A non-line-of-sight signal recognition method based on wavelet-Gram convolutional neural network

Abstract

The present invention relates to a non-line-of-sight signal recognition method based on a wavelet-Gram convolutional neural network. The method comprises the following steps: after obtaining channel impulse response (CIR) data, using discrete wavelet transform to take the low-frequency part with more obvious characteristics in the original CIR signal as the effective part, thereby improving the speed and efficiency of subsequent non-line-of-sight signal recognition training; then normalizing the extracted low-frequency part so that it is located in the (-1, 1) interval and performing Gram angle field transform to supplement the time characteristics of the signal without changing the original sequence of the signal; finally, putting the data set after the Gram angle field transform into a convolutional neural network for classification training, and testing the network training effect with a test set, thereby improving the ability of non-line-of-sight signal recognition.

Description

Technical Field

[0001] The invention belongs to the technical field of wireless positioning and relates to a non-line-of-sight signal recognition method based on wavelet convolutional neural network. Background Art

[0002] With the rapid development of the national economy, positioning technology has matured. Satellite navigation-based services are widely used in outdoor environments. In recent years, location-based service technologies have also been evolving towards indoor positioning systems. With the continuous advancement of modern society and the increasing number of large buildings, people spend 70% of their time indoors, resulting in a growing demand for indoor location-based services. For example, accurate indoor location information is required for locating special populations, office space management, and logistics warehousing. High-precision indoor positioning holds enormous potential for application in emergency scenarios, such as firefighting, public safety, and earthquake relief. Among current wireless positioning technologies, ultra-wideband positioning technology, unlike traditional communication technologies, does not utilize the carrier wave of conventional communication systems. Instead, it transmits data by sending and receiving extremely narrow pulses in the nanosecond or sub-nanosecond range, making it a leading indoor positioning technology.

[0003] In ultra-wideband indoor positioning systems, various obstacles indoors can block ultra-wideband signals, causing non-line-of-sight (NLOS) propagation. This NLOS propagation reduces the reliability of communication and positioning accuracy. Therefore, accurately identifying NLOS signals during positioning and taking appropriate measures can improve positioning accuracy. Traditional NLOS detection techniques can be categorized into three types: distance estimation-based methods, channel impulse response-based methods, and position estimation-based methods. Distance-based methods primarily distinguish between line-of-sight (LOS) and non-line-of-sight (NLOS) using probability density functions or the variance of range estimates. While seemingly simple, these methods often suffer from poor positioning performance due to limitations in prior distribution functions or time delays. Position-based methods identify NLOS during the position estimation process, and can even use the calculated position coordinates sequentially for NLOS detection. These methods can detect NLOS by comparing position estimates generated from different subsets of distance estimates when redundant distance estimates are available, but are ineffective when redundant distance estimates are unavailable. The method based on channel impulse response is widely used for NLOS signal identification because it can make full use of the statistical information of the received multipath components and use channel parameters to determine whether the received signal is an NLOS signal or a LOS signal. Summary of the Invention

[0004] The present invention addresses the problem that traditional methods based on channel impulse response are unable to automatically extract effective channel features, and proposes a non-line-of-sight signal recognition method based on a wavelet-Gram convolutional neural network. This method is based on the ability of the channel impulse response (CIR) to reflect channel characteristics, and utilizes the ability of the convolutional neural network to automatically extract features to classify and identify non-line-of-sight signals. However, since the convolutional neural network cannot reflect the time relationship of the signal, the Gram angular field is used to convert the one-dimensional time signal into a two-dimensional image, and the time relationship of the signal is retained. A discrete wavelet transform is performed before the Gram angular field transform to extract the low-frequency part with more obvious features, thereby improving the rate and efficiency of signal feature extraction, thereby achieving the purpose of improving non-line-of-sight signal recognition.

