A GPS positioning method based on CNN-BP neural network
By combining the CNN-BP model of convolutional neural network and fully connected neural network in GPS signal processing, the multipath effect is solved, and the problem of low GPS positioning accuracy in urban environments is achieved, and higher positioning accuracy and lower Cost, suitable for GPS mobile phone positioning.
Patent Information
- Application Number
- CN202210545295.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-19
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2042-05-19
AI Technical Summary
In an environment with dense urban high-rise buildings, GPS signal receivers are susceptible to multipath effect interference, resulting in reduced positioning accuracy, and the existing suppression methods are complex and the effect is not ideal.
Using a combination of convolutional neural network (CNN) and fully connected neural network (BP), a CNN-BP neural network prediction model is constructed to process multipath signals in GPS signals to reduce the impact of multipath effects. The model includes the input layer, the convolutional layer, the pooling layer, the BP neural network layer and the output layer. The network parameters are trained through the error backpropagation algorithm to improve the GPS positioning accuracy.
It significantly improves the positioning accuracy of GPS in urban environments, reduces cost and complexity, and makes this method suitable for the positioning mode of GPS mobile phones, with the prediction errors of longitude, latitude and altitude reduced by 4.91%, 22.2%, and 19.7%, respectively. .
Smart Images

Figure CN115097496B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to GPS (Global Positioning Navigation System) positioning and navigation technology, and mainly utilizes a neural network system to suppress the multipath effect generated in the process of GPS signal reception. Background Art
[0002] In recent years, with the development of transportation, positioning and navigation technology has emerged. It can guide us to the corresponding destination and provide real-time attitude information, speed information, location information and other navigation information during navigation. At present, the most widely used is global positioning and navigation technology (GPS). In an open environment without any obstruction, the GPS signal will directly reach the receiver and obtain relatively accurate positioning information. In the urban canyon environment with severe obstruction, due to the tall buildings on both sides of the road, the narrow streets are prone to multipath effects, which seriously interfere with the positioning accuracy of the receiver. Therefore, how to suppress the influence of multipath effects in GPS positioning and improve positioning accuracy is an extremely important point in the positioning process.
[0003] At present, there are three main methods for suppressing multipath effects at home and abroad: antenna enhancement technology, improved receiver technology and signal post-processing technology. Among them, antenna enhancement technology mainly suppresses the impact of multipath effects by adopting special types of antennas, antenna arrays combined with signal processing, or careful antenna site selection and installation. Improved receiver technology is to weaken the impact of multipath effects through the internal signal processing method of the receiver. Currently, there are mainly narrow correlation technology and multipath reduction technology. However, these two methods for suppressing multipath effects have high hardware requirements for the receiver, the amount of calculation is relatively large, and the space occupied is relatively large, so they do not have good application scenarios. The current mainstream multipath effect suppression technology is signal post-processing technology, which is a method of suppressing multipath effects by means of signal processing of the measurement data in the receiver. The main methods are based on algorithms such as signal-to-noise ratio, adaptive filtering, and wavelet transform. These algorithms are used to weaken the multipath in the pseudorange residual and pseudorange rate residual to suppress the multipath effect. However, these algorithms are relatively complex and the effects are not ideal.
[0004] With the development of computer technology, many scholars have studied the use of neural networks to deal with the impact of multipath effects on positioning accuracy, and have achieved some good results. On the basis of neural networks, the present invention proposes a method based on the combination of convolutional neural networks (CNN) and fully connected neural networks (BP) (hereinafter collectively referred to as CNN-BP) to deal with the impact of multipath effects in the GPS positioning process.
[0005] According to the search by patent applicants, the domestic invention patents for GPS multipath effect suppression methods include Shen Feng and Li Pingmin, who proposed a method for suppressing multipath effects in their patent "A GPS anti-multipath method based on a robust beamforming algorithm" (application number CN201310138055.4)[1]. Luo Ruidan and Xu Ying proposed using the sparse characteristics of multipath signals on the time axis to estimate multipath parameters based on a compressed sensing algorithm in their patent "A GPS multipath effect correction positioning method and system" (application number CN201410662660.6)[3], which calculates the position correction number based on the estimated distance difference between the reflected signal and the direct signal, and then corrects the position of the receiver, reducing the multipath effect and improving the positioning accuracy.
