Non-line-of-sight signal discrimination method and system based on convolutional transformation learning model

Abstract

The present invention discloses a non-line-of-sight signal discrimination method and system based on a convolutional transform learning model. The method includes: performing parameter preprocessing on the collected satellite signals to obtain output satellite signals; constructing a convolutional transform learning model based on a non-convex log regularization operator; inputting the output satellite signals into the convolutional transform learning model for feature extraction to obtain sparse features of the satellite signals; and performing test analysis on the sparse features of the satellite signals based on a support vector machine to obtain line-of-sight signals and non-line-of-sight signals. The system includes: a collection module, a construction module, an extraction module, and a classification module. By using the present invention, the influence of non-line-of-sight signals in satellite signals on the positioning performance can be excluded and the positioning accuracy can be further improved. As a non-line-of-sight signal discrimination method and system based on a convolutional transform learning model, the present invention can be widely applied to the field of Beidou navigation satellite positioning technology.

Description

Technical Field

[0001] The present invention relates to the field of Beidou navigation satellite positioning technology, and in particular to a non-line-of-sight signal identification method and system based on a convolution transformation learning model. Background Art

[0002] The global satellite navigation system is used for various location-based services, such as navigation positioning, timing, geographic exploration, precision measurement and many other fields. At present, my country's Beidou system has been networked, and many other countries and regions are also constantly building and improving their own global satellite navigation systems in order to improve the service performance of various places. However, despite the increasing number of positioning satellites, there are still large positioning errors in the urban canyon environment formed by dense and tall buildings. This positioning error originates from multipath signals, a signal reflected or diffracted by buildings and other objects. Multipath signals are divided into two types: line-of-sight signals and non-line-of-sight signals. Their effects on positioning results are significantly different. At present, countermeasures for non-line-of-sight multipath errors can be divided into the following four categories. The first is to reject global positioning results containing non-line-of-sight signal multipath errors. It detects jumps in the global satellite navigation system positioning results and excludes them by combining sensor trajectories, but it reduces the availability of the global satellite navigation system and is impractical; secondly, non-line-of-sight signals are identified from global satellite navigation system observations and excluded from positioning calculations. Large ranging errors in global satellite navigation system positioning pseudoranges detected by receiver autonomous integrity monitoring related technologies , and exclude signals with large ranging errors. However, in this case, the number of non-line-of-sight signals in the received signals is very small, and it cannot be applied to environments with more non-line-of-sight signals. The next step is to estimate and correct the multipath error of the non-line-of-sight signal, and to correct the multipath error of the non-line-of-sight signal by establishing a 3D city map. However, the correction of the multipath error of the non-line-of-sight signal is very complex and difficult, and it is not realistic in terms of accuracy. The fourth is to suppress the reception of non-line-of-sight signals in the RF segment. Hardware suppression can simplify the subsequent software algorithm processing flow, but a lot of hardware equipment is required in the RF suppression signal, which will increase the system complexity and operating cost. Someone has proposed a non-line-of-sight signal detection method based on convolutional neural network. This study uses global satellite navigation system observation data for non-line-of-sight signal detection. Compared with traditional methods, this method uses convolutional neural network to learn signal features and improves classification performance, but does not use signal-related output. However, convolutional neural network requires a large amount of labeled data for learning to improve the accuracy of classification, resulting in low efficiency in practical applications. When diffraction signals or reflection signals (such as mirror reflection) near the edge of a building are received, it is difficult to correctly distinguish non-line-of-sight signals. Summary of the invention

[0003] To solve the above technical problems, the objective of the present invention is to provide a non-line-of-sight signal discrimination method and system based on a convolutional transform learning model, which can eliminate the influence of non-line-of-sight signals in satellite signals on the positioning performance and further improve the positioning accuracy.