[0005] In order to achieve the above object, the present invention provides a non-line-of-sight signal recognition method based on a wavelet-Gram convolutional neural network, comprising the following steps:

[0006] (1) Obtaining training samples: Collect N channel impulse response samples with labeled signal categories as the training set S0 = {(x1(t), y1), (x2(t), y2), ..., (x N (t),y N )}; where x n (t)(n=1,2,…,N) represents the nth sample in the training set, which is a 1×T vector, where T represents the time dimension of the sample; y n Represents sample x n (t) label, y n ∈{0,1};

[0007] (2) Extracting the effective features of the original CIR signal: Performing discrete wavelet transform on each sample in the extracted training set S0, taking the low-frequency part with more obvious features as the effective part, and constructing a new training set in, Represents the effective part of the CIR signal after discrete wavelet transform, which is the nth sample of the training set S1. vector, T represents the time dimension of the original CIR signal; y n is a label, which is the same as the label in step (1);

[0008] (3) Feature preprocessing: Normalize each sample in the training set S1 so that it is in the interval (-1, 1) to satisfy the Gram angular field transform condition;

[0009] (IV) Converting one-dimensional signal to two-dimensional image: Using Gram angular field transform, each one-dimensional signal sample in the training set S1 is converted into two two-dimensional images containing temporal relationship features, namely Gram angular sum field (GASF) and Gram angular difference field (GADF), to construct a new training set in, Represents the two-dimensional image after Gram angular field transformation, which is the nth sample of the training set S. , T represents the time dimension of the original CIR signal; y n is a label, which is the same as the label in step (1);

[0010] (5) Convolutional neural network classification: The training set S is put into the convolutional neural network for classification training, and the test set is used to test the network training effect.

[0011] Furthermore, in the step (1), N channel impulse response samples with signal category labels are collected as a training set S0 = {(x1(t), y1), (x2(t), y2), ..., (x N (t),y N )}; where x n (t)(n=1,2,…,N) represents the nth sample in the training set, which is a 1×T vector, where T represents the time dimension of the sample; y n Represents sample x n (t) label, y n ∈{0,1};

[0012] Furthermore, in the step (ii), each sample in the extracted training set S0 is subjected to discrete wavelet transform, and the bior3.1 wavelet is used to perform discrete wavelet transform, and the original CIR signal is divided into a high-frequency part and a low-frequency part. After analysis, it is found that the low-frequency part after the transform contains more features than the original CIR signal and the decomposed low-frequency part. Therefore, the low-frequency part after discrete wavelet transform is continued to be processed as the effective part. Since the discrete wavelet transform adopts a downsampling process, the length of the low-frequency part after decomposition is half of the original CIR signal, and the length becomes Where T is the time dimension of the original CIR signal, and a new training set is constructed in, Represents the effective part of the CIR signal after discrete wavelet transform, which is the nth sample of the training set S1. vector, T represents the time dimension of the original CIR signal; y n is a label, which is the same as the label in step (1);

[0013] Furthermore, in step (3), the sample vectors in the training set S1 are normalized to reduce the low-frequency part after discrete wavelet transform to the range of -1 to 1 so as to meet the conditions of Gram angular field transform. The normalization formula is:

[0014]

[0015] Where X is the original time series data, X min is the minimum value in the original time series data, X max This is the maximum value in the original time series data. This X is also used when the test set samples are standardized. min With X max If the value is greater than 1 after normalization, it is set to 1; if the value is less than -1, it is set to -1;

[0016] Furthermore, in step (iv), each one-dimensional signal sample in the training set S1 is converted into two two-dimensional images containing temporal relationship features, namely, the Gram Angular Sum Field (GASF) and the Gram Angular Difference Field (GADF). The Gram Angular Field Transformation method is as follows:

[0017] (1) Standardize the one-dimensional time signal and scale it to the (-1, 1) interval. The standardization process has been completed in step (3), so there is no need to repeat the operation.

