A device and method for uwb l / o nlo identification

Abstract

The application provides a device and an algorithm for UWB line-of-sight / non-line-of-sight identification, the device comprising a UWB communication module and a TinyML supporting calculation module, the algorithm being that after complete CIR data and all channel characteristics are collected by the UWB communication module, key CIR data and channel characteristics are screened through feature selection to form an input data set, a pre-training model is built and weights are pre-trained, the input layer and feature extraction block weights of the pre-training model are frozen, a self-attention mechanism is added and the model is retrained, so that the model has effective representation for UWB line-of-sight / non-line-of-sight identification, finally, the retrained model is deployed on the TinyML supporting calculation module through model compression, quantization and TinyML deployment to realize UWB line-of-sight / non-line-of-sight identification. The application can realize fast, effective and robust UWB line-of-sight / non-line-of-sight identification on an embedded device with limited computing resources, and is helpful to further reduce UWB ranging error and improve positioning accuracy.

Description

Technical Field

[0001] This invention belongs to the field of wireless signal analysis, specifically relating to a device and method for UWB line-of-sight / non-line-of-sight identification. Background Technology

[0002] Ultra-wideband (UWB) is a pulse radio communication technology that enables ranging between two distinct UWB communication modules—a UWB tag and a UWB base station—by recording flight timestamps. UWB offers advantages such as strong penetration, low power consumption, good multipath resistance, high security, and low system complexity, providing centimeter-level accuracy in ranging. UWB chips contain various channel characteristics and related parameters, as well as the Channel Impulse Response (CIR), which reflects signal propagation quality and can be used to characterize the difference between line-of-sight and non-line-of-sight propagation.

[0003] In UWB-based wireless positioning systems, UWB line-of-sight / non-line-of-sight propagation is one of the main causes of anomalous measurement errors and affects the confidence level of positioning accuracy. UWB line-of-sight / non-line-of-sight identification is beneficial for rationally selecting base station positioning strategies or effectively compensating for observation errors under mixed line-of-sight and non-line-of-sight conditions. Therefore, UWB line-of-sight / non-line-of-sight identification is crucial for reliable and accurate wireless positioning estimation. UWB line-of-sight / non-line-of-sight identification is a binary classification task that largely relies on establishing strong representational relationships between input features and labels.

[0004] Currently, numerous machine learning and deep learning methods have achieved state-of-the-art results in UWB line-of-sight / non-line-of-sight recognition. However, effective and robust UWB line-of-sight / non-line-of-sight recognition remains a challenge on resource-constrained edge devices. Tiny Machine Learning (TinyML) is the intersection of IoT and machine learning, enabling the implementation of machine learning algorithms on low-resource, low-power microcontroller units (MCUs). Since many UWB positioning systems lack sufficient computing resources, achieving robust and effective UWB line-of-sight / non-line-of-sight recognition on resource-constrained embedded devices is crucial for wireless positioning systems. Summary of the Invention

[0005] To address the aforementioned issues, this invention discloses a device and method for UWB line-of-sight / non-line-of-sight identification. This invention not only enables high-performance UWB line-of-sight / non-line-of-sight identification on PCs or workstations without computing power limitations, but also enables fast, effective, and robust UWB line-of-sight / non-line-of-sight identification on embedded devices with limited computing resources. It is suitable for the application needs of many low-cost wireless positioning systems and helps to further reduce UWB ranging errors and improve positioning accuracy.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] A UWB line-of-sight / non-line-of-sight recognition device includes a UWB communication module and a computing module supporting TinyML.

[0008] The UWB communication module collects complete original CIR channel features and all channel features, and obtains relevant parameters from its internal registers, requiring that the line-of-sight or non-line-of-sight status between the UWB tag and the UWB base station is known.

[0009] The aforementioned TinyML-enabled computing module meets the minimum configuration requirements for deploying UWB line-of-sight / non-line-of-sight recognition methods.