[0004] The first technical solution adopted by the present invention is: a non-line-of-sight signal discrimination method based on a convolutional transform learning model, including the following steps:

[0005] Perform parameter preprocessing on the collected satellite signals to obtain output satellite signals;

[0006] Construct a convolutional transform learning model based on a non-convex log regularization operator;

[0007] Input the output satellite signals into the convolutional transform learning model for feature extraction to obtain sparse features of the satellite signals;

[0008] Perform test analysis on the sparse features of the satellite signals based on a support vector machine to obtain line-of-sight signals and non-line-of-sight signals.

[0009] Further, the step of performing parameter preprocessing on the collected satellite signals to obtain output satellite signals specifically includes:

[0010] Collect and decompose the satellite signals through the correlator of a global navigation satellite system receiver to obtain in-phase channel signals and quadrature channel signals;

[0011] Multiply the in-phase channel signals and the quadrature channel signals by their respective copy signals through a multiplier to obtain correlation signals;

[0012] Combine the correlation signals and perform a sum-of-squares calculation on the in-phase channel signals and the quadrature channel signals to obtain output satellite signals.

[0013] Further, the expressions of the in-phase channel signals and the quadrature channel signals are specifically as follows:

[0014]

[0015]

[0016] In the above formula, represents the Doppler estimation error, represents the phase estimation error, n I 、n Q represent two independent and identically distributed white noise components, T i represents the correlation integration time, τ represents the propagation delay, I represents the in-phase channel signal, Q represents the quadrature channel signal, and M represents the correlation coefficient proportional to the propagation delay error.

[0017] Further, the step of constructing a convolutional transform learning model based on a non-convex log regularization operator specifically includes:

[0018] Introduce a non-convex log regularization operator as a sparse constraint condition;

[0019] Based on the sparse constraint condition, learn a set of convolutional kernels, extract sparse features from the input data, and construct a convolutional transform optimization function based on log regularization according to the set of convolutional kernels and the sparse features;

[0020] Solve the convolutional transform optimization function based on log regularization through the proximal convex difference method to obtain a solution result;

[0021] According to the solution result, adopt an alternating update strategy to alternately iteratively optimize the set of convolutional kernels and the sparse features, and construct a convolutional transform learning model.

[0022] Further, the expression of the convolutional transform optimization function based on log regularization is specifically as follows:

[0023]

[0024] In the above formula, x represents the input data, D represents the set of convolutional kernels, D m represents the set of convolutional kernels in the m-th layer of the model, d k represents the k-th convolutional kernel in the set of convolutional kernels, z k represents the set of sparse features obtained by the k-th convolutional kernel, F represents the convolutional transform optimization function based on log regularization, K represents the number of convolutional kernels in the set of convolutional kernels, J D (D) represents the diversity regularization function of the set of convolutional kernels, J z (·) represents the log sparse regularization function.

[0025] Further, the convolutional transform optimization function based on log regularization includes a non-convex part and a convex part. The step of solving the convolutional transform optimization function based on log regularization through the proximal convex difference method to obtain a solution result specifically includes:

[0026] Perform differential decomposition on the non-convex part of the convolutional transform optimization function based on log regularization to obtain a decomposition result;

[0027] According to the decomposition result, perform proximal gradient descent solution on the convex part of the convolutional transform optimization function based on log regularization to obtain a solution result.

[0028] Further, the step of inputting the output satellite signal into the convolutional transform learning model for feature extraction to obtain the sparse features of the satellite signal specifically includes:

[0029] Input the output satellite signal into the convolutional transformation learning model, where the convolutional transformation learning model includes a sparse feature module and a convolutional kernel module;

[0030] Based on the convolutional kernel module, perform feature extraction processing on the output satellite signal to obtain the feature information of the output satellite signal;

[0031] Based on the sparse feature module, optimize and train the feature information of the output satellite signal to obtain the sparse feature of the satellite signal.

[0032] Furthermore, it further includes iteratively updating the convolutional transformation learning model according to the sparse feature of the satellite signal, which specifically includes:

[0033] Iteratively update the sparse feature module and the convolutional kernel module according to the sparse feature of the satellite signal;

[0034] Until both the sparse feature module and the convolutional kernel module meet the convergence condition, stop the update to obtain the final convolutional transformation learning model;

[0035] Perform feature extraction processing on the satellite signal according to the final convolutional transformation learning model.

