xn and Yn represent the original nth input eigenvector; Xmax, Ymax and Xmin, Ymin represent the maximum and minimum values ​​of the input eigenvector, respectively.

[0032] Selecting a suitable kernel function is the key to SVM classification. In the present invention, Gaussian radial basis function (RBF) is used as the kernel function of SVM, and its expression is:

[0033]

[0034] Where X and Y are two input vectors; ||XY|| represents the Euclidean distance between the two input vectors; σ is the kernel function parameter, which is used to control the width of the kernel function.

[0035] The SVM is trained using the training sample set. The SVM model obtained through training can classify the input phase spectrum and signal-to-noise ratio feature vector as direct signal or indirect signal, and determine the classification decision boundary. The decision function form is:

[0036]

[0037] Where ai is the Lagrange multiplier, each sample corresponds to an ai, n is the total number of samples, yi is the actual value of the predicted indicator, K is the kernel function, and b is the bias.

[0038] In the classification process, for a new GNSS signal, its phase spectrum and signal-to-noise ratio are extracted as input feature vectors, and classified using the trained SVM model to obtain the classification result of whether the signal is a direct signal or an indirect signal.

[0039] Step 4: Use a neural network to train the labeled direct and indirect signals, and adjust the weights of the two types of signals in the positioning calculation through the back propagation algorithm, so that the calculated positioning value is as close as possible to the initial reference positioning value, and the positioning error is less than the preset threshold (3 meters).

[0040] A multi-layer perceptron (MLP) neural network is trained on the labeled direct and indirect signals. Its architecture consists of an input layer, a hidden layer, and an output layer. The input layer receives feature data labeled with direct and indirect signals, the hidden layer contains several neurons, and the output layer outputs weighted values ​​for the two signal types. The network's input features are the phase spectrum and signal-to-noise ratio, and its output is the weighted values ​​for the two signal types, which are used to adjust the contribution of direct and indirect signals in positioning calculations.

[0041] First, the weights and biases of the neural network are initialized. The weights are randomly initialized, and the biases are initialized to 0. The training dataset is then divided into batches based on the acquisition location. A batch of input data is then fed into the network, and the predicted value of the output layer is obtained through forward propagation. The error between the predicted value and the initial reference positioning value is calculated, using the mean squared error (MSE) as the loss function, as follows:

[0042]

[0043] Among them, m is the number of samples in the training data, yi and where are the actual and predicted values ​​of the predicted indicator, respectively. The backpropagation algorithm is then used to calculate the gradient of the loss function with respect to the weights and biases of each layer. The weights and biases are updated using gradient descent, with a learning rate set to 0.001. This process is repeated until the model has run 200 iterations or the training error converges below a specified threshold. This optimizes the weight distribution, improves positioning accuracy, and ensures that the final positioning error is less than the preset threshold (3 meters).

[0044] The beneficial effects of the present invention include:

[0045] The present invention proposes a rotating antenna GNSS multipath error correction method based on machine learning. By rotating the antenna at the target point to collect GNSS signals, the Doppler frequency data phase spectrum and signal-to-noise ratio are extracted as signal features, and then support vector machines (SVM) and neural networks are used to classify and weight the signals, the positioning accuracy of GNSS in a multipath error environment is effectively improved. The problem that the existing correction methods have limited effect on non-line-of-sight and non-direct signal errors is solved. The method can be applied to high-precision positioning scenarios in complex environments such as urban canyons and high-rise building obstructions, providing more reliable positioning technology support for fields such as autonomous driving and logistics distribution, and providing reference significance for other satellite positioning multipath error elimination methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a flow chart of the rotating antenna GNSS multipath error correction method based on machine learning according to the present invention;

[0047] Figure 2 It is a schematic diagram of a rotating antenna device, which realizes the rotation of the GNSS antenna to collect data. DETAILED DESCRIPTION

[0048] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0049] As attached Figure 1 As shown, the present invention discloses a rotating antenna GNSS multipath error correction method based on machine learning, and the specific steps are as follows:

[0050] Step 1. Use the attached Figure 2 The device shown rotates the antenna at the target point to collect GNSS signals

[0051] At the target positioning point, the GNSS antenna is rotated horizontally at a preset radius and speed. The rotation radius of the rotating device described in the present invention is 25 cm and the speed is 27 rpm. It has a simple and compact structure and is easy to use. It can be placed in an environment blocked by high-rise buildings to receive GNSS signals.

[0052] Step 2. Window the collected Doppler frequency data and perform a Fourier transform to extract the phase spectrum as the signal feature. The collected Doppler frequency data is denoted as Xn, where n is the sampling point index.

