UWB positioning system NLOS identification method

Abstract

The invention relates to the technical field of indoor positioning, in particular to an NLOS identification method for a UWB positioning system, and the method comprises the following steps: S1, data preprocessing: carrying out the normalization operation of collected UWB channel pulse response data; s2, performing one-dimensional CNN feature extraction: inputting the data preprocessed in the step S1 into a convolution layer to perform convolution operation, and inputting the data output by the convolution operation into a batch normalization layer to perform normalization operation; s3, self-attention mechanism feature enhancement: inputting the one-dimensional CNN features extracted in the step S2 into a self-attention mechanism, and calculating and outputting attention output features; and S4, classification output: inputting the attention output features output in the step S3 into an input layer, mapping the features to classification labels through a softmax function, and outputting prediction probabilities obtained under data line-of-sight and non-line-of-sight scenes to realize NLOS classification identification. According to the invention, the NLOS state of the UWB signal can be effectively identified, and the precision and reliability of the UWB positioning system are improved.

Description

Technical Field

[0001] The present invention relates to the field of indoor positioning technology, and more particularly to a NLOS recognition method for a UWB positioning system. Background Art

[0002] With the development of smart mobile devices and the Internet of Things (IoT), location-based services (LBS) are becoming increasingly important in daily life and industrial production. Global Navigation Satellite Systems (GNSS) offer stable and highly accurate positioning outdoors. However, indoors, GNSS signals have difficulty penetrating obstacles such as buildings, significantly weakening or even failing to penetrate. Furthermore, multipath transmission within buildings can lead to errors in positioning calculations, making satellite signals often unavailable or unreliable indoors, significantly reducing positioning accuracy.

[0003] To address the limitations of GNSS indoor positioning, researchers have developed a variety of indoor positioning technologies, such as Wi-Fi, Bluetooth, geomagnetism, and UWB. Among them, UWB positioning technology has been widely used in industrial automation, smart homes, smart transportation, and other fields due to its advantages of large transmission bandwidth and high positioning accuracy. However, the NLOS transmission of signals seriously affects the accuracy of indoor positioning systems. Under NLOS conditions, the signal propagation path is blocked by indoor obstacles, resulting in longer ranging signal propagation time and reduced performance of ranging positioning algorithms based on time of arrival (TOA) and time difference of arrival (TDOA). Compared with other technologies, the NLOS problem is more prominent in UWB ranging algorithms because of their wide operating range and high positioning accuracy requirements.

[0004] Currently, methods for identifying and mitigating NLOS / LOS situations based on machine learning techniques can be divided into two main categories. Traditional NLOS scenario recognition methods do not use machine learning techniques, but instead manually analyze static parameter features extracted from the UWB channel impulse response (CIR) signal, such as kurtosis, skewness, maximum amplitude, and rise time. However, these methods are limited by prior knowledge, require complex data preprocessing, and may ignore the correlation between feature parameters. In recent years, the application of machine learning in NLOS recognition has developed. For example, traditional machine learning methods such as support vector machines (SVM) and random forests (RF) have been used for NLOS recognition. Although these methods reduce the workload and improve recognition accuracy, they still require manual extraction of vector features from the CIR time series, which cannot fully utilize the CIR information. With the development of deep learning, NLOS recognition methods based on deep learning, such as convolutional neural networks (CNN) and long short-term memory networks (LSTM), have attracted attention. However, existing methods have problems such as the CNN architecture's insufficient attention to feature information and the LSTM's limited ability to model local features of the data. Therefore, it is necessary to propose a new method to improve the accuracy and robustness of NLOS recognition for UWB signals. Summary of the Invention

[0005] The purpose of the present invention is to provide a non-line-of-sight / line-of-sight recognition method based on the CNN-SA module to solve the problems of low accuracy, insufficient utilization of data information, insufficient model stability and generalization ability in the existing UWB positioning system NLOS recognition method, and improve the positioning accuracy of the UWB positioning system in indoor environments. The present invention integrates the self-attention mechanism into the convolutional neural network and constructs the CNN-SA algorithm. The algorithm consists of three key modules: a one-dimensional convolutional neural network (1D-CNN) feature extraction module, a self-attention module, and an output layer. The 1D-CNN feature extraction module further enhances feature representation, dynamically adjusts the weights of different features, and focuses on key information; the output layer maps the features after feature extraction and self-attention weighting to classification labels to achieve NLOS classification decisions.