[0036] Furthermore, the expressions for updating the sparse feature module and the convolutional kernel module are as follows:

[0037]

[0038]

[0039] In the above formula, represents the updated sparse feature module, represents the updated convolutional kernel module, represents the sparse feature module at the t-th iterative update, represents the convolutional kernel module at the t-th iterative update, η z 、η d represent the gradient descent step size, represents the gradient of the continuously smooth part in the convex function, J D represents the orthogonal projection constraint function, represents the first closed convex function among the two closed convex functions obtained by using the convex difference algorithm.

[0040] The second technical solution adopted by the present invention is: a non-line-of-sight signal discrimination system based on a convolutional transformation learning model, including:

[0041] An acquisition module, configured to perform parameter preprocessing on the acquired satellite signal to obtain an output satellite signal;

[0042] A construction module constructs a convolutional transform learning model based on a non-convex log regularization operator;

[0043] An extraction module is used to input the output satellite signal into the convolutional transform learning model for feature extraction to obtain the sparse features of the satellite signal;

[0044] A classification module performs test analysis on the sparse features of the satellite signal based on a support vector machine to obtain line-of-sight signals and non-line-of-sight signals.

[0045] The beneficial effects of the method and system of the present invention are as follows: By inputting the acquired satellite signal into a global satellite navigation system signal correlator, decomposing and squaring and summing the satellite signal as the input data of the convolutional transform learning model, the quality of the output result of the convolutional transform learning model is strengthened. The convolutional transform learning model is constructed based on log regularization and proximal convex difference, can learn the satellite signal features, and uses SVM to classify signals, excluding the influence of non-line-of-sight signals on the positioning performance, further improving the positioning accuracy, and achieving the goal of accurate positioning of the global satellite navigation system receiver in complex urban canyons. Description of the Drawings

[0046] Figure 1 is a flowchart of the steps of the non-line-of-sight signal identification method based on the convolutional transform learning model of the present invention;

[0047] Figure 2 is a structural block diagram of the non-line-of-sight signal identification system based on the convolutional transform learning model of the present invention;

[0048] Figure 3 is a schematic diagram of the specific implementation process of signal processing according to the convolutional transform learning model of the present invention;

[0049] Figure 4 is a schematic diagram of the structure of the convolutional transform learning model constructed based on the non-convex log regularization operator of the present invention. Detailed Embodiments

[0050] The following further describes the present invention in detail with reference to the drawings and specific embodiments. For the step numbers in the following embodiments, they are only set for the convenience of description and explanation, and no limitation is imposed on the order between the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0051] Referring to Figure 1 and Figure 3 , the present invention provides a non-line-of-sight signal identification method based on a convolutional transform learning model, and the method includes the following steps:

[0052] S1. Perform parameter preprocessing on the collected satellite signal to obtain an output satellite signal;

[0053] Specifically, the signal correlator described in the present invention is located inside a global satellite navigation system receiver. Therefore, the operation of the receiver includes reading signals and solving correlations. First, the receiver reads the already obtained Doppler frequency shift, code delay, and the original signal, and divides the signal into an in-phase signal channel and a quadrature signal channel. The expressions of the in-phase channel signal and the quadrature channel signal are specifically as follows:

[0054]

[0055]

[0056] In the above formula, represents the Doppler estimation error, represents the phase estimation error, n I 、n Q represent two independent and identically distributed white noise components, T i represents the correlation integration time, τ represents the propagation delay, I represents the in-phase channel signal, Q represents the quadrature channel signal, and M represents the correlation coefficient proportional to the propagation delay error;

[0057] Among them,

[0058]

[0059] In the above formula, D represents the value of the navigation information bit that remains unchanged during the correlation integration time T i , represents the autocorrelation function of the PRN code, represents the local estimation of the propagation delay τ;

[0060] Furthermore, through the internal multiplier hardware, the signals of the two channels are correlated with the corresponding copy signals of the receiver itself to obtain the correlation output. Subsequently, the correlations of the two signals are integrated and dumped. Taking the Doppler frequency shift error amount and the propagation delay error solved for each signal as two orthogonal axes, take I(t) 2 +Q(t) 2 as the output of the correlator, and finally complete the conversion from the correlator output to the input of the convolutional transform model, that is, obtain the output satellite signal.

