Satellite multipath interference signal suppression method, system, device and storage medium

Abstract

The present invention provides a satellite multipath interference signal suppression method, system, device and storage medium, belonging to the field of image processing. The method comprises the following steps: converting a satellite time domain signal into a multipath signal image matrix; adding a channel attention mechanism after each layer of a densely connected block of a densely connected network to obtain an SE densely connected network, and then adding a long short-term memory network to obtain an SDLNet model; obtaining first and second multipath signal sample image data sets, inputting the images in the first and second multipath signal sample image data sets into the SDLNet model respectively, and respectively obtaining a multipath interference signal detection model and a multipath interference signal suppression model; inputting the images in the multipath signal image matrix into the multipath interference signal detection model to obtain a frequency deviation value, a time delay value and a phase shift value of the multipath interference signal; and simultaneously inputting the images and the multipath signal image into the multipath interference signal suppression model to obtain a de-interferenced image.

Description

Technical Field

[0001] The present invention belongs to the field of image processing, and in particular relates to a method, system, device and storage medium for suppressing satellite multipath interference signals. Background Art

[0002] The Global Navigation Satellite System (GNSS) provides global positioning, velocity measurement, and timing services, and is the primary tool for positioning and navigation in our daily lives. Currently, these systems include my country's independently developed Beidou system, the US GPS system, Russia's GLONASS system, and the EU's Galileo system. In recent years, with the steady progress of society and the continuous development of science and technology, the construction of smart cities, the implementation of autonomous driving, and autonomous drone navigation have placed higher demands on the positioning accuracy of navigation satellites. Therefore, how to overcome various interference in the channel and obtain more accurate location information has become a research focus for scholars.

[0003] GNSS primarily achieves positioning through two methods: pseudorange measurement and carrier phase ranging. Specifically, both methods calculate the relative distance between the transmission of a GNSS signal by a navigation satellite and its reception by a receiver. Finally, the receiver's spatial position is calculated by combining the intersection of the four satellite ranging spaces. The former uses pseudorange codes in the signal for ranging, while the latter uses the phase of the carrier wave. Pseudorange ranging is commonly used in navigation systems such as mobile phones, cars, and ships, and is a common sight in everyday life. However, due to the low power and weak penetration of GNSS signals and the complexity of the real-world environment, signal propagation and reception are often affected, resulting in reduced positioning accuracy and even drift. Factors that affect positioning accuracy include incomplete synchronization between the receiver and satellite clocks, refraction and reflection of signals through the atmosphere and troposphere, which lengthens the propagation path, errors in the satellite's own orbital position, and multipath effects in urban areas and mountains. Multipath effects of GNSS satellite signals are considered the primary factor affecting urban navigation positioning accuracy, and their detection and mitigation are currently a pressing issue.

[0004] In recent years, machine learning, a branch of artificial intelligence, has demonstrated remarkable performance in fields such as image processing, natural language processing, target detection, and autonomous driving. A growing number of researchers are applying it to multipath detection and mitigation, creating a new solution for addressing the multipath effect. Some researchers have used convolutional neural networks to predict multipath signal parameters using image representation. However, due to the limited number of multipath signal parameters and the random nature of multipath effects, the parameters are randomly distributed within a certain range, affecting the accuracy of multipath interference signal detection. Furthermore, existing neural network-based research has primarily focused on multipath detection, while relatively little research has been conducted on multipath mitigation. Summary of the Invention

[0005] In order to overcome the above-mentioned deficiencies in the prior art, the present invention provides a method for suppressing satellite multipath interference signals, comprising the following steps:

[0006] Obtain satellite time domain signals and convert them into multipath signal image matrix;

[0007] A channel attention mechanism SE module is added after each layer of the dense connection block of the dense connection network to obtain the SE dense connection network. A long short-term memory network LSTM is added after the SE dense connection network to obtain the SDLNet model.

[0008] Obtain a first multipath signal sample image dataset and a second multipath signal sample image dataset, input images in the first multipath signal sample image dataset into an SDLNet model, train the SDLNet model, and obtain a multipath interference signal detection model; input images in the second multipath signal sample image dataset into the SDLNet model, train the SDLNet model, and obtain a multipath interference signal suppression model;

[0009] The image in the multipath signal image matrix is ​​input into the multipath interference signal detection model to obtain the frequency deviation value, time delay value and phase shift value of the multipath interference signal; the frequency deviation value, time delay value, phase shift value and the multipath signal image are simultaneously input into the multipath interference signal suppression model to obtain the image after interference removal.

[0010] Preferably, the step of converting the satellite time domain signal into a multipath signal image matrix is ​​as follows: copying the one-dimensional satellite time domain signal into N t times, and then use the one-dimensional autocorrelation function to replicate the one-dimensional satellite time domain signal N times. f times, get two N t ×N f Matrix, and then two N t ×N f The matrix is ​​multiplied bit by bit to obtain a two-dimensional multipath signal image matrix, where N tN is the number of delay points to be observed. f is the number of frequency deviation points to be observed.

[0011] Preferably, the SE module is used to perform a squeezing operation and an excitation operation on the input feature map. The SE module is used to perform a squeezing operation and an excitation operation on the input feature map, wherein the squeezing operation squeezes the feature map into a 1*1*C feature vector through a global pooling operation; the excitation operation is completed through two linear connection layers, the first linear connection layer reduces the feature vector of size 1*1*C to 1*1*C / redu, and is connected to the first linear connection layer through a ReLu activation function, the first linear connection layer restores the dimension of the input feature vector to 1*1*C, and outputs a weight matrix through a Sigmoid function, and then multiplies the weight matrix with the original feature matrix to obtain a reconstructed feature map.

