UWB (ultra wide band) arrival angle estimation method and device based on feature fusion under non-line-of-sight condition

Abstract

The invention discloses a UWB angle of arrival estimation method and device based on feature fusion under a non-line-of-sight condition, and the method comprises the steps: obtaining a channel impulse response (CIR) through a CIR feature extraction module, the CIR feature extraction module comprises a CNN module and a Transform encoder, the CNN module extracts the spatial-temporal features of a signal through a convolution layer and a pooling layer, and the spatial-temporal features of the signal are obtained through the transform encoder; the spatial and temporal characteristics are enhanced by using the Transform encoder; based on the input double-antenna signal difference, a double-antenna feature module is used for extracting the difference feature between the double antennas, and the double-antenna feature module is used for enhancing feature expression through a multi-layer perceptron; and splicing output features of the CIR feature extraction module and the double-antenna feature module by using a feature fusion and output layer to generate an AOA estimation result. According to the invention, high-precision AOA estimation can be carried out in a complex propagation environment.

Description

Technical Field

[0001] The present invention belongs to the field of communication technology, and in particular relates to a method and device for estimating UWB angle of arrival based on feature fusion under non-line-of-sight conditions. Background Art

[0002] In the field of ultra-wideband (UWB) angle measurement, hardware impairments such as antenna coupling and interference in non-line-of-sight (NLOS) conditions severely impact angle estimation accuracy, especially in complex environments. Antenna coupling can cause signal interference, resulting in measurement errors. In NLOS environments, multipath propagation and reflections further exacerbate errors, making angle estimation more difficult.

[0003] Specifically, high-precision positioning is crucial for wireless communications, sensor networks, and commercial applications, providing critical support for precise navigation and location awareness. Ultra-wideband (UWB) is widely used in commercial devices due to its superior temporal resolution and excellent anti-interference capabilities. UWB transceivers equipped with antenna arrays can estimate the signal's AOA, adding an additional advantage to current UWB positioning systems.

[0004] Current research primarily estimates the angle of arrival (AOA) using the phase difference of arrival (PDOA) between antennas. However, hardware impairments to the antennas, such as antenna coupling, can lead to significant signal distortion, which is difficult to accurately model. Furthermore, under NLOS conditions, environmental factors have a more significant impact on positioning. Direct signals are attenuated by obstacles, and the received signal is typically the result of multipath propagation. Multipath effects complicate the signal propagation path, posing a greater challenge to AOA estimation, as the actual signal propagation path is difficult to accurately capture.

[0005] Traditional modeling methods often rely on complex physical models or assumptions, but in actual environments, due to the highly nonlinear characteristics of antenna damage and multipath propagation, these models are often difficult to adapt to various complex changes. Summary of the Invention

[0006] In response to the above problems, the present invention provides a method and device for estimating the angle of arrival of UWB based on feature fusion under non-line-of-sight conditions, aiming to improve the estimation accuracy in complex environments.

[0007] According to a first aspect of an embodiment of the present disclosure, a method for estimating UWB angle of arrival based on feature fusion under non-line-of-sight conditions is provided, the method comprising:

[0008] A CIR feature extraction module is used to obtain a channel impulse response (CIR). The CIR feature extraction module includes a CNN module and a Transformer encoder. The CNN module extracts the spatiotemporal features of the signal through convolutional layers and pooling layers, and the Transformer encoder is used to enhance the spatiotemporal features.

[0009] Based on the difference of the input dual-antenna signals, a dual-antenna feature module is used to extract the difference features between the dual antennas, wherein the dual-antenna feature module includes enhancing feature expression using a multi-layer perceptron;

[0010] The output features of the CIR feature extraction module and the dual-antenna feature module are spliced ​​together using a feature fusion and output layer to generate an AOA estimation result.

[0011] In some embodiments, the Transformer encoder uses a self-attention mechanism to assign weights to different parts of the input data to capture global context information.

[0012] In some embodiments, the original input data in the CNN module passes through the one-dimensional convolution operation of the convolution layer and then flows into the activation layer. The activation layer performs a nonlinear transformation on the data output by the convolution layer to capture deep features and the inherent complex structure of the data. The pooling layer reduces the dimension and filters the feature data. The fully connected layer flattens the two-dimensional feature map output by the pooling layer into a one-dimensional feature vector and maps the feature vector to the final output result.

[0013] In some embodiments, the self-attention mechanism input includes Query, Key and Value. For the input matrix X, it is mapped through a linear transformation layer to generate the corresponding query vector Q x , key vector K x Sum value vector V x , and then through a layer of matrix multiplication, the query vector Q x and key vector K x Perform matrix multiplication to calculate Q x and K x The similarity or correlation between them is used to obtain an attention score matrix and scale it; the scaled attention score matrix is ​​converted into attention weights based on probability distribution through the SoftMax function, and the attention weights are combined with the value vector V x Perform matrix multiplication to obtain a weighted value vector, which is the final output.