[0006] References:
[0007] [1] Shen Feng, Li Pingmin, Patent Name: A GPS Anti-Multipath Method Based on Robust Beamforming Algorithm, Application Number: CN201310138055.4
[0008] [2] Luo Ruidan, Xu Ying, Patent Name: A Multipath Parameter Estimation Method Based on Compressed Sensing Algorithm, Application Number: CN201811422470.1
[0009] [3] Chen Wu, Patent Name: A GPS multipath correction positioning method and system, Application Number: CN201410662660.6
[0010] [4] Li Yandong, Hao Zongbo, Lei Hang. A review of convolutional neural network research[J]. Journal of Computer Applications, 2016, 36(9): 2508-2515
[0011] [5] Li Tao, Yuan Gannan. Error analysis of GPS / INS deep integrated navigation based on BP neural network [C] / / . Electronic Proceedings of the 3rd Annual Conference on Satellite Navigation in China - S09 Integrated Navigation and New Navigation Methods. 2012: 124-128.
[0012] [6] Zhang Xiaoting. Research and optimization of multipath suppression technology for GPS system[D]. Shanghai Jiaotong University, 2016. Summary of the invention
[0013] The present invention provides a GPS positioning method based on CNN-BP neural network, which processes multipath signals by using the CNN-BP prediction model in the neural network, thereby suppressing the influence of multipath effect. The technical solution is as follows:
[0014] A GPS positioning method based on CNN-BP neural network includes the following steps:
[0015] The first step is to construct a GPS data set and perform data preprocessing including GPS position, speed, and satellite information;
[0016] The second step is to build a CNN-BP neural network prediction model: The CNN-BP neural network prediction model consists of five parts, namely input layer, convolution layer, pooling layer, BP neural network layer, output layer,
[0017] The input layer includes GPS position, speed and satellite information to form a 3×3 input matrix. The GPS position includes GPS longitude, latitude and altitude information. The speed includes acceleration, angular velocity and speed information. The satellite information includes the number of satellites, accuracy and satellite time information.
[0018] Convolution layer, used to extract and optimize input feature values. The convolution module consists of 2 1×1 convolution kernels, 2 2×2 convolution kernels, and 2 3×3 convolution kernels. For a 3×3 input matrix, the 1×1 convolution kernel connects each input data; the 2×2 convolution kernel partially connects the 2×2 part of the input data matrix; and the 3×3 convolution kernel fully connects the input data matrix.
[0019] After the input matrix is processed by the convolution layer, it enters the pooling layer for feature dimensionality reduction. The pooling module consists of 2 (2×2) pooling kernels and 2 (3×3) pooling kernels. After the convolution layer processing, 2 3×3 feature matrices, 2 2×2 feature matrices and 2 eigenvalues are obtained. For the 4 feature matrices, 2 2×2 pooling operations and 2 3×3 pooling operations are performed, and finally 6 optimized eigenvalues are obtained.
[0020] The 6 optimized feature values obtained by the pooling layer are used as input to the fully connected BP neural network layer for feature fitting; then the GPS prediction results are output through the output layer, including the longitude, latitude and altitude of GPS;
[0021] The third step is model training: input the GPS information in the prepared training set into the CNN-BP neural network prediction model. The convolution layer and the pooling layer are trained using the error back propagation algorithm. The BP neural network part is fully connected and needs to be adjusted layer by layer from top to bottom according to the chain rule. The network parameters of the CNN-BP neural network prediction model are continuously updated until the error no longer decreases. Then the trained network parameters are saved as the model.
[0022] The GPS positioning method based on CNN-BP neural network of the present invention has the following beneficial effects: a CNN-BP neural network model is proposed to solve this problem, and various parameters in the CNN-BP model, input layer structure, convolution layer structure, pooling layer structure, BP neural network layer structure and output layer structure are designed, and the corresponding CNN-BP model training is designed for the GPS system in this article. The influence of multipath effect on the GPS process is reduced, so that the positioning accuracy of GPS in the case of high-rise buildings in the city is greatly improved, and the method of the present invention is simple and low-cost, and can be directly used in the positioning mode of GPS mobile phones. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] Figure 1 Design idea diagram of CNN-BP model
[0024] Figure 2 Convolution module structure diagram
[0025] Figure 3 Pooling module structure diagram
[0026] Figure 4 Input to the pooling process
[0027] Figure 5 BP neural network layer structure
[0028] Figure 6 Schematic diagram of convolutional layer to feedforward neural network
[0029] Figure 7 BP neural network training method
[0030] Figure 8 Experimental Results
[0031] (a) Highly predicted situation
[0032] (b) Latitude prediction
[0033] (c) Longitude prediction DETAILED DESCRIPTION
[0034] The proposed method mainly includes: input layer structure design, convolution layer structure design, pooling layer structure design, BP neural network design and CNN-BP model training. The main steps include:
[0035] The first step is to build a data set and perform data preprocessing;
[0036] The data used in the present invention comes from the GPS data collected by the DB502 board. A total of 5689 groups of GPS information data are collected, and the real GPS data corresponding to each location point is output. 4000 groups of data are used as training data, and the remaining 1689 groups of data are used as test data for experiments.