[0012] Preferably, the number of samples in the first multipath signal sample image data set is 5,000 groups; the number of samples in the second multipath signal sample image data set is 50,000 groups.

[0013] Preferably, the performance of the multipath interference signal suppression model is evaluated by structural similarity.

[0014] Preferably, the performance of the multipath interference signal detection model is evaluated by a determination coefficient.

[0015] The present invention also provides a satellite multipath interference signal suppression system, comprising:

[0016] A signal conversion module is used to obtain satellite time domain signals and convert them into a multipath signal image matrix;

[0017] The basic model construction module is used to add a channel attention mechanism SE module after each layer of the dense connection block of the dense connection network to obtain the SE dense connection network, and add a long short-term memory network LSTM after the SE dense connection network to obtain the constructed SDLNet model;

[0018] a signal processing model construction module, configured to obtain a first multipath signal sample image dataset and a second multipath signal sample image dataset, input images from the first multipath signal sample image dataset into an SDLNet model, train the SDLNet model, and obtain a multipath interference signal detection model; and input images from the second multipath signal sample image dataset into the SDLNet model, train the SDLNet model, and obtain a multipath interference signal suppression model;

[0019] The signal interference removal module is used to input the image in the multipath signal image matrix into the multipath interference signal detection model to obtain the frequency deviation value, time delay value, and phase shift value of the multipath interference signal; and simultaneously input the frequency deviation value, time delay value, phase shift value and the multipath signal image into the multipath interference signal suppression model to obtain the image after interference removal.

[0020] The present invention also provides a computer device, comprising a memory and a processor; the memory stores a computer program, and the processor is used to run the computer program in the memory to execute the satellite multipath interference signal suppression method.

[0021] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor to execute the satellite multipath interference signal suppression method.

[0022] The satellite multipath interference signal suppression method, system, device, and storage medium provided by the present invention have the following beneficial effects:

[0023] The present invention adds a channel attention mechanism SE module after each layer of a densely connected block of a densely connected network, so that each layer of the network in the densely connected block connects and reuses all feature images of the upper layer, thereby enhancing the characterization capability of the densely connected network for image signal features, thereby greatly improving the prediction accuracy of the three parameters of the SE densely connected network in terms of frequency deviation, delay and phase shift of multipath interference signals; the present invention adds a long short-term memory network LSTM after the SE densely connected network to obtain an SDLNet model, which can learn the temporal correlation in the signal, remove multipath contaminated images, and make the reconstructed image have a high similarity with the interference-free direct signal image, indicating that the SDLNet model has a better multipath suppression effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] To more clearly illustrate the embodiments of the present invention and its design, the following briefly introduces the drawings required for this embodiment. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be derived from these drawings without inventive effort.

[0025] Figure 1 is a flow chart of an embodiment of the present invention;

[0026] Figure 2 Schematic diagram of residual connection;

[0027] Figure 3 Schematic diagram of dense connection;

[0028] Figure 4Schematic diagram of the internal structure of the densely connected block;

[0029] Figure 5 Perform operations inside the feature extraction F(x) layer in the densely connected block;

[0030] Figure 6 This is a schematic diagram of the specific structure of the transition layer and its network location;

[0031] Figure 7 Schematic diagram of feature map changes under SE attention mechanism;

[0032] Figure 8 Schematic diagram of the specific structure of the dense connection layer with channel attention mechanism;

[0033] Figure 9 This is the LSTM structure diagram;

[0034] Figure 10 The overall architecture of the network SDLNet of the present invention;

[0035] Figure 11 is a schematic diagram of multipath signal images and parameter samples, where: Figure 11 a.1, a.2 and a.3 are I branch images, Figure 11 b.1, b.2 and b.3 are Q branch images;

[0036] Figure 12 This is a flow chart of multipath detection and suppression;

[0037] Figure 13 This is a sample of channel feature visualization, where Figure 13 a.1, a.2, a.3 and a.4 are the outputs of the convolutional layer, Figure 13 b.1, b.2, b.3 and b.4 are the outputs of the SE dense connection block (1);

[0038] Figure 14 R 2 Compare the line chart, where Figure 14 (a) is the comparison of Δτ correlation coefficient; Figure 14 (b) Δf correlation coefficient comparison; Figure 14 (c) is the comparison of Δθ correlation coefficient;

[0039] Figure 15 The SIMM and loss values ​​of SDLNet in multipath suppression change with the training rounds; Figure 15 (a) is the relationship curve between the structural similarity of the image after removing multipath interference and the direct signal image and the training rounds; Figure 15 (b) is the changing curve of the loss function during the training process;

[0040] Figure 16 is the frequency slice diagram, where Figure 16 (a.1) SDLNet suppresses output 1; (a.2) SDLNet suppresses output 2; (b.1) ECAResNet suppresses output 1; (b.2) SDLNet suppresses output 2; (c.2) RegNet suppresses output 1; (c.3) RegNet suppresses output 2. DETAILED DESCRIPTION

[0041] In order to enable those skilled in the art to better understand the technical solution of the present invention and to be able to implement it, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solution of the present invention and are not intended to limit the scope of protection of the present invention.

[0042] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "axial", "radial", "circumferential" and the like to indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the technical solutions of the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0043] In addition, the terms "first", "second", etc. are used for descriptive purposes only and are not to be understood as indicating or implying relative importance. In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meaning of the above terms in the present invention can be understood according to the specific circumstances. In the description of the present invention, unless otherwise specified, "plurality" means two or more, which will not be described in detail here.