[0014] In some embodiments, the Transformer encoder processes multiple attention distributions in parallel, and the output of the multiple attention distributions is to convert the output of each attention head i Concatenate and multiply by WO , W O The expression is hd v ×d model , h is the number of multi-head self-attention, d v is the output length of self-attention, d model is the embedding dimension.

[0015] In some embodiments, the input to the dual antenna feature module includes the following features:

[0016] Estimating angle of arrival based on phase difference φ represents the phase difference obtained by the dual-antenna array;

[0017] Direct path peak difference: F2 = peak1-peak2, where peak1 and peak2 represent the direct path peaks of the dual-antenna array.

[0018] Signal direct path power difference: F3 = fppower1 - fppower2, where fppower1 and fppower2 represent the direct path power of the signal of the dual-antenna array respectively;

[0019] Signal power difference: F4 = power1-power2, where power1 and power2 represent the signal power of the dual-antenna array respectively;

[0020] The real part of the correlation coefficient of the dual-antenna channel impulse response: F5 = Corr(CIR1, CIR2), is calculated as follows:

[0021]

[0022] Among them, CIR1 and CIR2 are the dual-antenna channel impulse responses respectively.

[0023] According to a second aspect of an embodiment of the present disclosure, a UWB arrival angle estimation device based on feature fusion under non-line-of-sight conditions is provided, the device comprising:

[0024] A CIR feature extraction module is used to obtain a channel impulse response (CIR). The CIR feature extraction module includes a CNN module and a Transformer encoder. The CNN module extracts the spatiotemporal features of the signal through convolutional layers and pooling layers, and enhances the spatiotemporal features using the Transformer encoder.

[0025] A dual-antenna feature module is used to extract differential features between the dual antennas based on the difference in the input dual-antenna signals, and the dual-antenna feature module includes enhancing feature expression using a multi-layer perceptron;

[0026] The feature fusion and output layer is used to generate an AOA estimation result after splicing the output features of the CIR feature extraction module and the dual-antenna feature module.

[0027] In some embodiments, the Transformer encoder in the CIR feature extraction module uses a self-attention mechanism to assign weights to different parts of the input data to capture global context information.

[0028] According to a third aspect of an embodiment of the present disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of the above-mentioned method for estimating the angle of arrival of UWB based on feature fusion under non-line-of-sight conditions are implemented.

[0029] According to a fourth aspect of an embodiment of the present disclosure, a non-temporary computer-readable storage medium is provided, on which computer instructions are stored. When the instructions are executed by a processor, the steps of the UWB arrival angle estimation method based on feature fusion under non-line-of-sight conditions are implemented.

[0030] The embodiments of the present disclosure provide a method and device for estimating UWB angle of arrival based on feature fusion under non-line-of-sight conditions, which have the following beneficial effects:

[0031] 1. A feature fusion network is proposed that fully utilizes the complementary information provided by CIR features and dual-antenna features. The former covers rich details of the channel propagation environment, while the latter provides direct indication of the signal source direction. Through the feature fusion network structure, CIR features and dual-antenna data are extracted and effectively integrated at the fusion layer, enabling this method to perform high-precision AOA estimation in complex propagation environments.

[0032] 2. The Transformer encoder model is introduced to deeply mine complex signal features, which is more superior than LSTM and ResNet.

[0033] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present invention and, together with the description, serve to explain the principles of the present invention.

[0035] Figure 1 is a schematic diagram of signal arrival in a signal model according to an embodiment of the present invention;

[0036] Figure 22. It is a schematic diagram of the AOA estimation result of PDOA in the NLOS environment according to an embodiment of the present invention;

[0037] Figure 3 This is a flow chart of a method for estimating UWB angle of arrival based on feature fusion under non-line-of-sight conditions in an embodiment of the present invention;

[0038] Figure 4 This is a principle block diagram of a method for estimating the angle of arrival of UWB based on feature fusion under non-line-of-sight conditions according to an embodiment of the present invention;

[0039] Figure 5 is a flowchart of a Transformer encoder according to an embodiment of the present invention;

[0040] Figure 6 is a flow chart of the self-attention mechanism in an embodiment of the present invention;

[0041] Figure 7 A schematic diagram of the structure of a UWB angle-of-arrival estimation device based on feature fusion under non-line-of-sight conditions according to an embodiment of the present invention;

[0042] Figure 8 is a schematic diagram of an electronic device according to an embodiment of the present invention;

[0043] Figure 9 This is a diagram of the situation collected by the test equipment of the embodiment of the present invention in different experimental scenarios;

[0044] Figure 10 2. It is a schematic diagram comparing the performance of different models in the training set according to the embodiment of the present invention;

[0045] Figure 11 2 is a schematic diagram comparing the performance of different models in the test set according to the embodiment of the present invention. DETAILED DESCRIPTION

[0046] The present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are intended only to illustrate the present invention and are not intended to limit the present invention. It should also be noted that, for ease of description, the accompanying drawings only illustrate portions relevant to the present invention, not all structures.