[0053] The present invention adopts Hanning window and is calculated as follows:

[0054]

[0055] X wn=Xn·W(n)

[0056] Where W(n) is the value of the window function, n is the index of the sampling point (0≤n≤N-1), and N is the length of the window (the number of sampling points).

[0057] Perform fast Fourier transform (FFT) on the windowed data:

[0058]

[0059] Where, X ik is the complex result after Fourier transform, k is the frequency index (0≤k≤N-1), and j is the imaginary unit.

[0060] Calculate the argument of the Fourier transform result and extract the phase spectrum:

[0061] φ i (k) = arg(X ik )

[0062] Where, φ i (k) is the phase spectrum, X ik is the complex result after Fourier transform, arg(X ik ) represents the complex number X ik 's argument.

[0063] Step 3: Use support vector machine (SVM) to classify the extracted phase spectrum and signal-to-noise ratio, distinguish direct signals from indirect signals, and determine the classification decision boundary.

[0064] First, the extracted phase spectrum and signal-to-noise ratio are used as input feature vectors to construct a training sample set. The phase spectrum of the nth sample is denoted as Xn, and the signal-to-noise ratio is denoted as Yn. Each sample corresponds to a GNSS signal and is labeled as either direct or indirect.

[0065] Normalize the training sample set so that the value range of the input feature vector is unified to the interval [0,1], thereby improving the training efficiency and classification performance of the SVM. The normalization formula is as follows:

[0066]

[0067] Where Pn and Qn represent the normalized phase spectrum and signal-to-noise ratio of the nth input eigenvector; Xn and Yn represent the original nth input eigenvector; Xmax, Ymax and Xmin, Ymin represent the maximum and minimum values ​​of the input eigenvector, respectively.

[0068] Selecting a suitable kernel function is the key to SVM classification. In the present invention, Gaussian radial basis function (RBF) is used as the kernel function of SVM, and its expression is:

[0069]

[0070] Where X and Y are two input vectors; ||XY|| represents the Euclidean distance between the two input vectors; σ is the kernel function parameter, which is used to control the width of the kernel function.

[0071] The SVM is trained using the training sample set. The SVM model obtained through training can classify the input phase spectrum and signal-to-noise ratio feature vector as direct signal or indirect signal, and determine the classification decision boundary. The decision function form is:

[0072]

[0073] Where ai is the Lagrange multiplier, and each sample corresponds to an a i , n is the total number of samples, yi is the actual value of the predicted indicator, K is the kernel function, and b is the bias.

[0074] In the classification process, for a new GNSS signal, its phase spectrum and signal-to-noise ratio are extracted as input feature vectors, and classified using the trained SVM model to obtain the classification result of whether the signal is a direct signal or an indirect signal.

[0075] Step 4: Use a neural network to train the labeled direct and indirect signals, and adjust the weights of the two types of signals in the positioning calculation through the back propagation algorithm, so that the calculated positioning value is as close as possible to the initial reference positioning value, and the positioning error is less than the preset threshold (3 meters).

[0076] A multi-layer perceptron (MLP) neural network is trained on the labeled direct and indirect signals. Its architecture consists of an input layer, a hidden layer, and an output layer. The input layer receives feature data labeled with direct and indirect signals, the hidden layer contains several neurons, and the output layer outputs weighted values ​​for the two signal types. The network's input features are the phase spectrum and signal-to-noise ratio, and its output is the weighted values ​​for the two signal types, which are used to adjust the contribution of direct and indirect signals in positioning calculations.

[0077] First, the weights and biases of the neural network are initialized. The weights are randomly initialized, and the biases are initialized to 0. The training dataset is then divided into batches based on the acquisition location. A batch of input data is then fed into the network, and the predicted value of the output layer is obtained through forward propagation. The error between the predicted value and the initial reference positioning value is calculated, using the mean squared error (MSE) as the loss function, as follows:

[0078]

[0079] Among them, m is the number of samples in the training data, y i and where are the actual and predicted values ​​of the predicted indicator, respectively. The backpropagation algorithm is then used to calculate the gradient of the loss function with respect to the weights and biases of each layer. The weights and biases are updated using gradient descent, with a learning rate set to 0.001. This process is repeated until the model has run 200 iterations or the training error converges below a specified threshold. This optimizes the weight distribution, improves positioning accuracy, and ensures that the final positioning error is less than the preset threshold (3 meters).

[0080] It should be noted that the above content merely illustrates the technical idea of ​​the present invention and cannot be used to limit the scope of protection of the present invention. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications all fall within the scope of protection of the claims of the present invention.