[0044] Example

[0045] The present invention provides a method for suppressing satellite multipath interference signals, specifically as follows Figure 1 As shown, the following steps are included:

[0046] Step 1: Obtain satellite time domain signals and convert them into multipath signal image matrix.

[0047] The satellite time-domain signal generated by the Global Navigation Satellite System (GNSS) typically consists of three parts: a carrier, a pseudo-random number (PRN), and a navigation message. To achieve pseudo-range measurement, the ideal form of the direct signal received by the receiving antenna after the satellite time-domain signal propagates through space can be expressed as:

[0048]

[0049] Where A represents the received power of the signal, D(t) represents the binary navigation message at time t, and C(t) represents the PRN sequence, which is determined by the satellite that sends the signal and is f c Carrier frequency.

[0050] However, due to the influence of various interference factors in the actual channel during the propagation of GNSS signals in space, phenomena such as time delay, phase offset, and power attenuation will occur. Therefore, the signal received by the receiver cannot be directly expressed by equation (1). In addition, due to the high-speed operation of navigation satellites, they usually have a large relative motion with the ground receiver, and the resulting Doppler frequency offset is often not negligible. Taking all the above factors into consideration, the actual received signal can be modeled as equation (2):

[0051]

[0052] Where A represents the received power of the signal, D(t-τ) represents the binary navigation message at time t-τ, C(t-τ) represents the PRN sequence at time t-τ, τ represents the time delay, δf represents the Doppler frequency deviation, and θ represents the phase offset.

[0053] The replica signal in the multipath effect is caused by refraction, reflection and other phenomena. In most cases, refraction and reflection will not change the main structure of the signal. Therefore, the form of the multipath interference signal should be similar to that of formula (1), but with differences in parameters such as amplitude, phase, and delay. Therefore, the multipath replica signal is expressed by formula (3):

[0054]

[0055] Among them, A mp , τ mp ,δf mp ,θ mp They represent the received power, signal delay, Doppler frequency deviation and phase shift of different multipath replica signals respectively.

[0056] The signal received by the sending and receiving end consists of the direct signal and the multipath replica signal, so the superposition of the two can obtain the received signal S r (t), its formula is described as formula (4):

[0057]

[0058] Due to the thermal effect of the hardware circuit and other reasons, there will be interference such as thermal noise in the received signal. Therefore, additive white Gaussian noise b(t) is used in equation (4) to simulate this influence.

[0059] The received signal is processed by radio frequency such as power amplification and then enters the correlator loop for digital signal processing. d2 (t) is divided into two paths and multiplied with the local oscillator signal and the signal shifted by 90 degrees respectively to obtain the signals of the in-phase branch I and the orthogonal branch Q after removing the carrier. Then, the correlator output of the two-branch signal is completed through coherent integration.

[0060] The signal output by the correlator can be expressed as:

[0061]

[0062] in, Indicates the signal's carrier-to-noise ratio, T i represents the time of coherent integration, a is the amplitude related to the navigation message, Δτ is the error between the estimated delay and the actual delay of the signal, R(*) represents the autocorrelation function of the PRN code, Δf is the frequency offset error between the local oscillator signal and the actual signal, Δθ is the phase error between the local oscillator signal and the actual signal, and n I 、n Q The additive white Gaussian noise of I and Q branches, sinc(πΔf T i ) is shown in formula (7).

[0063]

[0064] Since the entire system is a linear system, according to the linear superposition principle of linear systems, the I and Q branch outputs of the actual received signal can be expressed by equations (8) and (9), which are equivalent to the superposition of the direct signal and the multipath signal:

[0065]

[0066] Where I MP , Q MP The forms of are consistent with those of equations (5) and (6), respectively, except that the parameters a, estimation errors Δτ, Δf, and Δθ have different values.

[0067] The signal without multipath interference is constructed by equations (5) and (6), and the multipath signal with interference is constructed by equations (8) and (9). In GNSS signals, PRN codes mainly include C / A code and P code. C / A code is commonly used for civilian positioning, and P code is commonly used for military positioning. This paper takes C / A code as an example. The code rate of C / A code is 1.023MHz, the code length is 1023, and the code period is 1ms. Due to the discrete nature of digital signal processing, the signal needs to be sampled and processed, and the sampling frequency is set to 20MHz.

[0068] Set the multipath interference delay range to (-2 / 3T c , 5T c ), the range of Doppler frequency deviation is δf<|min(5.5 / T i ,800+5 / T i )|, where T c is the chip width, T i Coherent integration time. Because the sinc function value is much smaller than its maximum value after exceeding this interval, it can be ignored. The coherent integration time is 20ms, and the phase offset range is (0, 2pi). Specific data are shown in Table 1.

[0069] Table 1 Data parameter settings

[0070] Simulation parameters Value <![CDATA[C / A code rate F c > 1.023MHz Code length 1023 Code period 1ms <![CDATA[Sampling rate F s > 20MHz <![CDATA[Coherent integration time T i > 0.02s Delay range (-1.5 / Fc, 5 / Fc) Frequency error range (-275Hz, 275Hz) Phase error range (0，2π) Amplitude range (0.6，0.9)

[0071] It is worth noting that the process of converting time domain data into image data, taking the in-phase branch as an example, before performing the in-phase branch correlation calculation, the one-dimensional time domain signal is copied N times. t times, the one-dimensional autocorrelation function is replicated N times. f times, get two N t ×N f The two matrices are then multiplied to obtain the processed two-dimensional image matrix. t is the number of delay points to be observed, N f The number of frequency deviation observation points to be observed is set to 80. The number of delay observation points is determined by the delay observation range. The number of replications described above is determined by the number of observation points, which is equal to the number of replications.