[0010] A method for UWB line-of-sight / non-line-of-sight recognition includes:

[0011] 1) Feature selection;

[0012] 2) Build a pre-trained model and pre-train the weights;

[0013] 3) Freeze the input layer and feature extraction block weights of the pre-trained model, add a self-attention mechanism, and retrain the model to achieve UWB line-of-sight / non-line-of-sight recognition;

[0014] 4) Model compression, quantization, and TinyML deployment work together to achieve UWB line-of-sight / non-line-of-sight recognition on a TinyML-supporting computing module.

[0015] Furthermore, the feature selection described in step 1) is used to filter out key CIR signal features and key channel features from the original CIR signal and all channel features collected by the UWB communication module to form the input dataset.

[0016] Using feature selection to filter out key CIR data and key channel features is highly relevant to UWB line-of-sight / non-line-of-sight identification tasks.

[0017] Furthermore, the pre-trained model and pre-trained weights described in step 2) are used to construct a complex implicit relationship between the input dataset obtained by feature selection and UWB line-of-sight / non-line-of-sight recognition through a neural network model; the pre-trained model is pre-trained using a supervised training method of the neural network to obtain the pre-trained weights.

[0018] The input of the pre-trained model is the input dataset obtained by the feature selection method, and the output is the UWB line-of-sight / non-line-of-sight recognition result; the model is a multilayer perceptron model including an input layer, a feature extraction module, a classifier module and an output layer; the training labels are the binary labels corresponding to the line-of-sight / non-line-of-sight states of each data point in the known input dataset; the pre-trained model uses a supervised training method of neural networks to train the weights.

[0019] Furthermore, step 3) involves freezing the input layer and feature extraction block weights of the pre-trained model to retain the effective features extracted by the pre-trained model from the input dataset obtained using feature selection, along with the dimensionality-reduced representation.

[0020] The addition of a self-attention mechanism and retraining of the model are used to re-evaluate the effective features extracted from the input dataset obtained by feature selection and the dimensionality reduction representation of the pre-trained model and to learn effective classification representations. The model is retrained by freezing the input layer and feature extraction block weights of the pre-trained model and adding the attention mechanism through a supervised training method of neural networks, so that it has effective representations for UWB line-of-sight / non-line-of-sight recognition.

[0021] The process of adding a self-attention mechanism and retraining the model involves first freezing the input layer and feature extraction block weights of the pre-trained model; secondly, adding a self-attention mechanism between the feature extraction block and the classifier module of the pre-trained model, and retraining the model weights using a supervised neural network training method; wherein, the training labels are the binary labels corresponding to the line-of-sight or non-line-of-sight states for each data point in the known input dataset.

[0022] The mathematical description of the self-attention mechanism is as follows:

[0023] ;

[0024] ;

[0025] Among them, query (Q), key (K), and value (V) are the three key elements of the self-attention mechanism;

[0026] Where X is the input data, and f1, f2, and f3 are the mapping rules corresponding to Q, K, and V, respectively;

[0027] Where W1, W2, and W3 are respectively calculated linearly... The linear transformation matrices that transform a vector into a new vector have the same structure, are independent of each other, and are obtained through model training;

[0028] Here, B1, B2, and B3 are bias vectors that do not change the structure of the Q, K, and V vectors, and are obtained through model training.

[0029] Furthermore, the model compression, quantization, and TinyML deployment described in step 4) are used to perform volume compression, weight quantization, and TinyML deployment on models with effective representations of UWB line-of-sight / non-line-of-sight recognition, and deploy them on devices that support TinyML.

[0030] Among them, model compression uses volume compression to compress the original neural network model into a TFLite file to reduce the size of the original model and make it easier to deploy in TinyML.

[0031] Quantization is performed as weight quantization, which quantizes the TFLite file of the compressed model into all integers, making it more suitable for online inference on computing modules that support TinyML and further reducing the model size;

[0032] Using TinyML for deployment, the TFLite file, which has been compressed in size and quantized in weight, is converted into C / C++ source code and burned onto a computing module that supports TinyML to complete the deployment.