Claims

1. A rotating antenna GNSS multipath error correction method based on machine learning, characterized by: The following steps are involved: Step 1: Rotate the antenna at the target point to collect GNSS signals; Step 2: Perform Fourier transform on the Doppler frequency data after windowing, and extract the phase spectrum as the signal feature; Step 3: Use support vector machine to classify the extracted phase spectrum and signal-to-noise ratio, distinguish direct signals from indirect signals, and determine the classification decision boundary; Step 4: Use a neural network to train the labeled direct and indirect signals, and adjust the weights of the two types of signals in the positioning calculation through the back propagation algorithm, so that the calculated positioning value is as close as possible to the initial reference positioning value, and the positioning error is less than the preset threshold.

2. The method for correcting multipath errors of a rotating antenna GNSS based on machine learning according to claim 1, wherein: Step 1 is as follows: At the target positioning point, the GNSS antenna is rotated horizontally at a preset radius and rotation speed. The rotation radius of the rotating device is 25 cm and the rotation speed is 27 rpm. It is placed in an environment blocked by high-rise buildings to receive GNSS signals.

3. The method for correcting multipath errors of a rotating antenna GNSS based on machine learning according to claim 1, wherein: Step 2 is as follows: The collected Doppler frequency data is recorded as Xn, where n is the sampling point index; Use the Hanning window and calculate as follows: X wn =Xn·W(n) Where W(n) is the value of the window function, n is the index of the sampling point (0≤n≤N-1), and N is the length of the window; Perform fast Fourier transform on the windowed data: Where, X ik is the complex result after Fourier transform, k is the frequency index (0≤k≤N-1), and j is the imaginary unit; Calculate the argument of the Fourier transform result and extract the phase spectrum: φ i (k)=arg(X ik ) Where, φ i (k) is the phase spectrum, X ik is the complex result after Fourier transform, arg(X ik ) represents the complex number X ik 's argument.

4. The method for correcting multipath errors of a rotating antenna GNSS based on machine learning according to claim 1, wherein: Step 3 is as follows: First, the extracted phase spectrum and signal-to-noise ratio are used as input feature vectors to construct a training sample set. The phase spectrum of the nth sample is denoted as Xn, and the signal-to-noise ratio is denoted as Yn. Each sample corresponds to a GNSS signal and is marked as a direct signal or an indirect signal. Normalize the training sample set so that the value range of the input feature vector is unified to the interval [0,1] to improve the training efficiency and classification performance of SVM. The normalization formula is as follows: Where Pn and Qn represent the normalized phase spectrum and signal-to-noise ratio of the nth input eigenvector; Xn and Yn represent the original nth input eigenvector; Xmax, Ymax and Xmin, Ymin represent the maximum and minimum values ​​of the input eigenvector, respectively; The Gaussian radial basis function is used as the kernel function of SVM, and its expression is: Where X and Y are two input vectors; ||XY|| represents the Euclidean distance between the two input vectors; σ is the kernel function parameter, which is used to control the width of the kernel function; The SVM is trained using the training sample set. The SVM model obtained through training classifies the input phase spectrum and signal-to-noise ratio feature vectors into direct signals or indirect signals, and determines the classification decision boundary. The decision function form is: Where ai is the Lagrange multiplier, each sample corresponds to an ai, n is the total number of samples, yi is the actual value of the predicted indicator, K is the kernel function, and b is the bias; In the classification process, for a new GNSS signal, its phase spectrum and signal-to-noise ratio are extracted as input feature vectors, and classified using the trained SVM model to obtain the classification result of whether the signal is a direct signal or an indirect signal.

5. The method for correcting multipath errors of a rotating antenna GNSS based on machine learning according to claim 1, wherein: Step 4 is as follows: A multi-layer perceptron neural network is used to train the labeled direct and indirect signals. Its structure includes an input layer, a hidden layer, and an output layer. The input layer receives feature data labeled with direct and indirect signals, the hidden layer contains several neurons, and the output layer outputs weighted values ​​for the two types of signals. The network's input features are phase spectrum and signal-to-noise ratio, and its output is the weighted values ​​of the two types of signals, which are used to adjust the contribution of direct and indirect signals in positioning calculations. First, the weights and biases of the neural network are initialized. The weights are randomly initialized, and the biases are initialized to 0. Then, the training data set is divided into several batches according to the collection location. Then, a batch of input data is input into the network, and the predicted value of the output layer is obtained through forward propagation calculation. Calculate the error between the predicted value and the initial reference positioning value, using the mean square error as the loss function, the formula is: Among them, m is the number of samples in the training data, yi and are the actual value and predicted value of the prediction indicator respectively; then the gradient of the loss function with respect to the weights and biases of each layer is calculated through the back propagation algorithm, and the weights and biases are updated using the gradient descent method, with the learning rate set to 0.001; the above process is repeated until the model as a whole runs 200 iterations or the training error converges below the specified threshold, thereby optimizing the weight distribution, improving the positioning accuracy, and ensuring that the final positioning error is less than the preset threshold.

Images