[0072] Step 2: Add a channel attention mechanism SE module after each layer of the densely connected block of the densely connected network to obtain the SE densely connected network, and add a long short-term memory network LSTM after the SE densely connected network to obtain the constructed SDLNet model.

[0073] Deep learning is an important branch of artificial intelligence, providing new solutions to many complex problems. In recent years, deep learning has continued to develop, especially in the fields of image recognition, video behavior detection, natural language processing, etc., and has made important breakthroughs. The deep learning network provides an "input-output" end-to-end black box data processing model, which not only avoids a large number of complex feature extraction processes, but also improves the accuracy of recognition and prediction compared to traditional methods. In order to better apply deep learning to multipath signal processing and improve the effect of multipath detection and suppression, the present invention has converted the time domain waveform signal into image data and stored it in a matrix during the data preprocessing process. Each group of images also includes labels such as time delay, frequency deviation, phase shift, navigation message, etc. for subsequent model training.

[0074] The existing technology compares MLP and CNN. In terms of CNN, a network composed of 4 convolutional layers and 3 fully connected layers is mainly constructed to classify pure noise signals and useful signals containing noise signals. Since there are fewer classification targets, a shallow network can achieve good results. Some existing technologies combine the histogram loss to construct a neural network similar to the VGG16 structure to estimate the multipath effect, and the effect is better than the traditional CNN regression model. However, experiments have found that when using image data to directly restore signals contaminated by multipath, it is necessary to learn feature information of more scales in order to obtain a better suppression effect. Therefore, the present invention intends to construct a deep learning model (SE-Dense-LSTM Net, SDLNet) that integrates channel attention dense connection network and LSTM to simultaneously realize multipath parameter estimation and multipath pollution suppression. The network details will be described in detail below.

[0075] The basic idea of ​​deep learning is to build a suitable network model, use a large amount of known data, and use the back-propagation algorithm to continuously adjust and optimize the network parameters from the difference between the predicted results and the true labels, so as to obtain a mapping from input to output, so that the predicted results gradually approach the true labels. At the beginning of the application of deep learning, the problems dealt with were relatively simple and the number of network layers was relatively shallow. As the research continued to deepen, scholars found that the reasonable construction of deeper neural networks can extract higher-dimensional features that are difficult to identify by human power, which is conducive to dealing with more complex practical problems. Initially, in the deep learning network, the input of this layer only came from the output of the previous layer, and the output of this layer was only used as the input of the next layer. This single input source caused the neural network to have gradient vanishing or gradient explosion during the back-propagation process, which greatly reduced the network learning effect. The emergence of residual networks not only provides a new solution to such problems, but also brings the number of network layers to a new depth. The core of the residual network is the idea of ​​residual connection, that is, the original input features of the previous learning unit will also be used as the input features of the next unit. After being added to the output features of this unit, they are input to the next layer to participate in the operation of the next unit. Its residual connection method is as follows Figure 2 shown.

[0076] like Figure 3 As shown in the figure, the dense connection network draws on the basic idea of ​​residual connection in the residual network to build a dense connection block based on forward splicing. The output of each layer in the dense connection block will participate in the subsequent calculation in the form of cascade splicing. Specifically, the internal structure diagram of a dense connection block is as follows Figure 4 As shown in the figure, the dense connection block contains 3 layers (the number of layers is determined by the actual learning problem). The input of each layer contains the feature maps of all previous layers. Assuming that the original feature input to the dense connection block is X0, the feature extraction process of each layer is recorded as function F(*), and the feature map obtained after processing the i-th layer is recorded as X i , according to the dense connection principle described above, we can get formula (10).

[0077] X i =F([X0, X1, X2, ..., X n-1 ]) (10)

[0078] The feature map is batch normalized (BN), nonlinearly activated (ReLu), and image convolution (CONV) in F(*). Each layer of the network performs this operation once. The specific operations are as follows: Figure 5As shown. Due to the reuse of shallow network feature data and image convolution operations, the distribution of data features will change after being processed by a layer of network (this phenomenon is called internal co-vector transfer). In order to alleviate the above phenomenon and ensure the consistency of the input feature distribution of each layer of the network, the input feature map is first batch normalized in the BN layer in each layer. Since the distribution of input features of each layer of the network is consistent, there is no need to adapt to the new data distribution during network learning, so the BN operation also speeds up the convergence of the network. The image of the ReLu function is shown in the figure, and its analytical expression can be expressed as f(x) = max(0, x). Since the data of the feature map of the research object in the present invention are all positive values, the use of this activation function not only improves the network's ability to characterize nonlinear factors, but also simplifies the network's calculation process due to its relatively simple derivative, thereby speeding up the update speed of weights and the learning speed of the network. Image convolution is mainly used to extract image features.

[0079] By observing formula (10), we can find that as the number of network layers increases, the dimension of the feature map continues to increase. If the output features of the densely connected blocks are not reasonably reduced in dimension, the accumulation of a large number of redundant features will not only be of no help in improving the learning ability of the network, but will also greatly increase the learning burden of the network and affect the convergence time of the network. Therefore, the densely connected network adds a transition layer between every two densely connected blocks. The transition layer reduces the number of channels of the feature map through a 1*1 convolution operation, and reduces the length and width of the feature map to half of the original through an average pooling layer with a step size of 2. The transition layer simplifies the parameters of the network and makes the network structure more compact. Its specific structure and placement are as follows: Figure 6 shown.