[0006] In order to achieve the above object, the present invention is implemented through the following technical solutions:

[0007] A NLOS identification method for a UWB positioning system comprises the following steps:

[0008] S1. Data preprocessing: collecting UWB channel impulse response data and normalizing the UWB channel impulse response data;

[0009] S2. One-dimensional CNN feature extraction: Extract and output one-dimensional CNN features using a feature extraction module of a one-dimensional convolutional neural network, wherein the feature extraction module has a convolution layer and a batch normalization layer, and includes the following steps:

[0010] S21: Input the data preprocessed in step S1 into the convolution layer for convolution operation, capture the local patterns and features of the preprocessed data, and import them into the data model;

[0011] S22. Normalize the data output by the convolutional layer through the batch normalization layer. The calculation formula is as follows:

[0012] S3. Self-attention mechanism feature enhancement:

[0013] The features output by the feature extraction module of the one-dimensional convolutional neural network are input into the self-attention mechanism, and the input data is set to X = {x1, x2, x3, ..., x N}, and calculate the query Q, key K and value V respectively through independent convolution functions f1(X), f2(X) and f3(X), and then calculate the attention output of the self-attention mechanism;

[0014] S4, classification output: The features output by the self-attention module are input into the output layer, the features are mapped to classification labels through the softmax function, and the prediction probabilities obtained in the line-of-sight and non-line-of-sight scenarios are output to achieve NLOS classification and recognition.

[0015] In a preferred embodiment, the UWB channel impulse response data is cleaned and then normalized.

[0016] In a preferred embodiment, the cleaning includes noise removal and abnormal data processing.

[0017] In a preferred embodiment, the feature extraction module of the one-dimensional convolutional neural network includes a convolution layer, a batch normalization layer and a pooling layer, and the convolution layer performs a convolution operation on the input data through multiple filters.

[0018] In the preferred embodiment, the data model in step S21 is as follows:

[0019]

[0020] Where * represents the convolution operation, N is the number of kernels in the μ-1 layer, is the i-th feature map of μ-1 layer, represents the weight, is the bias of the jth convolution kernel in the μ layer, represents the feature map of the j-th convolution output of the μ layer, f is the LeakRELU activation function, and the expression of f is as follows:

[0021]

[0022] Where β = 0.01,

[0023] In the preferred embodiment, the normalization process in step S22 is calculated using the following formula:

[0024]

[0025] in, is the value of the t-th neuron corresponding to the i-th feature map in the μ layer, ω is the width of the pooling layer, is the value of the neuron in the μ+1 layer.

[0026] In the preferred embodiment, in step S3, query Q, key K and value V are calculated respectively by independent convolution functions f1(X), f2(X) and f3(X), and the formula is as follows:

[0027]

[0028] Among them, W1, W2, and W3 represent weight matrices; B1, B2, and B3 represent bias vectors.

[0029] In the preferred embodiment, the attention output of the self-attention mechanism is calculated in step S3, and the formula is as follows:

[0030]

[0031] Among them, d k is the number of channels of X;

[0032] Out=γ×Attention(Q,K,V)+V (7)

[0033] Here, γ is a learning parameter.

[0034] In the preferred solution, in step S3, the softmax function is used to convert the input adjacent elements into probability distribution values ​​between 0 and 1. The calculation formula is as follows:

[0035]

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] The UWB signal NLOS recognition method based on the CNN-SA module of the present invention can effectively identify the NLOS state of the UWB signal, improve the accuracy and reliability of the UWB positioning system, and provide strong support for the development of indoor positioning technology. Multi-layer convolution can extract rich local features from CIR data; the self-attention mechanism can capture the global dependencies in the signal and enhance the feature expression capability; the LeakyReLU activation function can enable the model to better handle negative inputs during training and avoid the gradient vanishing problem; global average pooling reduces feature loss and improves classification effect. Compared with traditional manual feature extraction, it avoids the manual extraction of static parameter features of UWB signals and complex data preprocessing; compared with traditional machine learning methods, it can make full use of channel impulse response (CIR) information; compared with traditional deep learning (CNN, LSTM, etc.), it can take into account the processing of complex local features and global information. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a schematic diagram of the overall structure of the ultra-wideband CNN-SA non-line-of-sight recognition method proposed in the present invention;

[0039] Figure 2 Comparison chart of recognition accuracy curves of different NLOS recognition methods under the same training batch. DETAILED DESCRIPTION

[0040] The following is a clear and complete description of the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0041] The present invention proposes a non-line-of-sight / line-of-sight recognition method based on a CNN-SA module, and the specific steps are as follows: Step 1: Clarify the NLOS propagation interference problem in indoor UWB positioning, and determine the key significance of NLOS / LOS recognition for improving positioning accuracy; Step 2: Design a CNN-SA algorithm, integrate a multi-layer convolutional neural network with a self-attention mechanism, and extract and enhance signal features from CIR data; Step 3: In the process of layer-by-layer extraction of CIR features by the multi-layer convolutional network, introduce the LeakyReLU function to effectively reduce information loss. At the same time, a global average pooling layer is used to replace the fully connected layer to achieve optimization and simplification of the model structure; Step 4: Use an open source dataset to train and test the algorithm and reasonably divide the dataset; Step 5: Analyze the algorithm parameters and compare them with traditional algorithms; This method avoids the problems of traditional recognition methods such as reliance on prior knowledge, complex preprocessing and insufficient feature utilization, and only uses CIR data to reduce dependence on the model; The self-attention mechanism and the multi-layer convolutional network work together to improve recognition accuracy.