[0047] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flow charts. Although the flow charts describe the steps as sequential processes, many of the steps can be implemented in parallel, concurrently, or simultaneously. In addition, the order of the steps can be rearranged. The process can be terminated when its operation is completed, but can also have additional steps not included in the accompanying drawings. The process can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0048] The present invention provides a method and apparatus for UWB angle-of-arrival (AOA) estimation based on feature fusion under non-line-of-sight (NLOS) conditions. This method proposes a feature fusion method that significantly enhances the robustness of the model in complex environments and under hardware impairments, thereby achieving more accurate angle-of-arrival (AOA) estimation. This method not only effectively improves estimation accuracy under adverse conditions but also has strong practical application value.

[0049] Deep learning can automatically extract complex features and perform effective modeling in a data-driven manner, especially when dealing with nonlinear relationships and high-dimensional data, showing stronger adaptability and accuracy than traditional methods.

[0050] To address challenges in complex environments, such as antenna hardware damage and multipath propagation in non-line-of-sight conditions, this embodiment proposes a deep learning-based AOA estimation method. By automatically learning the complex relationship between signal characteristics and AOA, this method overcomes the limitations of traditional models in hardware damage and harsh environments, achieving more accurate AOA estimation.

[0051] The machine learning-based ultra-wideband signal arrival angle estimation method of the present invention has two main innovations: first, a feature fusion network is proposed to fully utilize the complementary information provided by the channel impulse response (CIR) features and dual-antenna features; second, a Transformer encoder model is introduced to deeply mine complex signal features.

[0052] In some application embodiments, the anchor receives the UWB signal transmitted by the tag and records the CIR. The CIR at antenna m can be expressed as in and are the complex amplitude and delay of the mth antenna at path l, T is the number of samples of the CIR length, L is the total number of paths, δ is the impulse function, and n(t) is Gaussian white noise. Qorvo's existing DW3120 and DW3220 can not only perform two-way ranging (TWR) ranging, but also provide dual-antenna PDoA, which can simultaneously measure the AoA and arrival time (ToF) of the received signal. In the 2-D plane, the arriving signal reaches the dual antennas, Figure 1 Schematic diagram of signal arrival.

[0053] At this time, the phase difference obtained by the dual antenna array can be expressed as: To avoid phase entanglement, the dual antenna spacing needs to be set to less than λ / 2. In this embodiment, d is preferably set to 2 cm, CH5 is used, and the center frequency is 6489.6 MHz. After obtaining the phase difference, the AOA can be estimated as:

[0054] In NLOS environments, since the signal is affected by obstacle reflection and multipath effects, the method based on phase difference angle of arrival (PDOA) is usually difficult to achieve accurate AOA estimation. Figure 2 As shown in the figure, although the AOA estimation based on PDOA shows an overall linear variation within the range of -60° to 40° in NLOS environments, its accuracy and stability are poor, with significant deviations. This is primarily due to the influence of NLOS interference. In other angular ranges, the nonlinearity of the error is even more pronounced. This instability is closely related not only to hardware impairments, especially signal distortion caused by antenna coupling, but also to phase difference measurement bias caused by multipath effects. Therefore, relying solely on the PDOA method cannot meet the requirements of high-precision positioning, and its applicability in complex environments is significantly limited. More advanced models are urgently needed to overcome these challenges.

[0055] The feature fusion-based AOA estimation algorithm involved in the embodiments of the present invention aims to combine the unique advantages of different features to achieve more accurate AOA estimation. This algorithm fully utilizes the complementary information provided by CIR features and dual-antenna features. The former covers rich details of the channel propagation environment, while the latter provides direct indication of the signal source direction. Through a feature fusion network structure, CIR features and dual-antenna data are extracted and effectively integrated at the fusion layer, enabling this method to perform highly accurate and robust AOA estimation in complex propagation environments.

[0056] Example 1 is used to illustrate a UWB arrival angle estimation method based on feature fusion under non-line-of-sight conditions. Figure 3 As shown, the following steps are included:

[0057] S1. Obtain a channel impulse response (CIR) using a CIR feature extraction module. The CIR feature extraction module includes a CNN module and a Transformer encoder. The CNN module extracts spatiotemporal features of the signal through convolutional layers and pooling layers, and enhances the spatiotemporal features using the Transformer encoder.