[0080] like Figure 7As shown in Figure 1, the various parameters of multipath signals are hidden in the image, but it is difficult to observe the connection between them with the naked eye. For example, the highlighted position of the image contains rich useful information, while the boundary position contains less information. In order to dynamically adjust the weights of each part of the learning, it is possible to consider introducing an attention mechanism. The SE module belongs to the category of channel attention. Since the convolution operation and feature splicing process will generate a large number of channel features, and the importance of various channel features for detecting and suppressing multipath signals is inconsistent, in order to extract relevant important features and suppress non-important redundant features, a channel attention mechanism can be used. During the training process of the neural network, the SE module will generate a weight matrix to measure the importance of each channel and reconstruct the features through the weight matrix to achieve the purpose of highlighting the key points and omitting the redundancy. The SE module mainly performs squeeze and excitation operations. Specifically, after the feature map is input into the SE module, a global pooling operation is used to squeeze the original feature map into a 1*1*C vector. At this time, a channel is represented by a value, that is, a value has the global receptive field of the channel. The formula form of this operation is shown in Equation (11):

[0081]

[0082] Where X is the input feature, W and H are the width and height of the input feature map, and X(i, j) is the feature map.

[0083] The excitation operation is completed by two linear connection layers. The first linear connection layer reduces the feature vector of size 1*1*C to 1*1*C / redu, and connects to the next fully connected layer through the ReLu activation function. This layer restores the dimension of the feature vector to 1*1*C, and outputs the weight matrix through the Sigmoid function. The reconstructed feature map is obtained by multiplying the weight matrix with the original feature matrix. In the channel weight extraction, the feature map change process is as follows Figure 7 As shown, it is deeply integrated with the dense connection layer to construct a dense connection network layer G(x) with channel attention mechanism. The specific implementation structure is as follows Figure 8 shown.

[0084] After the time domain signal is mapped into the image signal, the horizontal axis of the image is the expansion of the time delay observation point, and the vertical axis is the expansion of the frequency deviation observation point. Therefore, there is still some time series information hidden in the image. In order to reasonably utilize the time domain information contained in the image, it is considered to introduce the long short-term memory network (LSTM) into the deep learning network. The long short-term memory network is composed of several unit memory cells. Each unit cell has a consistent structure. The state of the feature vector is transmitted in these unit cells. The features learned by the previous unit cell are used as memory input to the next unit cell and fused with the input features at this moment to jointly determine the feature information to be learned. Specifically, the unit cell is composed of an input gate, a forget gate, and an output gate. The input gate (i t ) determines the amount of current input information that needs to be retained in the unit cell; the forget gate (f t ) selectively determines the redundant information to be discarded in the features of the previous unit cell input, retains the useful feature information, and the output gate (o t ) controls the update of the current cell information and the amount of characteristic information transmitted to the next cell. The specific structure is as follows Figure 9 shown.

[0085] Based on the above hierarchy, the present invention constructs a densely connected network with channel attention, which also integrates LSTM to extract implicit temporal relationships. The overall architecture of the network is as follows: Figure 10 shown.

[0086] In order to adapt to different learning tasks, the present invention changes the number of convolutional layers in the SE dense connection block in the above architecture, and obtains two sizes of learning networks. The specific structure of their hierarchies is shown in Table 2.

[0087] Table 2 Specific hierarchical structures of the two networks

[0088]

[0089] Step 3: Obtain a first multipath signal sample image dataset and a second multipath signal sample image dataset, input the images in the first multipath signal sample image dataset into the SDLNet model, train the SDLNet model, and obtain a multipath interference signal detection model; input the images in the second multipath signal sample image dataset into the SDLNet model, train the SDLNet model, and obtain a multipath interference signal suppression model.

[0090] The present invention can generate a data set according to the method of step 1. In the generated data set, each sample unit contains 9 sets of data, which are the GNSS direct signal presented in the form of an image matrix in the I and Q branches, the superimposed signal contaminated by the multipath effect presented in the form of an image matrix in the I and Q branches, and the time delay, Doppler frequency deviation, phase shift, navigation message and amplitude of the multipath signal. The generated signal image and multipath parameters are as follows: Figure 11 As shown, Figure 11 a.1, a.2 and a.3 are I branch images, Figure 11 b.1, b.2 and b.3 are Q branch images, Figure 11 a.1 and Figure 11 The multipath parameters in b.1 are mp_τ = 0.968 ms, mp_f = 42.931 Hz, and mp_θ = 4.051 rad; Figure 11 a.2 and Figure 11 The multipath parameters in b.2 are mp_τ = 1.263 ms, mp_f = -35.666 Hz, and mp_θ = 4.207 rad; Figure 11 a.3 and Figure 11 The multipath parameters in b.3 are mp_τ = 1.165ms, mp_f = 66.6331Hz and mp_θ = 1.860rad.

[0091] The generated data set includes two data sets: a first multipath signal sample image data set and a second multipath signal sample image data set. By inputting the first multipath signal sample image data set and the second multipath signal sample image data set into the SDLNet model respectively and training the SDLNet model, a multipath interference signal detection model and a multipath interference signal suppression model can be generated accordingly. When training the generated multipath interference signal detection model for multipath parameter estimation, the amount of data required for model convergence is much smaller than the amount of data required for multipath suppression training of the multipath interference signal suppression model. This is because multipath parameter estimation extracts features from the original image and ultimately regresses a multipath parameter, while multipath suppression optimizes and reorganizes the contaminated image data and ultimately regresses an image. The latter is more complicated, and a large number of structural parameters need to be determined and calculated, so more data is required. The process of multipath detection and suppression is as follows: Figure 12 As shown in the figure, for multipath detection, the total number of samples is set to 5,000; for multipath mitigation, the total number of samples is set to 50,000. The training set, validation set, and test set account for 60%, 20%, and 20% of the total number of samples, respectively.