[0061] S2. Construct a convolutional transform learning model based on a non-convex log regularization operator, and input the output satellite signal into the convolutional transform learning model for feature extraction to obtain the sparse features of the satellite signal;

[0062] Specifically, refer to Figure 4, which is the network structure diagram of the learning model based on log regularization and proximal convex difference convolution transform of the present invention. This network mainly includes a plurality of convolution modules and sparse constraint modules with the same structure. In this embodiment, it is composed of two layers of convolution modules and sparse constraint modules with the same structure alternately combined. The convolution module in the first layer extracts convolution features from all input data, and the sparse constraint module optimizes and trains the data with preliminarily extracted features to obtain sparser features that more strictly meet the requirements. The sparser features extracted from the previous layer are used as the input signal of the next layer, and the sparser features of the previous layer are sparsely encoded through the convolution module and the sparse constraint module, and finally sparser features with more discriminability and richness are extracted;

[0063] Furthermore, describe the sparse constraint function in the convolution transform learning model, introduce a non-convex log regularization operator as the sparse constraint function, and the expression of the sparse constraint function is as follows:

[0064]

[0065] In the above formula, λ represents the sparse coefficient, δ represents the log constant, α represents the input vector, and α i represents the i-th element of the input vector, R represents the vector space, and N represents the total number of elements contained in the input vector;

[0066] Construct a sparse constraint function based on log regularization, and at the same time initialize the convolution kernel D: = [d 1 …d k , sparse features {z k} ∈ R n . For the non-convex optimization problem of the regularization constraint function, convert the objective function into a convex optimization part and a non-convex optimization part, decompose the non-convex optimization part into the difference of two convex functions through the convex difference algorithm, and finally convert it into an optimization problem of the combination of three functions, that is, the convolution transform optimization function based on log regularization. The expression of the convolution transform optimization function based on log regularization is specifically as follows:

[0067]

[0068] In the above formula, x represents the input data, D represents the convolution kernel group, D m represents the convolution kernel group of the m-th layer in the model, d k represents the k-th convolution kernel in the convolution kernel group, z k represents the set of sparse features obtained by the k-th convolution kernel, F represents the convolution transform optimization function based on log regularization, K represents the number of convolution kernels in the convolution kernel group, J D (D) represents the diversity regularization function of the convolution kernel group, and J z (·) represents the log sparse regularization function;

[0069] After being transformed into a convex optimization problem, the solution still faces the problem that the function is not fully differentiable or non-differentiable. Therefore, the proximal gradient method is adopted to replace the gradient at the non-fully differentiable or non-differentiable part, which is equivalent to using the proximal projection operator as an approximate gradient for gradient descent. Since the two variables do not affect each other, the alternating optimization method is used for iterative update;

[0070] The expression for convex difference decomposition of the non-convex function in the optimization function is specifically as follows:

[0071] J(α) = J 1 (α) - J 2 (α)

[0072] In the above formula, J 1 (α) and J 2 (α) both represent closed convex functions in R n ;

[0073] The expression for solving the convex function by proximal gradient descent is as follows:

[0074]

[0075] In the above formula, η represents the descent step size, and h(α) represents the differentiable part in the convex function;

[0076] Furthermore, the expressions for updating the sparse feature module and the convolution kernel module are as follows:

[0077]

[0078]

[0079] In the above formula, represents the updated sparse feature module, represents the updated convolution kernel module, represents the sparse feature module at the t-th iterative update, represents the convolution kernel module at the t-th iterative update, η z 、η d represent the gradient descent step sizes, represents the gradient of the continuously smooth part in the convex function, J D represents the orthogonal projection constraint function, represents the first closed convex function among the two closed convex functions obtained by the convex difference algorithm;

[0080] It should be noted that the convolution kernels learned by the convolution transformation learning model through an unsupervised method do not have the problem of relying on data labels, overcoming the bottleneck when facing unlabeled or sparsely labeled data. At the same time, the diversity can be constrained through the regularization of the convolution module, avoiding the extraction of similar and redundant features for the same signal.