[0042] Example 1

[0043] A UWB positioning system NLOS identification method includes the following steps:

[0044] S1. Data Acquisition and Preprocessing: Acquire UWB channel impulse response data (CIR data). This CIR data characterizes multipath propagation between the transmitting and receiving devices, reflecting the signal characteristics and environmental conditions of the propagation path. Preprocess the acquired CIR data, including normalization, to improve data quality and lay the foundation for subsequent processing. The normalization formula is as follows:

[0045]

[0046] Among them, x represents the original data (CIR data), x min and x max Respectively represent the minimum and maximum values ​​of CIR data, x i Indicates normalized CIR data.

[0047] S2, one-dimensional convolutional neural network (1D-CNN) feature extraction: the CIR data x after normalization in step S1 i The input is fed into the 1D-CNN feature extraction module. This feature extraction module consists of multiple convolutional layers, batch normalization layers, and pooling layers. The convolutional layer performs convolution operations on the input data through multiple filters to capture local patterns and features in the data. The mathematical model of the convolution operation is:

[0048]

[0049] Where * represents the convolution operation, N is the number of kernels in the μ-1 layer, is the i-th feature map of μ-1 layer, represents the weight, is the bias of the jth convolution kernel in the μ layer, represents the feature map of the j-th convolution output of the μ layer, f is the LeakReLU activation function, and the expression is:

[0050]

[0051] Among them, β is generally set to 0.01. This effectively solves the "dead ReLU" problem of the LeakReLU function. The batch normalization layer normalizes the data output by the convolutional layer, accelerating the training process and enhancing model stability. The pooling layer uses the max pooling function to downsample the convolutional layer output, reducing the data dimension and computational complexity while retaining basic feature information. The max pooling function expression is as follows:

[0052]

[0053] in, is the value of the t-th neuron corresponding to the i-th feature map in the μ layer, ω is the width of the pooling layer, is the value of the neuron in the μ+1 layer; j represents the output index;

[0054] S3, Self-attention mechanism feature enhancement: The features output by the 1D-CNN feature extraction module are input into the self-attention mechanism. Assume that the input data is X = {x1, x2, x3, ..., x N}, the query Q, key K and value V are calculated respectively by independent convolution functions f1(X), f2(X) and f3(X), as follows:

[0055]

[0056] Among them, W1, W2, and W3 represent weight matrices, which are used to perform linear transformation on the input X and extract different features; B1, B2, and B3 represent bias vectors, which are used to adjust the results after linear transformation.

[0057] The calculated attention output value is:

[0058]

[0059] Among them, d k is the number of channels of X. The softmax function converts the input into a probability distribution whose output values ​​range from 0 to 1 and the sum of the output values ​​is 1.

[0060] The calculation formula of the softmax function in the above steps is as follows:

[0061]

[0062] where x i is the element of the input vector, and e is the base of the natural logarithm. The final output of the self-attention module is:

[0063] Out=γ×Attention(Q,K,V)+V (7)

[0064] Where γ is a learning parameter. After the self-attention layer, layer normalization (LN) and dropout layers are added to enhance the stability and generalization ability of the model.

[0065] S4, classification output: The features output by the self-attention module are input into the output layer, and the features are mapped to classification labels through the softmax function. The output data is the predicted probability obtained in the line-of-sight (LOS) and non-line-of-sight (NLOS) scenarios to achieve NLOS classification and recognition.

[0066] Data collection and preprocessing: Using a DW1000 UWB module compliant with the IEEE 802.15.4-2011 IR-UWB standard, we conducted 3,000 line-of-sight (LOS) channel measurements and 3,000 non-line-of-sight (NLOS) channel measurements in seven different indoor environments (Office 1, Office 2, a small apartment, a small workshop, a kitchen with a living room, a bedroom, and a boiler room), acquiring a total of 42,000 measurement data. The collected CIR data was cleaned to remove outliers and noise, and then normalized using the aforementioned normalization formula. The preprocessed data was divided into training, test, and validation sets with a 70%, 20%, and 10% ratio.

[0067] Model Construction and Training: Build a CNN-SA model, including the 1D-CNN feature extraction module, the self-attention module, and the output layer. In the 1D-CNN feature extraction module, appropriately set the parameters of the convolutional, batch normalization, and pooling layers, such as the kernel size, stride, and pooling window size. Each layer uses a different kernel size, typically 3, 5, or 7, and the stride is typically set to 1.