[0092] Step 4: Input the image in the multipath signal image matrix into the multipath interference signal detection model to obtain the frequency deviation value, delay value, and phase shift value of the multipath interference signal; input the frequency deviation value, delay value, phase shift value, and multipath signal image into the multipath interference signal suppression model at the same time to obtain the image after interference removal.

[0093] R 2 Also known as the coefficient of determination, it represents the proportional relationship between the regression square and the total deviation square sum, and is often used to evaluate the degree of fit between the regression model and the observed data. Its calculation formula is shown in formula (12), where SST is the total deviation square sum, which represents the sum of the differences between the observed values ​​and the data mean, and SSR is the regression square sum, which represents the sum of the differences between the predicted values ​​and the data mean. The calculation formulas of the two correspond to the numerator and denominator of the second equation respectively. In order to simplify the calculation, it is often simplified to the form of 1-SSE / SST, as shown in formula (13). R 2 The value range is (0, 1). In extreme cases, when R 2 = 0, indicating that the prediction model is completely inconsistent with the observed data. 2 =1 means that the prediction model completely fits the observed data and has a good explanation effect on the observed data. 2 The larger the value and the closer it is to 1, the better the prediction effect of the model.

[0094]

[0095] Where, is the model prediction value, y i is the actual observed value, is the mean of the observations.

[0096] In order to clarify the features extracted by a single layer network in the deep learning network, the convolutional layer and SE dense connection block (1) shown in Table 3 randomly extract the output features of a certain channel and present them in the form of 40*40 and 20*20 pictures. The visualization results of these feature maps are shown in Figure 13 As shown, Figure 13 a.1, a.2, a.3 and a.4 are the outputs of the convolutional layer, Figure 13 b.1, b.2, b.3 and b.4 are the outputs of the SE dense connection block (1).

[0097] Table 3 shows a comparison of the performance of the network proposed in this paper with other mainstream networks in the test set. In terms of multipath prediction, the network proposed in this paper has a determination coefficient of 99.01%, 99.12%, and 82.65% for the estimated values ​​of the three parameters of frequency offset, time delay, and phase shift, respectively. It has the highest prediction accuracy among all experimental models and is R higher than the second best network ECAResnet in predicting the three multipath parameters.2 They are 0.49%, 1.07%, and 20.54% higher respectively. This is due to the fact that each layer of the network in the dense connection block connects and reuses all the feature images of the upper layer, which enhances the network's ability to represent image signal features. It is worth mentioning that compared with the residual network, the dense connection network shows better performance in estimating the three parameters. This is because compared with the superposition algorithm of residual connection, the dense connection concatenate can extract more original features from the image signal and has better generalization ability.

[0098] In addition, in order to verify the effectiveness of each part, an ablation experiment was introduced to remove the channel attention module and the long short-term memory network of the network of the present invention for multipath prediction. The experimental results show that when only the channel attention mechanism of the network is removed, the determination coefficients of the three multipath parameter predictions of Δτ, Δf, and Δθ decrease by 0.14%, 0.8%, and 8.22%, respectively. This is because the channel attention mechanism can give different weights to the channels according to the importance of the features, helping the network to focus on learning more meaningful channel features, which is beneficial to improving the accuracy of the network multipath prediction. When the channel attention mechanism and the long short-term memory network are removed at the same time, the R 2 They fell by 4.4%, 1.61% and 6.26% respectively.

[0099] It's worth noting that a network with 172 SE densely connected layers and a network with 124 SE densely connected layers achieve nearly identical estimation results in multipath detection. However, the 172-layer network has approximately 1.8 times the number of parameters as the 124-layer network, and its training time is significantly longer than that of the 124-layer network. This is because the detection task only requires outputting a final estimate, and the complexity of the feature extraction task is not very high. Therefore, a 124-layer network can be used for multipath detection. Compared with a 161-layer network, this network achieves faster training convergence and higher multipath parameter estimation accuracy.

[0100] Table 3 Comparison of determination coefficients of different network models in path detection

[0101]

[0102] In order to show the comparison results more intuitively, it will be drawn into a line graph, such as Figure 14 As shown, Figure 14 (a) is the comparison of Δτ correlation coefficient; Figure 14 (b) Δf correlation coefficient comparison; Figure 14 (c) is the comparison of Δθ correlation coefficient.

[0103] Structural Similarity (SSIM) is an indicator that constructs a functional relationship based on brightness, contrast, and structure to measure the similarity between two images. The relationship can be expressed as:

[0104] SIMM(x,y)=f(I(x,y),C(x,y),S(x,y)) (14)

[0105] SIMM(x,y)=[I(x,y)] α [C(x,y)] β [S(x,y)] γ (15)

[0106] Where x and y represent the two images being compared, I(x, y) represents the brightness contrast function using the mean grayscale value as the estimate, C(x, y) represents the contrast function using the standard deviation as the estimate, and S(x, y) represents the structural contrast function using the image divided by its own standard deviation as the estimate. α, β, and γ are constants greater than 0, and their values ​​are used to adjust the importance of each of the three dimensions. The specific calculation formulas for each contrast function are shown in Equations (16), (17), and (18):

[0107]

[0108] Where μ x 、μ y Represent the average values ​​of X and Y, σ x , σ y Indicates the standard deviation of X and Y, σ xy Represents the covariance of X and Y. C1, C2, and C3 are constants whose function is to prevent the denominator from being zero.