[0058] S2. Based on the input dual-antenna signal difference, using a dual-antenna feature module to extract the difference features between the dual antennas, wherein the dual-antenna feature module includes enhancing feature expression using a multi-layer perceptron;

[0059] S3. Use feature fusion and output layer to splice the output features of the CIR feature extraction module and the dual-antenna feature module to generate an AOA estimation result.

[0060] Specifically, the principle block diagram of the AOA estimation algorithm based on feature fusion is as follows: Figure 4 shown.

[0061] The original input data in the CNN module undergoes a one-dimensional convolution operation in the convolution layer and then flows into the activation layer. The activation layer performs a nonlinear transformation on the data output by the convolution layer to capture deep features and the inherent complex structure of the data. The pooling layer reduces the dimensionality and filters the feature data. The fully connected layer flattens the two-dimensional feature map output by the pooling layer into a one-dimensional feature vector and maps the feature vector to the final output result.

[0062] Specifically, the CIR feature extraction module consists of a CNN and a Transformer encoder. The CNN module extracts the spatiotemporal features of the signal through its convolutional and pooling layers, thereby generating a deep representation for subsequent analysis. The convolutional layer extracts local features through a sliding window operation, which can effectively capture local patterns in the CIR signal. The pooling layer ensures the integrity of key information while reducing the amount of computation, further improving the generalization ability of the model. After the CNN module, the Transformer encoder is introduced to enhance the ability to capture temporal features. The self-attention mechanism used in the Transformer structure can assign weights to different parts of the input data, allowing the network to focus on key features in the signal, improving the performance of AOA estimation. The introduction of the self-attention mechanism ensures that the model can capture global context information, making the output more robust and accurate.

[0063] The core component of a CNN is the convolution layer, which performs the convolution operation for feature extraction. A one-dimensional convolution operation involves the interaction between the original signal and the convolution kernel. Below is a detailed explanation and calculation of a one-dimensional convolution operation: The original input data (signal) is a vector of size N. The convolution kernel is a vector w of size M, where N ≤ M. The output vector y stores the result of the convolution operation. The convolution result is as follows:

[0064]

[0065] After the one-dimensional convolution operation of the convolutional layer, the data usually flows into the activation layer. The main function of the activation layer is to introduce nonlinearity. By performing nonlinear transformations on the data output by the convolutional layer, the network can learn and simulate more complex functions and patterns. Nonlinear mapping allows the network to abstract from simple linear relationships and capture deeper features and the inherent complex structure of the data. The activation function selected by the network is ReLU (Rectified Linear Unit), specifically:

[0066] ReLU(x)=max(0,x)

[0067] Before entering subsequent layers, data often needs to be processed by a pooling layer. The primary purpose of the pooling layer is to reduce the dimensionality of feature data and filter it, removing redundant information so that the network can more effectively capture the key features of the original data. Pooling layers are divided into maximum pooling and average pooling. After the dimensionality reduction processing of the pooling layer, the data typically enters the fully connected layer. Its function is to flatten the two-dimensional feature maps output by the pooling layer into one-dimensional feature vectors and map these feature vectors to the final output.

[0068] The specific CNN network parameter settings are as follows:

[0069] Table 1 CNN parameter settings

[0070]

[0071] The Transformer encoder uses a self-attention mechanism to assign weights to different parts of the input data to capture global context information.

[0072] Specifically, based on CNN, the Transformer encoder is introduced to solve the problem of multipath effect and deep mining of complex signal features in non-line-of-sight environments. The flowchart of the Transformer encoder is as follows Figure 5 As shown in Figure 2. One of its core components is the attention mechanism, which plays a key role in capturing information. The self-attention mechanism allows the model to directly focus on any two positions in the sequence when processing sequence data, without the need for a fixed context vector. The process of the self-attention mechanism is as follows: Figure 6 Its input usually consists of three parts, namely Q (Query), K (Key), and V (Value). For a given input matrix X, each part will be mapped through a linear transformation layer to generate the corresponding query vector, key vector, and value vector. This process can be expressed as the following formula:

[0073] Q x =Wq X

[0074] K x =W k X

[0075] V x =W ν X

[0076] W q 、W k 、W ν They are the corresponding linear transformation coefficient matrices, and then pass through a layer of matrix multiplication. This layer converts the query vector Q x and key vector K x Perform matrix multiplication. Its purpose is to calculate Q x and K x The similarity or correlation between them is used to obtain an attention score matrix, such as Figure 6 shown.