[0018] (2) The normalized one-dimensional time signal is converted from rectangular coordinates to polar coordinates. The conversion formula is:

[0019]

[0020]

[0021] Where i represents the timestamp of the one-dimensional time signal, represents the signal amplitude at the i-th time stamp, θ i Indicates the angle of the signal at the i-th time stamp converted to polar coordinates, R i It represents the radius of the polar coordinate representation of the signal at the i-th timestamp, and N represents the signal length of the one-dimensional time signal.

[0022] (3) In the principle of Gram matrix, the cosine function is used to represent the correlation between two vectors. Therefore, the Gram angle field uses the cosine function to represent the correlation between two angles, which is the Gram angle sum field. The sin function is used to represent the difference between the two angles, which is the Gram angle difference field. The transformation formula is:

[0023]

[0024]

[0025] Where GASF represents the Gram angle sum field matrix, and GADF represents the Gram angle difference field matrix.

[0026] Package GASF and GADF to construct a new training set in, Represents the two-dimensional image after Gram angular field transformation, which is the nth sample of the training set S. , T represents the time dimension of the original CIR signal; y n is a label, which is the same as the label in step (1);

[0027] Furthermore, in step (5), the training set S is put into the convolutional neural network for classification training, and the network training effect is tested with the test set.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] The present invention provides a non-line-of-sight signal recognition method based on a wavelet-Gram convolutional neural network. Based on the channel impulse response (CIR), the method utilizes the convolutional neural network's ability to automatically extract features. The method extracts channel features reflected in the CIR signal to perform non-line-of-sight signal recognition. Since conventional convolutional neural networks cannot reflect the temporal relationship characteristics of signals, a Gram angle field is used to convert a one-dimensional time signal into a two-dimensional image. This method not only preserves the original signal information but also reflects the correlation between each signal time stamp through the Gram angle sum field and the Gram angle difference field, supplementing the signal's temporal features for convolutional neural network classification. Furthermore, a discrete wavelet transform is added to this method. The discrete wavelet transform is used to extract the more prominent low-frequency components and remove some invalid features, thereby improving the efficiency of convolutional neural network training. Furthermore, due to the downsampling process of the discrete wavelet transform, the original CIR signal is shortened by half, increasing the speed of convolutional neural network training. Furthermore, because the temporal features are preserved, the purpose of improving non-line-of-sight signal recognition is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a flow chart of the non-line-of-sight signal recognition method based on the Wavelet-Gram convolutional neural network of the present invention; DETAILED DESCRIPTION

[0031] The present invention is described in detail below by way of exemplary embodiments, but it should be understood that elements, structures, and features of one embodiment may be beneficially combined in other embodiments without further description.

[0032] See also Figure 1 The present invention discloses a non-line-of-sight signal recognition method based on a wavelet-Gram convolutional neural network, comprising the following steps:

[0033] (1) Obtaining training samples: Collect N channel impulse response samples with labeled signal categories as the training set S0 = {(x1(t), y1), (x2(t), y2), ..., (x N (t),yN )}; where x n (t)(n=1,2,…,N) represents the nth sample in the training set, which is a 1×T vector, where T represents the time dimension of the sample; y n Represents sample x n (t) label, y n ∈{0,1}; For example, the DWM1000 module can be used for data acquisition, with a pulse repetition frequency of 64 MHz, and each data has 1016 samples, and the time resolution is about half a period of the 499.2 MHz fundamental frequency;

[0034] (2) Extracting the effective features of the original CIR signal: Performing discrete wavelet transform on each sample in the extracted training set S0, taking the low-frequency part with more obvious features as the effective part, and constructing a new training set in, Represents the effective part of the CIR signal after discrete wavelet transform, which is the nth sample of the training set S1. vector, T represents the time dimension of the original CIR signal; y n is a label, which is the same as the label in (1);

[0035] (III) Feature preprocessing: The sample vectors in the training set S1 are normalized to reduce the low-frequency part after discrete wavelet transform to the range of -1 to 1 so as to meet the conditions of Gram angular field transform. The normalization formula is:

[0036]