[0081] S3. Test and analyze the sparse features of satellite signals based on a support vector machine to obtain line-of-sight signals and non-line-of-sight signals.

[0082] Specifically, according to the learned features, train a support vector machine. The training data is divided into two categories: line-of-sight signals and non-line-of-sight signals. The specific expression of the training is as follows:

[0083]

[0084] In the above formula, y i represents the data label, w represents the scaling coefficient, and X i represents the training sample set. Among them, X i = [x 1 , x 2 , …, x n T ,

[0085] That is, test and analyze the sparse features of satellite signals based on a support vector machine. Randomly and evenly divide the sparse features into a training set and a test set. Obtain the classification results of line-of-sight signals and non-line-of-sight signals through training and testing, and exclude non-line-of-sight signals from the positioning calculation.

[0086] Referring to Figure 2 , the non-line-of-sight signal discrimination system based on the convolution transformation learning model includes:

[0087] An acquisition module for preprocessing the parameters of the acquired satellite signals to obtain output satellite signals;

[0088] A construction module for constructing a convolution transformation learning model based on a non-convex log regularization operator;

[0089] An extraction module for inputting the output satellite signals into the convolution transformation learning model for feature extraction to obtain the sparse features of the satellite signals;

[0090] A classification module for testing and analyzing the sparse features of satellite signals based on a support vector machine to obtain line-of-sight signals and non-line-of-sight signals.

[0091] The content in the above method embodiments is applicable to the system embodiments. The functions specifically implemented by the system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those of the above method embodiments. ​

[0092] The above is a specific description of the preferred embodiment of the present invention. However, the present invention is not limited to the described embodiment. Those skilled in the art can make various equivalent deformations or substitutions without departing from the spirit of the present invention, and these equivalent deformations or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A method for identifying non-line-of-sight signals based on a convolutional transform learning model, characterized in that, it includes the following steps: Perform parameter preprocessing on the collected satellite signals to obtain output satellite signals; Construct a convolutional transform learning model based on a non-convex log regularization operator; Input the output satellite signals into the convolutional transform learning model for feature extraction to obtain sparse features of the satellite signals; Based on a support vector machine, conduct test analysis on the sparse features of the satellite signals to obtain line-of-sight signals and non-line-of-sight signals; The step of performing parameter preprocessing on the collected satellite signals to obtain output satellite signals specifically includes: Collect and decompose satellite signals through the correlator of a global satellite navigation system receiver to obtain in-phase channel signals and quadrature channel signals; Multiply the in-phase channel signals and the quadrature channel signals by their respective corresponding replica signals through a multiplier to obtain correlation signals; Combine the correlation signals and perform a sum-of-squares calculation on the in-phase channel signals and the quadrature channel signals to obtain output satellite signals; The expressions of the in-phase channel signals and the quadrature channel signals are specifically as follows: In the above formula, represents the Doppler estimation error, represents the phase estimation error, and n I , n Q represent two independent and identically distributed white noise components, T i represents the correlation integration time, τ represents the propagation delay, I represents the in-phase channel signal, Q represents the quadrature channel signal, and M represents the correlation coefficient proportional to the propagation delay error.