[0037] The second step is to build a CNN-BP neural network prediction model: the main design idea is: the GPS position and the number of satellites are normalized and then input into the convolution layer, and then the eigenvalue and eigenvalue matrix are generated after the convolution operation, and then the eigenvalue matrix is reduced in dimension through the pooling layer to obtain the eigenvalue, and then these eigenvalues are input into the BP neural network layer, and finally the relevant GPS position prediction value is output. The CNN-BP neural network prediction model mainly consists of five parts, which will be introduced in detail below:
[0038] (1) The design structure diagram of the CNN-BP neural network prediction model is shown in the attached figure. Figure 1 As shown in the figure, the five parts are input layer, convolution layer, pooling layer, BP neural network layer and output layer.
[0039] (2) According to the actual input situation, the input layer parameters include GPS position, speed and satellite information, forming a 3×3 input. GPS position includes three GPS position information, speed packet includes three speed information, and satellite information includes satellite number, accuracy and satellite time.
[0040] (3) The data processed from the input layer will pass through the convolution layer. The convolution layer is equivalent to disconnecting some of the connections of the fully connected neural network and converting it into a locally connected neural network. The parameters of the convolution kernel in the convolution layer are the weights of the locally connected neural network. Its main function is to extract and optimize the input feature values. The convolution module consists of 2 (1×1) convolution kernels with a step size of 1, 2 (2×2) convolution kernels with a step size of 1, and 2 (3×3) convolution kernels with a step size of 1. The structure diagram of the convolution layer module is shown in the figure. Figure 2 As shown. The idea of this design is: the input is a 3×3 data matrix, the 1×1 convolution kernel connects each input data; the 2×2 convolution kernel is equivalent to partially connecting the 2×2 part of the input data matrix; the 3×3 convolution kernel is equivalent to fully connecting the input data matrix. After convolution processing by the convolution module, the 1×1 convolution kernel convolves each input data to obtain 2 3×3 feature matrices, the 2×2 convolution kernel is equivalent to convolving the 2×2 part of the input data matrix to obtain 2 2×2 feature matrices, and the 3×3 convolution kernel convolves the input fully connected data matrix to obtain 2 eigenvalues.
[0041] (4) After convolution processing, the data enters the pooling layer. The main function of the pooling layer is to reduce the dimension of features. The pooling module consists of 2 (2×2) pooling kernels and 2 (3×3) pooling kernels, such as Figure 3 In order to convert two 3×3 feature matrices and two 2×2 feature matrices into four eigenvalues, two 2×2 pooling operations and two 3×3 pooling operations are required. Specifically, two 3×3 feature matrices are passed through two (3×3) pooling kernels to obtain two eigenvalues, and two 2×2 feature matrices are passed through two (2×2) pooling kernels to obtain two eigenvalues. Finally, adding the two eigenvalues obtained in the convolution process, a total of six optimized eigenvalues can be obtained, which are input into the next BP neural network layer and input into the pooling process as shown in the figure. Figure 4 shown.
[0042] (5) Design the BP neural network layer structure. The 6 eigenvalues output by the pooling layer are used as input to the fully connected BP neural network layer, and 20 neurons are set. The BP neural network layer structure is shown in the figure below: Figure 5 shown.
[0043] (6) Design the output layer structure. The output layer is the actual GPS location. In the present invention, three indicators are determined, namely longitude, latitude, and altitude.
[0044] The third step is CNN-BP model training. CNN-BP adds convolutional layers and pooling layers to the BP neural network to optimize the original eigenvalues. It is a multi-layer feedforward neural network that uses the error back propagation algorithm for training. The pooling layer has no parameters, so it does not need to be trained. The BP neural network part is fully connected and needs to be adjusted layer by layer from top to bottom according to the chain rule. The convolution layer also needs to be trained through error back propagation. CNN-BP model training includes two parts, the first is the error back propagation method of the convolution layer, and the second is the error back propagation method of the BP neural network layer.
[0045] a) The principle of the back propagation training method of the convolution layer is to convert the convolution operation into the form of a feedforward neural network. X represents the neuron input variable, W represents the convolution kernel parameter, that is, the weight, and Y represents the result of the convolution operation. It is obvious from the figure that the convolution operation is actually the partial connection between the second layer of the neural network and the output layer neurons, and the value W in the convolution kernel is the weight between them. The weight adjustment of the convolution kernel is the same as that of a multi-layer neural network, starting from the connection weight of the upper layer and adjusting layer by layer.