[0037] Where X is the original time series data, X min is the minimum value in the original time series data, X max This is the maximum value in the original time series data. This X is also used when the test set samples are standardized. min With X max If the value is greater than 1 after normalization, it is set to 1; if the value is less than -1, it is set to -1;

[0038] (IV) Converting one-dimensional signals into two-dimensional images: Each one-dimensional signal sample in the training set S1 is converted into two two-dimensional images containing temporal relationship features using the Gram angular sum field (GASF) and the Gram angular difference field (GADF). The Gram angular field transform method is as follows:

[0039] (1) Standardize the one-dimensional time signal and scale it to the (-1, 1) interval. The standardization process has been completed in step (3), so there is no need to repeat the operation.

[0040] (2) The normalized one-dimensional time signal is converted from rectangular coordinates to polar coordinates. The conversion formula is:

[0041] θ i =arccos(y i ) (7)

[0042]

[0043] Where i represents the timestamp of the one-dimensional time signal, θ i Indicates the angle of the signal at the i-th time stamp converted to polar coordinates, R i It represents the radius of the polar coordinate representation of the signal at the i-th timestamp, and N represents the signal length of the one-dimensional time signal.

[0044] (3) In the principle of Gram matrix, the cosine function is used to represent the correlation between two vectors. Therefore, the Gram angle field uses the cosine function to represent the correlation between two angles, which is the Gram angle sum field. The sin function is used to represent the difference between the two angles, which is the Gram angle difference field. The transformation formula is:

[0045]

[0046]

[0047] Where GASF represents the Gram angle sum field matrix, and GADF represents the Gram angle difference field matrix.

[0048] Package GASF and GADF to construct a new training set in, Represents the two-dimensional image after Gram angular field transformation, which is the nth sample of the training set S. , T represents the time dimension of the original CIR signal; y n is a label, which is the same as the label in step (1);

[0049] (5) Convolutional neural network classification: The training set S is put into the convolutional neural network for classification training, and the test set is used to test the network training effect.

[0050] The non-line-of-sight signal recognition method of the present invention first uses discrete wavelet transform to extract the low-frequency part of the channel impulse response (CIR) signal with more obvious characteristics, thereby improving the speed and efficiency of non-line-of-sight signal recognition. The extracted low-frequency signal with more obvious characteristics is then normalized to reduce the signal amplitude to the interval (-1, 1) to meet the Gram angle field transform conditions. Then, a Gram angle field transform is performed to retain the time relationship between the original signal sequence and each timestamp in the Gram angle sum field (GASF) and the Gram angle difference field (GADF), thereby supplementing the time features for subsequent training. The GASF and GADF are then put into a convolutional neural network as two channels for non-line-of-sight signal recognition, and the recognition effect is tested using a test set to obtain a non-line-of-sight signal recognition result.

[0051] The non-line-of-sight signal recognition method of the present invention uses discrete wavelet transform and Gram angle field to help convolutional neural network extract features and recognize non-line-of-sight signals, thereby improving the performance and efficiency of recognizing non-line-of-sight signals.

[0052] In order to more clearly illustrate the beneficial effects of the non-line-of-sight signal recognition method of the present invention, the non-line-of-sight signal recognition method of the present invention is further described below with reference to embodiments.

[0053] Example:

[0054] The data used in this example comes from the elastic Wireless Networking Experimentation (eWINE), a subproject of the EU's Horizon 2020 research program. The program leverages artificial intelligence to address current challenges in dense and dynamic wireless networks. The data was measured using a DW1000 ultra-wideband transceiver chip, configured with four channels (500MHz bandwidth, 3993.6MHz center frequency), a data length of 1024, and a data rate of 110kb / s. The dataset contains 42,000 samples across seven different measurement scenarios: Office Environment 1, Office Environment 2, Apartment, Workshop, Living Room, Bedroom, and Boiler Room. To prevent overfitting in specific environments, strict node positioning was not performed during the measurement process. Instead, a UWB node was placed at a random fixed location within the measurement environment, and measurements were performed while the other UWB node was moved throughout the environment. For each environment, 3,000 measurements were collected under line-of-sight (LOS) channel conditions, followed by 3,000 measurements under non-line-of-sight (NLOS) channel conditions within the same environment. This example analyzes 6,000 samples from the dataset, including 5,000 training samples and 1,000 test samples. Each row in the dataset is 200 characters long and represents the characteristic CIR700-CIR900 signal range, totaling 200 channel impulse response data points.