[0077] After matrix multiplication, the result is usually scaled. This is because in the self-attention mechanism, Q x and K x The dimension of is usually large, and the directly calculated attention score may be very large, leading to the gradient disappearance or gradient explosion problem. Scaling is usually done by dividing the score by a constant, usually Where D is K x The mask layer is optional and is used to mask out the attention scores of certain positions. The scores are converted into probability distributions through the SoftMax function. The sum of the attention weights of all positions is 1, and the output of each position is a weighted sum, where the contribution of each input position is determined by its corresponding attention weight. Finally, the attention weight output by SoftMax is combined with the value vector V x Perform matrix multiplication. The result of this step is a weighted value vector, the final output, which integrates the information of all input positions, but the weights are dynamically determined by the attention mechanism. The output value of the self-attention mechanism is:

[0078]

[0079] The Transformer encoder uses a multi-self-attention mechanism, which is an extension of the self-attention mechanism. It enhances the expression and generalization capabilities of the model by processing multiple attention distributions in parallel. The output of each attention head i As shown in the following formula, the output of multiple heads is to combine each head i Concatenate and multiply by W O , as shown in the following formula.

[0080]

[0081] MultiHeadAttention(Q x ,K x ,V x )=Concat(head1,...head h )W O

[0082] Among them, W i Q 、 W i V Denotes the weight matrices of the Q, K, and V matrices of the i-th head, h is the number of multi-head self-attention, and W O The shape is hd v ×d model , the output length of self-attention is d v , there are h outputs, of which d model is the embedding dimension, d v =d model / h.

[0083] Table 2 details the key parameters of the Transformer encoder in the model, including the embedding dimension, the number of attention heads, the feedforward layer dimension, the number of encoder layers, and other core settings.

[0084] Table 2 Transformer encoder parameter configuration

[0085] Parameter name Parameter value Embedding Dimension 128 Number of attention heads 4 Feedforward layer dimension 128 Number of encoder layers 2

[0086] The dual-antenna feature module, primarily composed of fully connected layers, is designed to extract differential features between the two antennas. By introducing additional features such as peak difference and phase difference between the two antennas, this module can more comprehensively capture the spatial information between the antennas. This module uses a multi-layer perceptron (MLP) to enhance the expressive power of features through nonlinear transformations, thereby improving the model's adaptability and predictive performance in complex environments.

[0087] The input of this module focuses on the signal difference between the two antennas, and the following features are selected as input:

[0088] 1. PDOA-AOA:

[0089] 2. Direct path peak difference: F2 = peak1-peak2

[0090] 3. Signal direct path power difference: F3 = fppower1 - fppower2

[0091] 4. Signal power difference: F4 = power1-power2

[0092] 5. The real part of the correlation coefficient of the two antenna CIR (Channel impulse response): F5 = Corr(CIR1, CIR2), calculated as follows: Where power1 and power2 are the two signal powers, fppower1 and fppower2 are the direct path powers of the two antennas, and peak1 and peak2 are the direct path peak values ​​of the two antennas.

[0093] After feature extraction, the CIR and dual-antenna output features are concatenated to form a comprehensive feature vector. This network leverages the strengths of both types of features to achieve more accurate AOA estimation. This comprehensive feature vector is then processed through a series of fully connected layers, ultimately outputting a predicted value—the desired AOA estimation result.

[0094] Another embodiment is used to illustrate a UWB arrival angle estimation device based on feature fusion under non-line-of-sight conditions. The device 700 includes:

[0095] A CIR feature extraction module 710 is used to obtain a channel impulse response (CIR). The CIR feature extraction module includes a CNN module and a Transformer encoder. The CNN module extracts the spatiotemporal features of the signal through convolutional layers and pooling layers, and enhances the spatiotemporal features using the Transformer encoder.

[0096] A dual-antenna feature module 720 is configured to extract differential features between the dual antennas based on the difference in the input dual-antenna signals, wherein the dual-antenna feature module includes enhancing feature expression using a multi-layer perceptron;

[0097] The feature fusion and output layer 730 is used to generate an AOA estimation result after splicing the output features of the CIR feature extraction module and the dual-antenna feature module.

[0098] The Transformer encoder in the CIR feature extraction module 710 uses a self-attention mechanism to assign weights to different parts of the input data to capture global context information.

[0099] In addition to the above modules, the apparatus 700 may further include other components. However, since these components are irrelevant to the contents of the embodiments of the present disclosure, their illustration and description are omitted here.

[0100] The other specific working processes of the UWB arrival angle estimation device 700 based on feature fusion under non-line-of-sight conditions refer to the description of the embodiment of the UWB arrival angle estimation method based on feature fusion under non-line-of-sight conditions, which will not be repeated here.