[0068] The model was trained using the training set, with an initial learning rate of 0.001 and a training batch size of 64. During the training process, the model parameters were continuously adjusted through the backpropagation algorithm to minimize the loss function, allowing the model to gradually learn the characteristic patterns of NLOS / LOS in UWBCIR data.

[0069] Model testing and optimization: During the training process, the model is regularly tested using the test set to observe changes in model performance indicators, such as accuracy and loss. Based on the test results, the model parameters are fine-tuned, such as adjusting the learning rate and increasing or decreasing the number of training rounds, to improve model performance.

[0070] After training, the model was fully evaluated using the validation set, and performance metrics such as accuracy, precision, recall, specificity, and F1 score were calculated. The CNN-SA model of the present invention was compared with other comparative models (such as CNN, LSTM, CNN-LSTM, etc.) on the same dataset to verify the superiority of the model of the present invention.

[0071] Through the above specific implementation methods, the UWB signal NLOS recognition method based on the CNN-SA module of the present invention can effectively identify the NLOS state of the UWB signal, improve the accuracy and reliability of the UWB positioning system, and provide strong support for the development of indoor positioning technology. Figure 2 As shown in the figure, the classification and recognition accuracy is significantly improved compared with existing algorithms.

[0072] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A NLOS identification method for a UWB positioning system, characterized in that: The steps include: S1. Data preprocessing: collecting UWB channel impulse response data and normalizing the UWB channel impulse response data; S2. One-dimensional CNN feature extraction: Extract and output one-dimensional CNN features using a feature extraction module of a one-dimensional convolutional neural network, wherein the feature extraction module has a convolution layer and a batch normalization layer, and includes the following steps: S21: Input the data preprocessed in step S1 into the convolution layer for convolution operation, capture the local patterns and features of the preprocessed data, and import them into the data model; S22, normalize the data output by the convolution layer through the batch normalization layer; S3, self-attention mechanism feature enhancement: The features output by the feature extraction module of the one-dimensional convolutional neural network are input into the self-attention mechanism, and the input data is set to X = {x1, x2, x3, ..., x N }, and calculate the query Q, key K and value V respectively through independent convolution functions f1(X), f2(X) and f3(X), and then calculate the attention output of the self-attention mechanism; S4, classification output: The features output by the self-attention module are input into the output layer, the features are mapped to classification labels through the softmax function, and the prediction probabilities obtained in the line-of-sight and non-line-of-sight scenarios are output to achieve NLOS classification and recognition.

2. The NLOS identification method of the UWB positioning system according to claim 1, characterized in that: The UWB channel impulse response data is cleaned and then normalized.

3. The NLOS identification method of the UWB positioning system according to claim 2, characterized in that: The cleaning includes noise removal and abnormal data processing.

4. The NLOS identification method of the UWB positioning system according to claim 1, characterized in that: The feature extraction module of the one-dimensional convolutional neural network includes a convolution layer, a batch normalization layer and a pooling layer, and the convolution layer performs a convolution operation on the input data through multiple filters.

5. The NLOS identification method of the UWB positioning system according to claim 1, characterized in that: The data model in step S21 is as follows: Where * represents the convolution operation, N is the number of kernels in the μ-1 layer, is the i-th feature map of μ-1 layer, represents the weight, is the bias of the jth convolution kernel in the μ layer, represents the feature map of the j-th convolution output of the μ layer, f is the LeakRELU activation function, and the expression of f is as follows: Where β = 0.01, 6. The NLOS identification method of the UWB positioning system according to claim 1, characterized in that: The normalization process in step S22 is calculated using the following formula: in, is the value of the t-th neuron corresponding to the i-th feature map in the μ layer, ω is the width of the pooling layer, is the value of the neuron in the μ+1 layer.

7. The NLOS identification method of the UWB positioning system according to claim 1, characterized in that: In step S3, the query Q, key K, and value V are calculated respectively by independent convolution functions f1(X), f2(X), and f3(X), as follows: Among them, W1, W2, and W3 represent weight matrices; B1, B2, and B3 represent bias vectors.

8. The NLOS identification method of the UWB positioning system according to claim 7, characterized in that: Step S3 calculates the attention output of the self-attention mechanism, and the formula is as follows: Among them, d k is the number of channels of X; Out=γ×Attention(Q,K,V)+V (7) Here, γ is a learning parameter.

9. The NLOS identification method of the UWB positioning system according to claim 8, characterized in that: In step S3, the softmax function is used to convert the input adjacent elements into probability distribution values ​​between 0 and 1. The calculation formula is as follows:

Images