[0046] Δw 2jk =η(r k -y k )y k (1-y k ) j (1)
[0047] According to the chain rule of formula (1), X 11 and Y 31 The connection weight between The adjustment value is shown in the following formula.
[0048]
[0049] The above formula (2) consists of three parts: j The sum of the derivatives of the error function and activation function of the connected previous layer, z j The derivative of the activation function, and w 1jk The relevant input x i .
[0050] X 12 and Y 32 The connection weights between Its adjustment value is shown in formula (3).
[0051]
[0052] Since the same convolution kernel is used to perform convolution operations on multiple positions, the final adjustment value can be obtained by accumulating the connection weight adjustment values of all units, as shown in the formula.
[0053]
[0054] All convolution kernels in the convolutional neural network are processed according to the above mentioned method. From the above discussion, it can be seen that after the convolution kernel of the convolution layer is regarded as the weight of the feedforward neural network, the convolution kernel element can be adjusted according to the adjustment value of the unit. The schematic diagram of the convolution layer to the feedforward neural network is as follows Figure 6 shown.
[0055] b) The back propagation training method used by the BP neural network in the CNN-BP neural network is as follows Figure 7 As shown, first, the connection weight w between the output layer and the intermediate hidden layer is 2jk Taking the derivative, we get formula (5):
[0056]
[0057] In formula (3-5), the output y is k Derivative, output y k For the activation value u 2k Derivative, activation value u 2k The connection weight w 2jk After taking the derivative, we can get formula (6).
[0058]
[0059] Then the connection weight w between the input layer and the intermediate layer 1jk Derivative, as shown in formula (7).
[0060]
[0061] As shown in formula (7), the error function E performs a weighted summation of the derivatives of all output units. As in the previous process, the derivatives of the error function and the sigmoid function are substituted into the formula to obtain formula (8).
[0062]
[0063] Since the connection weight w 1ij Only for the middle layer z j The state of , so the remaining part of the formula is derived to obtain formula (9).
[0064]
[0065] Activation value u 2k Right j Derivation to obtain the connection weight w 2ij , combined with formula (10), we can get the connection weight w between the input layer and the middle layer 1ij The adjustment value is shown in formula (11).
[0066]
[0067]
[0068] The fourth step is to input the GPS information affected by the multipath effect, load the model trained in the third step, and output the GPS prediction information.
[0069] The fifth step is to compare the experimental results of the present invention with those of the traditional BP algorithm. Figure 8 The CNN-BP algorithm reduces the error in longitude prediction by 4.91%, the error in latitude by 22.2%, and the error in altitude prediction by 19.7%.
Claims
1. A GPS positioning method based on CNN-BP neural network, comprising the following steps: The first step is to construct a GPS data set and perform data preprocessing including GPS position, speed, and satellite information; The second step is to build a CNN-BP neural network prediction model: The CNN-BP neural network prediction model consists of five parts, namely the input layer, convolution layer, pooling layer, BP neural network layer and output layer, among which, The input layer includes GPS position, speed and satellite information to form a 3×3 input matrix. The GPS position includes GPS longitude, latitude and altitude information. The speed includes acceleration, angular velocity and speed information. The satellite information includes the number of satellites, accuracy and satellite time information. Convolution layer, used to extract and optimize input feature values. The convolution module consists of 2 1×1 convolution kernels, 2 2×2 convolution kernels, and 2 3×3 convolution kernels. For a 3×3 input matrix, the 1×1 convolution kernel connects each input data; the 2×2 convolution kernel partially connects the 2×2 part of the input data matrix; and the 3×3 convolution kernel fully connects the input data matrix. After the input matrix is processed by the convolution layer, it enters the pooling layer for feature dimensionality reduction. The pooling module consists of 2 (2×2) pooling kernels and 2 (3×3) pooling kernels. After the convolution layer processing, 2 3×3 feature matrices, 2 2×2 feature matrices and 2 eigenvalues are obtained. For the 4 feature matrices, 2 2×2 pooling operations and 2 3×3 pooling operations are performed to obtain 6 optimized eigenvalues. The 6 optimized feature values obtained by the pooling layer are used as input to the fully connected BP neural network layer for feature fitting; then the GPS prediction results are output through the output layer, including the longitude, latitude and altitude of GPS; The third step is model training: input the GPS information in the prepared training set into the CNN-BP neural network prediction model. The convolution layer and the pooling layer are trained using the error back propagation algorithm. The BP neural network part is fully connected and needs to be adjusted layer by layer from top to bottom according to the chain rule. The network parameters of the CNN-BP neural network prediction model are continuously updated until the error no longer decreases. Then the trained network parameters are saved as the model.