[0101] Another embodiment is used to illustrate that the system of the present invention can also be used with the help of Figure 8 The architecture of the computing device shown is implemented. Figure 8 The architecture of the computing device is shown in FIG. Figure 8 As shown, a computer system 810, a system bus 830, one or more CPUs 840, an input / output 820, a memory 850, etc. The memory 850 can store various data or files used for computer processing and / or communication, as well as program instructions executed by the CPU, including the method for estimating the angle of arrival of UWB based on feature fusion under non-line-of-sight conditions of the embodiment. Figure 8 The architecture shown is only exemplary and may be adjusted based on actual needs when implementing different devices. Figure 8 One or more components in the system. The memory 850, as a computer-readable storage medium, can be used to store software programs, computer executable programs, and modules, such as the program instructions / modules corresponding to the UWB arrival angle estimation method based on feature fusion under non-line-of-sight conditions in the embodiment of the present invention (for example, the CIR feature extraction module 710, the dual-antenna feature module 720, and the feature fusion and output layer 730 in the UWB arrival angle estimation device 700 based on feature fusion under non-line-of-sight conditions). One or more CPUs 840 execute various functional applications and data processing of the system of the present invention by running the software programs, instructions, and modules stored in the memory 850, that is, to implement the above-mentioned UWB arrival angle estimation method based on feature fusion under non-line-of-sight conditions, which includes:

[0102] A CIR feature extraction module is used to obtain a channel impulse response (CIR). The CIR feature extraction module includes a CNN module and a Transformer encoder. The CNN module extracts the spatiotemporal features of the signal through convolutional layers and pooling layers, and the Transformer encoder is used to enhance the spatiotemporal features.

[0103] Based on the difference of the input dual-antenna signals, a dual-antenna feature module is used to extract the difference features between the dual antennas, wherein the dual-antenna feature module includes enhancing feature expression using a multi-layer perceptron;

[0104] The output features of the CIR feature extraction module and the dual-antenna feature module are spliced ​​together using a feature fusion and output layer to generate an AOA estimation result.

[0105] Of course, the processor of the server provided in the embodiment of the present invention is not limited to executing the method operations described above, but can also execute relevant operations in the UWB arrival angle estimation method based on feature fusion under non-line-of-sight conditions provided in any embodiment of the present invention.

[0106] The memory 850 may primarily include a program storage area and a data storage area, wherein the program storage area may store an operating system and applications required for at least one function; the data storage area may store data created based on the use of the terminal, etc. Furthermore, the memory 850 may include high-speed random access memory and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state memory device. In some instances, the memory 850 may further include memory remotely located relative to one or more CPUs 840, and these remote memories may be connected to the device via a network. Examples of the aforementioned networks include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0107] The input / output 820 may be used to receive input digital or character information and generate key signal input related to user settings and function control of the device. The input / output 820 may also include a display device such as a display screen.

[0108] The embodiments of the present invention also provide a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for estimating the angle of arrival of UWB based on feature fusion under non-line-of-sight conditions described in the above embodiments is implemented. The computer-readable storage medium of the embodiments of the present invention may adopt any combination of one or more computer-readable media. The computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0109] A computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, which carries computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium that can transmit, propagate, or transport a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0110] The program code embodied on the storage medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

[0111] In addition, other specific working processes of a non-transitory computer-readable storage medium refer to the description of the embodiment of the UWB arrival angle estimation method based on feature fusion under non-line-of-sight conditions mentioned above, and are not repeated here.

[0112] In order to verify the effectiveness of the method of the present invention, the following verification examples are provided in the embodiment:

[0113] The datasets used were collected by the DW3000 device in different experimental scenarios, such as Figure 9 The main purpose of the experiment is to evaluate the performance of the feature fusion strategy for AOA estimation in complex NLOS environments. To this end, the embodiment uses a real UWB dataset that contains key information such as CIR and PDOA.

[0114] A total of 13,430 sets of data were collected in the experiment, of which the training set accounted for 70% of the total data volume and was used to update and optimize the model parameters; the validation set accounted for 15% and was used to evaluate the performance of the model on unseen data after each round of training to prevent overfitting; the test set accounted for the remaining 15% and was used for the final evaluation of the model after training to ensure the reliability of the model performance. During the entire training process, the training set and validation set were used for learning and adjusting the model, while the test set was used for the final model evaluation. The deep learning model of this patent was built under the pytorch framework. The batch size was set to 32. During the training process, the Adam optimizer was used. In order to ensure the training effect of the model, the maximum training cycle was set to 200 times, and the early stopping mechanism was introduced.

[0115] For the hyperparameter selection of the model proposed in this method, experiments tested [2, 3, 4] convolutional layers and kernel sizes of [2, 3, 4] for the CNN. For the Transformer, hidden dimensions of [64, 128, 256] were tested, and the number of multi-head attention heads was set to [4, 8, 16]. The best performing hyperparameter combination was ultimately selected by evaluating performance metrics on the validation set.