[0055] Convolutional neural networks can effectively and automatically extract features, but they cannot extract the positional relationship between each feature. For the channel impulse response (CIR) signal, this is the time feature contained in the signal. As a time series signal, the time feature of the CIR signal is also very important for non-line-of-sight signal recognition. The convolutional neural network's neglect of time features will affect the accuracy of non-line-of-sight signal recognition. Considering this problem, this patent introduces the wavelet Gram, which utilizes the characteristics of the Gram angle field to not only preserve the original signal sequence, but also supplement the time feature representation of the signal in the Gram angle sum field (GASF) and the Gram angle difference field (GADF). It also introduces the discrete wavelet transform to extract the low-frequency part with more obvious features. Due to downsampling, the signal length is shortened to half of the original length. The convolutional neural network then classifies and recognizes the non-line-of-sight signal. The addition of the wavelet Gram improves the efficiency and speed of non-line-of-sight signal recognition, and further improves the ability of non-line-of-sight signal recognition.

[0056] In this example, Method 1: Only convolutional neural network is used to classify non-line-of-sight signals, and the results of ten trainings are taken and the average value is calculated; Method 2: The method proposed in this patent is used to identify non-line-of-sight signals, and the results of ten trainings are taken and the average value is calculated. The recognition accuracy is shown in Table 1.

[0057] Table 1

[0058] method Dharma One Method 2 Recognition rate 78.11％ 81.34％

[0059] As can be seen from Table 1, the method provided by the present invention achieves the best recognition result and has the best non-line-of-sight signal recognition performance.

[0060] Based on the above analysis, the method provided by the present invention, by introducing the wavelet, not only provides time features for the convolutional neural network, but also extracts the parts with more obvious features, shortens the signal length, improves the efficiency and speed of training, and further improves the recognition ability of non-line-of-sight signals. The method has good non-line-of-sight signal recognition ability.

[0061] The above embodiments are only used to illustrate the present invention and are not intended to limit the scope of protection of the present invention. Within the scope of the technical solution described in the present invention, various simple deformations and modifications made by technicians in the relevant technical field should be included in the scope of the above patent application.

Claims

1. A non-line-of-sight signal recognition method based on wavelet-Gram convolutional neural network, characterized in that: Contains the following steps: (1) Obtaining training samples: Collect N channel impulse response (CIR) signal samples with labeled signal categories as the training set S0 = {(x1(t), y1), (x2(t), y2), ..., (x N (t),y N )}; where x n (t), n = 1, 2, ..., N, represents the nth sample in the training set, which is a 1 × T vector, where T represents the time dimension of the sample; y n Represents sample x n (t) label, y n ∈{0,1}; (2) Extracting the effective features of the original CIR signal: Performing discrete wavelet transform on each sample in the extracted training set S0, taking the low-frequency part with more obvious features as the effective part, and constructing a new training set in, Represents the effective part of the CIR signal after discrete wavelet transform, which is the nth sample of the training set S1. vector, T represents the time dimension of the original CIR signal; y n is a label, which is the same as the label in step (1); (3) Feature preprocessing: Normalize each sample in the training set S1 so that it is in the interval (-1, 1) to satisfy the Gram angular field transform condition; (IV) Converting one-dimensional signal to two-dimensional image: Using Gram angular field transform, each one-dimensional signal sample in the training set S1 is converted into two two-dimensional angular field images containing time relationship features, including Gram angular sum field GASF and Gram angular difference field GADF, to construct a new training set in, Represents the two-dimensional image after Gram angular field transformation, which is the nth sample of the training set S. , T represents the time dimension of the original CIR signal; y n is a label, which is the same as the label in step (1); (5) Convolutional neural network classification: The training set S is put into the convolutional neural network for classification training, and the test set is used to test the network training effect.