[0109] When calculating the structural similarity of two images, it is usually stipulated that α = β = γ = 1, C2 = 2C3, then Equation (15) can be calculated as:

[0110]

[0111] The value range of SIMM is [0, 1]. The closer the SIMM value is to 1, the more similar the two images are.

[0112] according to Figure 12In this process, multipath contamination removal is performed separately on the I and Q branch image signals. Taking the in-phase branch as an example, the three parameters of translation, delay, and phase shift are replicated two-dimensionally and expanded into three 80*80 data matrices. These are then stacked with the in-phase branch image containing multipath contamination as three additional feature channels and used as raw feature inputs into a deep learning network, trained on the uncontaminated direct signal. Because image reconstruction in multipath mitigation is more complex than parameter estimation in detection, increasing the depth of the network structure achieves better decontamination results.

[0113] During the multipath suppression training process, the relationship curve between the loss function of the output image and the training rounds is as follows: Figure 15 As shown in the figure; the relationship curve between the structural similarity (SIMM) of the image after removing multipath interference and the direct signal image and the training rounds is shown in the figure Figure 15 As shown in (a), the change of the loss function during the training process is as follows Figure 15 As shown in (b), the loss decreases and the structural similarity increases with the number of training rounds, demonstrating the effectiveness of the model. After approximately 50 iterations, the growth of structural similarity slows down, and the model begins to converge.

[0114] In order to demonstrate the advantages of the algorithm of the present invention, the current mainstream network of deep learning is selected for comparison with the network of the present invention. The SIMM indicators in the test set are shown in Table 4.

[0115] Table 4 Comparison of multipath suppression indicators in different networks

[0116] Network Model I branch SIMM Q branch SIMM EfficientNet 0.7583 0.7621 RegNet 0.7867 0.7646 DenseNet 0.9115 0.9272 ECAResNet 0.9535 0.9701 SDLNet(172) 0.9881 0.9946

[0117] Observation and comparison results show that the multipath-free image output by the network proposed in this invention has the highest SIMM value, both on the I branch and the Q branch. The average SIMM values ​​in the test set reached 98.81% and 99.46%, respectively. This shows that the reconstructed image has a high similarity with the interference-free direct signal image, proving the effectiveness of the constructed deep learning network. Compared with ECAResNet, the average SIMM value on the in-phase branch and the orthogonal branch increased by 3.46% and 2.45%, respectively. This is attributed to the splicing of the densely connected network to achieve the reuse of the original features. During the learning process, the network will inactivate some features. If an important high-dimensional feature can only be extracted by the later network layer, but the front network layer happens to abandon it when extracting low-dimensional features, it will greatly limit the maximum accuracy of the network prediction and suppression. Feature reuse is a good way to solve the above problems while also alleviating the gradient disappearance phenomenon. Residual networks add the original input information from the previous layer to the output of the current layer. Although this reuses the feature information of shallow layers, the addition process still changes the feature information to a certain extent, which limits the upper limit of network learning. Densely connected networks, on the other hand, parallelize upstream source information with network output features, preserving more original feature information during the dimension-by-dimension feature interpretation process. Therefore, compared to the feature reuse of residual networks, the feature reuse of densely connected networks is more effective in multipath suppression.

[0118] On the other hand, it also benefits from the introduction of the SE attention mechanism and long short-term memory network. During the feature learning process, SE attention can help the network focus on important channels in the image features from the global receptive field, accelerating the network convergence while also improving the network's inhibitory effect. The long short-term memory network can learn the temporal correlation in the signal, thereby improving the network's effectiveness. The fact that the network of the present invention is 7.66% and 6.74% higher than the SIMM of the ordinary densely connected network in the I and Q branches, respectively, further illustrates the effectiveness of the introduction of the attention mechanism and LSTM.

[0119] In order to more intuitively show the suppression effect, save the image after removing multipath interference, slice the signal at the same frequency perspective, and draw the images of the network output signal, direct signal and multipath signal in the same figure. Select three networks with good multipath suppression performance to suppress the same multipath signal respectively, and draw the images of the output signals for comparison. The output signal slices are as follows: Figure 16 As shown, Figure 16(a.1) SDLNet suppresses output 1; (a.2) SDLNet suppresses output 2; (b.1) ECAResNet suppresses output 1; (b.2) SDLNet suppresses output 2; (c.2) RegNet suppresses output 1; (c.3) RegNet suppresses output 2.

[0120] from Figure 16 It can be clearly seen that multipath significantly interferes with the direct signal, but the output signal of the algorithm presented in this paper almost overlaps with the direct signal, demonstrating the algorithm's excellent multipath mitigation capabilities. Furthermore, compared with other networks, the output signal curve has the highest degree of overlap with the direct signal curve, further demonstrating the algorithm's significant advantage in multipath mitigation.

[0121] In summary, the present invention, by constructing the SDLNet model, is able to achieve multipath detection and suppression of GNSS signals. In the multipath detection task, an SE-densely connected model with a 124-layer network architecture is used to estimate the time delay, frequency deviation, and phase shift of the multipath effect. The determination coefficients compared to the true values ​​reach 99.01%, 99.12%, and 82.65%, respectively. In the multipath suppression task, an SE-densely connected model with a 172-layer network architecture is used to deinterlace the multipath-interfered signal. The structural similarity (SIMM) between the deinterlaced I and Q signal images and the direct signal image reaches 98.81% and 99.46%, respectively. This model has higher prediction accuracy and better suppression effect than existing algorithms.