[0116] To highlight the superiority of the proposed method, we conducted a detailed analysis of the performance of various AOA estimation methods in NLOS scenarios. The experiments focused on MAE, STD, and the AOA estimation success rate within the ±3° and ±5° ranges. These metrics comprehensively reflect the estimation accuracy and stability of each method. The following commonly used AOA estimation methods were used for comparison:

[0117] (1) Single PDOA method: The PDOA method is a classic AOA estimation technique that estimates the angle by measuring the phase difference of the signal. This method has a relatively simple calculation process and can provide relatively accurate AOA estimation under ideal conditions.

[0118] (2) Feature fusion model based on Long Short-Term Memory (LSTM): LSTM is a deep learning model for processing time series data that can effectively capture the temporal dependencies of signals. In NLOS environments, the LSTM model can extract potential patterns from time series signals and provide support for AOA estimation.

[0119] (3) Feature fusion model based on Residual Network (ResNet): ResNet is a deep convolutional neural network that overcomes the vanishing gradient problem by introducing residual connections. In the AOA estimation task, this model effectively improves positioning accuracy by virtue of its deep feature extraction capabilities.

[0120] (4) Dual-antenna feature fully connected model: This method estimates AOA by fully connecting the features of the dual-antenna received signals and combining them with the spatial information of the dual antennas.

[0121] Next, we analyze the performance on the training and test sets. First, on the training set, as shown in Table 3, the experimental results show that the feature fusion (CNN) model performs well on the training set, with a MAE of 0.3287°, an RMSE of 1.1596°, and an STD of 1.1596°. However, it shows a significant decline on the validation set, with a MAE of 4.3009°, an RMSE of 8.6724°, and an STD of 8.6666°, indicating that the model is overfitting and lacks the ability to extract generalized features from the training data. In contrast, the feature fusion (CNN + Transformer Encoder) model outperforms the other models on both the training and validation sets. Although its MAE was 1.6478°, RMSE was 3.1644°, and STD was 3.1590° on the training set, it performed particularly well on the validation set, with MAE of 2.7683°, RMSE of 6.9141°, and STD of 6.9133°, demonstrating stronger generalization and robustness, significantly improving AOA estimation accuracy. Feature fusion (CNN+LSTM) and feature fusion (ResNet) performed poorly, especially in terms of accuracy. The LSTM model had significantly higher MAE and RMSE.

[0122] Table 3 Performance comparison of different models in the training set

[0123]

[0124] At the same time, the ±3° accuracy, ±5° accuracy and training time of the training set are compared, as shown in Table 4 and Figure 10 As shown. Experimental results show that the feature fusion model proposed in this application, namely the CNN+Transformer encoder, is significantly better than other models in terms of accuracy. In terms of training time, it is also more efficient than feature fusion (CNN+LSTM). Overall, the feature fusion model proposed in this patent performs best in terms of accuracy and training efficiency. Although the dual-antenna feature fully connected model has a shorter training time, its accuracy is far inferior to feature fusion (CNN+Transformer encoder).

[0125] Table 4 Performance comparison of different models on the training set 2

[0126]

[0127] From the test set results, as shown in Table 5 and Figure 11As shown, the feature fusion (CNN+Transformer encoder model) performs well, significantly outperforming other models with a MAE of 2.5844° and an accuracy of 89.21% within ±5°. Its low error index and high accuracy fully demonstrate its robustness and superiority in complex NLOS environments. In addition, the test time is only 0.54 seconds, demonstrating excellent efficiency. In comparison, the feature fusion CNN+LSTM and feature fusion ResNet are inferior in terms of both error and accuracy. The traditional PDOA method is affected by multipath interference, resulting in high error and severely limited applicability.

[0128] Table 5 Performance comparison of different models on the test set

[0129]

[0130] In summary, the technical solutions provided by the above embodiments provide a method and device for estimating UWB arrival angle based on feature fusion under non-line-of-sight conditions, which have the following beneficial effects:

[0131] 1. A feature fusion network is proposed that fully utilizes the complementary information provided by CIR features and dual-antenna features. The former covers rich details of the channel propagation environment, while the latter provides direct indication of the signal source direction. Through the feature fusion network structure, CIR features and dual-antenna data are extracted and effectively integrated at the fusion layer, enabling this method to perform high-precision AOA estimation in complex propagation environments.

[0132] 2. The Transformer encoder model is introduced to deeply mine complex signal features, which is more superior than LSTM and ResNet.

[0133] In this document, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a step or method that comprises a series of elements includes not only those elements, but also includes other elements not expressly listed, or also includes elements inherent to such step or method.