2. The method for identifying non-line-of-sight signals based on a convolutional transform learning model according to claim 1, characterized in that, The step of constructing a convolutional transform learning model based on a non-convex log regularization operator specifically includes: Introduce a non-convex log regularization operator as a sparse constraint condition; Based on the sparse constraint condition, learn a set of convolutional kernels, perform sparse feature extraction on the input data, and construct a convolutional transform optimization function based on log regularization according to the set of convolutional kernels and the sparse features; Solve the convolutional transform optimization function based on log regularization through the proximal convex difference method to obtain a solution result; According to the solution result, adopt an alternating update strategy to alternately iteratively optimize the set of convolutional kernels and the sparse features, and construct a convolutional transform learning model.

3. The method for identifying non-line-of-sight signals based on a convolutional transform learning model according to claim 2, characterized in that, The expression of the convolutional transform optimization function based on log regularization is specifically as follows: In the above formula, x represents the input data, D represents the convolution kernel group, D m represents the convolution kernel group of the m-th layer in the model, d k represents the k-th convolution kernel in the convolution kernel group, z k represents the sparse feature set obtained by the k-th convolution kernel, F represents the convolution transform optimization function based on log regularization, K represents the number of convolution kernels in the convolution kernel group, J D (D) represents the diversity regularization function of the convolution kernel group, J z (·) represents the log sparse regularization function.

4. The method for identifying non-line-of-sight signals based on a convolutional transform learning model according to claim 3, characterized in that, The convolutional transform optimization function based on log regularization includes a non-convex part and a convex part. The step of solving the convolutional transform optimization function based on log regularization through the proximal convex difference method to obtain a solution result specifically includes: Perform differential decomposition on the non-convex part of the convolutional transform optimization function based on log regularization to obtain a decomposition result; According to the decomposition result, perform proximal gradient descent solution on the convex part of the convolutional transform optimization function based on log regularization to obtain a solution result.

5. The method for identifying non-line-of-sight signals based on a convolutional transform learning model according to claim 4, characterized in that, The step of inputting the output satellite signals into the convolutional transform learning model for feature extraction to obtain sparse features of the satellite signals specifically includes: Input the output satellite signals into the convolutional transform learning model, and the convolutional transform learning model includes a sparse feature module and a convolutional kernel module; Perform feature extraction processing on the output satellite signal based on the convolution kernel module to obtain the feature information of the output satellite signal; Perform optimization training on the feature information of the output satellite signal based on the sparse feature module to obtain the sparse features of the satellite signal.

6. The non-line-of-sight signal discrimination method based on the convolution transform learning model according to claim 5, wherein, it further includes iteratively updating the convolution transform learning model according to the sparse features of the satellite signal, which specifically includes: Iteratively update the sparse feature module and the convolution kernel module according to the sparse features of the satellite signal; Until both the sparse feature module and the convolution kernel module meet the convergence condition, stop the update to obtain the final convolution transform learning model; Perform feature extraction processing on the satellite signal according to the final convolution transform learning model.

7. The non-line-of-sight signal discrimination method based on the convolution transform learning model according to claim 6, wherein, The expressions for updating the sparse feature module and the convolution kernel module are as follows: In the above formula, represents the updated sparse feature module, represents the updated convolutional kernel module, represents the sparse feature module at the t-th iterative update, represents the convolutional kernel module at the t-th iterative update, η z 、η d represent the gradient descent step size, represents the gradient of the continuously smooth part in the convex function, J D represents the orthogonal projection constraint function, represents the first closed convex function among the two closed convex functions obtained by using the convex difference algorithm.

8. A non-line-of-sight signal discrimination system based on a convolution transform learning model, wherein, used to execute the non-line-of-sight signal discrimination method based on the convolution transform learning model according to claim 1, including the following modules: An acquisition module for performing parameter preprocessing on the acquired satellite signal to obtain an output satellite signal; A construction module for constructing a convolution transform learning model based on a non-convex log regularization operator; An extraction module for inputting the output satellite signal into the convolution transform learning model for feature extraction to obtain the sparse features of the satellite signal; A classification module for performing test analysis on the sparse features of the satellite signal based on a support vector machine to obtain line-of-sight signals and non-line-of-sight signals.

Images