[0122] The present invention also provides a satellite multipath interference signal suppression system, which includes a signal conversion module, a basic model construction module, a signal processing model construction module and a signal interference removal module. The signal conversion module is used to obtain satellite time domain signals and convert them into multipath signal image matrices; the basic model construction module is used to add a channel attention mechanism SE module after each layer of the densely connected block of the densely connected network to obtain the SE densely connected network, and add a long short-term memory network LSTM after the SE densely connected network to obtain a constructed SDLNet model; the signal processing model construction module is used to obtain a first multipath signal sample image dataset and a second multipath signal sample image dataset, input the images in the first multipath signal sample image dataset into the SDLNet model, train the SDLNet model, and obtain a multipath interference signal detection model; input the images in the second multipath signal sample image dataset into the SDLNet model, train the SDLNet model, and obtain a multipath interference signal suppression model; the signal de-interference module is used to input the images in the multipath signal image matrix into the multipath interference signal detection model to obtain the frequency deviation value, time delay value, and phase shift value of the multipath interference signal; the frequency deviation value, time delay value, phase shift value, and multipath signal image are simultaneously input into the multipath interference signal suppression model to obtain a de-interferenced image.

[0123] The present invention also provides a computer device, comprising a memory and a processor; the memory stores a computer program, and the processor is used to run the computer program in the memory to execute the satellite multipath interference signal suppression method.

[0124] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor to execute the satellite multipath interference signal suppression method.

[0125] The above-described embodiments are only preferred specific implementation methods of the present invention, and the protection scope of the present invention is not limited thereto. Any simple changes or equivalent replacements of the technical solutions that can be obviously obtained by any technician familiar with the field within the technical scope disclosed in the present invention fall within the protection scope of the present invention.

Claims

1. A method for suppressing satellite multipath interference signals, characterized in that: The steps include: Obtain satellite time domain signals and convert them into multipath signal image matrix; A channel attention mechanism SE module is added after each layer of the dense connection block of the dense connection network to obtain the SE dense connection network. A long short-term memory network LSTM is added after the SE dense connection network to obtain the SDLNet model. Obtain a first multipath signal sample image dataset and a second multipath signal sample image dataset, input images in the first multipath signal sample image dataset into an SDLNet model, train the SDLNet model, and obtain a multipath interference signal detection model; input images in the second multipath signal sample image dataset into the SDLNet model, train the SDLNet model, and obtain a multipath interference signal suppression model; The image in the multipath signal image matrix is ​​input into the multipath interference signal detection model to obtain the frequency deviation value, time delay value and phase shift value of the multipath interference signal; the frequency deviation value, time delay value, phase shift value and the multipath signal image are simultaneously input into the multipath interference signal suppression model to obtain the image after interference removal.

2. The satellite multipath interference signal suppression method according to claim 1, characterized in that: The method of converting the satellite time domain signal into a multipath signal image matrix is ​​as follows: copy the one-dimensional satellite time domain signal to N t times, and then use the one-dimensional autocorrelation function to replicate the one-dimensional satellite time domain signal N times. f times, get two N t ×N f Matrix, and then two N t ×N f The matrix is ​​multiplied bit by bit to obtain a two-dimensional multipath signal image matrix, where N t N is the number of delay points to be observed. f is the number of frequency deviation points to be observed.

3. The satellite multipath interference signal suppression method according to claim 1, wherein: The SE module is used to perform squeezing and excitation operations on the input feature map, wherein the squeezing operation squeezes the feature map into a 1*1*C feature vector through a global pooling operation; the excitation operation is completed through two linear connection layers, the first linear connection layer reduces the feature vector of size 1*1*C to 1*1*C / redu, and is connected to the first linear connection layer through the ReLu activation function. The first linear connection layer restores the dimension of the input feature vector to 1*1*C, and outputs a weight matrix through the Sigmoid function, and then multiplies the weight matrix with the original feature matrix to obtain a reconstructed feature map.

4. The satellite multipath interference signal suppression method according to claim 1, wherein: The number of samples in the first multipath signal sample image data set is 5,000 groups; the number of samples in the second multipath signal sample image data set is 50,000 groups.

5. The satellite multipath interference signal suppression method according to claim 1, characterized in that: The performance of the multipath interference signal suppression model is evaluated by structural similarity.

6. The satellite multipath interference signal suppression method according to claim 1, characterized in that: The performance of the multipath interference signal detection model is evaluated by the coefficient of determination.

7. A satellite multipath interference signal suppression system, characterized in that: include: A signal conversion module is used to obtain satellite time domain signals and convert them into a multipath signal image matrix; The basic model construction module is used to add a channel attention mechanism SE module after each layer of the dense connection block of the dense connection network to obtain the SE dense connection network, and add a long short-term memory network LSTM after the SE dense connection network to obtain the constructed SDLNet model; a signal processing model construction module, configured to obtain a first multipath signal sample image dataset and a second multipath signal sample image dataset, input images from the first multipath signal sample image dataset into an SDLNet model, train the SDLNet model, and obtain a multipath interference signal detection model; and input images from the second multipath signal sample image dataset into the SDLNet model, train the SDLNet model, and obtain a multipath interference signal suppression model; The signal interference removal module is used to input the image in the multipath signal image matrix into the multipath interference signal detection model to obtain the frequency deviation value, time delay value, and phase shift value of the multipath interference signal; and simultaneously input the frequency deviation value, time delay value, phase shift value and the multipath signal image into the multipath interference signal suppression model to obtain the image after interference removal.

8. A computer device, characterized in that: It includes a memory and a processor; the memory stores a computer program, and the processor is used to run the computer program in the memory to execute the satellite multipath interference signal suppression method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor to execute the satellite multipath interference signal suppression method according to any one of claims 1 to 6.

Images