2. The non-line-of-sight signal recognition method based on wavelet-gram convolutional neural network according to claim 1, characterized in that: In the step (1), N channel impulse response samples with signal category labels are collected as a training set S0 = {(x1(t), y1), (x2(t), y2), ..., (x N (t),y N )}; where x n (t), n = 1, 2, ..., N, represents the nth sample in the training set, which is a 1 × T vector, where T represents the time dimension of the sample; y n Represents sample x n (t) label, y n ∈{0,1}.

3. The non-line-of-sight signal recognition method based on wavelet-gram convolutional neural network according to claim 2, characterized in that: In the step (ii), each sample in the extracted training set S0 is subjected to discrete wavelet transform, and the bior3.1 wavelet is used to perform discrete wavelet transform, and the original CIR signal is divided into a high-frequency part and a low-frequency part. The low-frequency part after the transform contains more features than the original CIR signal and the high-frequency part after decomposition, so the low-frequency part after discrete wavelet transform is continued to be processed as the effective part. Since the discrete wavelet transform adopts a downsampling process, the length of the low-frequency part after decomposition is half of the original CIR signal, and the length becomes Where T is the time dimension of the original CIR signal, and a new training set is constructed in, Represents the effective part of the CIR signal after discrete wavelet transform, which is the nth sample of the training set S1. vector, T represents the time dimension of the original CIR signal; y n is a label, which is the same as the label in step (1).

4. The non-line-of-sight signal recognition method based on wavelet-gram convolutional neural network according to claim 3, characterized in that: In step (3), the sample vectors in the training set S1 are normalized to reduce the low-frequency part after discrete wavelet transform to the range of -1 to 1 so as to meet the conditions of Gram angular field transform. The normalization formula is: Where X is the original time series data, X min is the minimum value in the original time series data, X max This is the maximum value in the original time series data; this X is also used when the test set samples are standardized. min With X max If the value is greater than 1 after normalization, it is set to 1; if the value is less than -1, it is set to -1.

5. The non-line-of-sight signal recognition method based on wavelet-gram convolutional neural network according to claim 4, characterized in that: In step (iv), each one-dimensional signal sample in the training set S1 is converted into two two-dimensional angular field images containing time relationship features using the Gram angle field transform, including the Gram angle sum field GASF and the Gram angle difference field GADF. The Gram angle field transform method is as follows: (1) Standardize the one-dimensional time signal and scale it to the (-1, 1) interval. The standardization process has been completed in step (3), so there is no need to repeat the operation. (2) The normalized one-dimensional time signal is converted from rectangular coordinates to polar coordinates. The conversion formula is: Where i represents the timestamp of the one-dimensional time signal, represents the signal amplitude at the i-th time stamp, θ i Indicates the angle of the signal at the i-th time stamp converted to polar coordinates, R i Indicates the radius of the signal at the i-th timestamp converted to polar coordinates, and N represents the signal length of the one-dimensional time signal; (3) In the principle of Gram matrix, the cosine function is used to represent the correlation between two vectors. Therefore, the Gram angle field uses the cosine function to represent the correlation between two angles, which is the Gram angle sum field. The sin function is used to represent the difference between the two angles, which is the Gram angle difference field. The transformation formula is: Where GASF represents the Gram angle sum field matrix, and GADF represents the Gram angle difference field matrix; GASF and GADF are packaged to construct a new training set in, Represents the two-dimensional image after Gram angular field transformation, which is the nth sample of the training set S. , T represents the time dimension of the original CIR signal; y n is a label, which is the same as the label in step (1).

6. The non-line-of-sight signal recognition method based on wavelet-gram convolutional neural network according to claim 5, characterized in that: In the step (5), the training set S is put into the convolutional neural network for classification training, and the network training effect is tested with the test set.

Images