[0134] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. A UWB arrival angle estimation method based on feature fusion under non-line-of-sight conditions, characterized by: The method comprises: A CIR feature extraction module is used to obtain a channel impulse response (CIR). The CIR feature extraction module includes a CNN module and a Transformer encoder. The CNN module extracts the spatiotemporal features of the signal through convolutional layers and pooling layers, and the Transformer encoder is used to enhance the spatiotemporal features. Based on the difference of the input dual-antenna signals, a dual-antenna feature module is used to extract the difference features between the dual antennas, wherein the dual-antenna feature module includes enhancing feature expression using a multi-layer perceptron; The output features of the CIR feature extraction module and the dual-antenna feature module are spliced ​​together using a feature fusion and output layer to generate an AOA estimation result.

2. The UWB arrival angle estimation method based on feature fusion under non-line-of-sight conditions according to claim 1 is characterized in that: The Transformer encoder uses a self-attention mechanism to assign weights to different parts of the input data to capture global context information.

3. The UWB arrival angle estimation method based on feature fusion under non-line-of-sight conditions according to claim 1 is characterized in that: The original input data in the CNN module undergoes a one-dimensional convolution operation in the convolution layer and then flows into the activation layer. The activation layer performs a nonlinear transformation on the data output by the convolution layer to capture deep features and the inherent complex structure of the data. The pooling layer reduces the dimension and filters the feature data. The fully connected layer flattens the two-dimensional feature map output by the pooling layer into a one-dimensional feature vector and maps the feature vector to the final output result.

4. The method for estimating the angle of arrival of UWB based on feature fusion under non-line-of-sight conditions according to claim 2, characterized in that: The self-attention mechanism input includes Query, Key and Value. For the input matrix X, it is mapped through a linear transformation layer to generate the corresponding query vector Q x , key vector K x Sum value vector V x , and then through a layer of matrix multiplication, the query vector Q x and key vector K x Perform matrix multiplication to calculate Q x and K x The similarity or correlation between them is used to obtain an attention score matrix and scale it; the scaled attention score matrix is ​​converted into attention weights based on probability distribution through the SoftMax function, and the attention weights are combined with the value vector V x Perform matrix multiplication to obtain a weighted value vector, which is the final output.

5. The method for estimating the angle of arrival of UWB based on feature fusion under non-line-of-sight conditions according to claim 2, characterized in that: The Transformer encoder processes multiple attention distributions in parallel, and the output of multiple attention distributions is the output of each attention head. i Concatenate and multiply by W O , W O The expression is hd v ×d model , h is the number of multi-head self-attention, d v is the output length of self-attention, d model is the embedding dimension.

6. The method for estimating the angle of arrival of UWB based on feature fusion under non-line-of-sight conditions according to claim 1, characterized in that: The input of the dual antenna feature module includes the following features: Estimating angle of arrival based on phase difference φ represents the phase difference obtained by the dual-antenna array; Direct path peak difference: F2 = peak1-peak2, where peak1 and peak2 represent the direct path peaks of the dual-antenna array. Signal direct path power difference: F3 = fppower1 - fppower2, where fppower1 and fppower2 represent the direct path power of the signal of the dual-antenna array respectively; Signal power difference: F4 = power1-power2, where power1 and power2 represent the signal power of the dual-antenna array respectively; The real part of the correlation coefficient of the dual-antenna channel impulse response: F5 = Corr(CIR1, CIR2), is calculated as follows: Among them, CIR1 and CIR2 are the dual-antenna channel impulse responses respectively.

7. A UWB arrival angle estimation device based on feature fusion under non-line-of-sight conditions, characterized in that: The device comprises: A CIR feature extraction module is used to obtain a channel impulse response (CIR). The CIR feature extraction module includes a CNN module and a Transformer encoder. The CNN module extracts the spatiotemporal features of the signal through convolutional layers and pooling layers, and enhances the spatiotemporal features using the Transformer encoder. A dual-antenna feature module is used to extract differential features between the dual antennas based on the difference in the input dual-antenna signals, and the dual-antenna feature module includes enhancing feature expression using a multi-layer perceptron; The feature fusion and output layer is used to generate an AOA estimation result after splicing the output features of the CIR feature extraction module and the dual-antenna feature module.

8. The UWB arrival angle estimation device based on feature fusion under non-line-of-sight conditions according to claim 7, characterized in that: The Transformer encoder in the CIR feature extraction module uses a self-attention mechanism to assign weights to different parts of the input data to capture global context information.

9. An electronic device, characterized in that: The method comprises a memory and a processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the processor performs the steps of the UWB arrival angle estimation method based on feature fusion under non-line-of-sight conditions as described in any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed, the method for estimating the UWB angle of arrival based on feature fusion under non-line-of-sight conditions is implemented as described in any one of claims 1 to 6.

Images