[0033] The working principle of this invention is:

[0034] After collecting complete CIR data and all channel features using a UWB communication module, key CIR data and channel features are selected through feature selection to form the input dataset. A pre-trained model is then built and its weights are pre-trained. The input layer and feature extraction block weights of the pre-trained model are frozen, a self-attention mechanism is added, and the model is retrained to provide an effective representation for UWB line-of-sight / non-line-of-sight recognition. Finally, through model compression, quantization, and TinyML deployment, the retrained model is deployed on a TinyML-enabled computing module to achieve UWB line-of-sight / non-line-of-sight recognition.

[0035] The beneficial effects of this invention are:

[0036] (1) The present invention can realize fast, effective and robust UWB line-of-sight / non-line-of-sight identification on embedded devices with limited computing resources. It is suitable for the application needs of many low-cost wireless positioning systems and helps to further reduce UWB ranging error and improve positioning accuracy.

[0037] (2) The UWB communication module can collect complete original CIR channel features and all channel features, and obtain relevant parameters of its internal registers. Feature selection is used to filter key CIR data and key channel features from the original CIR data and all channel features collected by the UWB communication module to form the input dataset. Compared with using complete original CIR information, the feature selection method significantly reduces the dimensionality of the input dataset without losing signal features closely related to UWB line-of-sight / non-line-of-sight recognition. A pre-trained model is used to construct the complex relationship between the input dataset and the UWB line-of-sight / non-line-of-sight recognition task. The constructed pre-trained model is pre-trained using a supervised neural network training method to obtain pre-training weights. By freezing the input layer and feature extraction block weights in the pre-trained model, the effective features extracted from the input dataset obtained using feature selection and the dimensionality-reduced representations are retained. A self-attention mechanism is added to re-evaluate the effective features extracted from the input dataset and the dimensionality-reduced representations, and to learn effective classification representations. The model with frozen input layer and feature extraction block weights, along with the added attention mechanism, is retrained using a supervised neural network training method, resulting in effective representations for UWB line-of-sight / non-line-of-sight recognition. This enables high-performance UWB line-of-sight / non-line-of-sight recognition on PCs or workstations without computational resource constraints. Model compression, quantization, and TinyML deployment allow the UWB line-of-sight / non-line-of-sight recognition method to run efficiently, robustly, and effectively on computationally limited embedded devices with low power consumption. Attached Figure Description

[0038] Figure 1 This is an overall schematic diagram of the UWB line-of-sight / non-line-of-sight recognition method proposed in this invention.

[0039] Figure 2 This is a schematic diagram of an embodiment of the present invention.

[0040] Figure 3 This is a schematic diagram illustrating the use of feature selection to obtain the 57-dimensional input dataset in step one of the implementation examples.

[0041] Figure 4 This is a schematic diagram of the neural network model structure in steps two and three of the implementation example.

[0042] Figure 5 A schematic diagram of step four in implementing the example.

[0043] Figures 6-7 This document presents performance metrics for UWB line-of-sight / non-line-of-sight recognition on a PC with unrestricted computing power, and compares its performance with various strategies.

[0044] Figures 8-9 This document describes the performance metrics of UWB line-of-sight / non-line-of-sight recognition on an Arduino Nano module, and compares its performance with various other strategies.

[0045] Figure 10 The device parameters for the Arduino Nano module used in the implementation example.

[0046] Figure 11 This is a schematic diagram of the line-of-sight propagation and non-line-of-sight propagation states of the UWB ranging module. Detailed Implementation

[0047] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0048] As shown in the figure, the UWB line-of-sight / non-line-of-sight recognition method of the present invention includes: 1) feature selection, 2) building a pre-trained model and pre-training weights, 3) freezing the input layer and feature extraction block weights of the pre-trained model, adding a self-attention mechanism and retraining the model, and 4) model compression, quantization and TinyML deployment.

[0049] The present invention also introduces a UWB line-of-sight / non-line-of-sight recognition device, including a UWB communication module and a computing module supporting TinyML.

[0050] The UWB communication module is required to be able to collect complete original CIR channel characteristics and all channel characteristics, and to obtain relevant parameters of its internal registers.

[0051] The aforementioned TinyML-supporting computing module is required to meet the minimum configuration requirements for deploying UWB line-of-sight / non-line-of-sight recognition methods;

[0052] The feature selection is used to filter out key CIR signal features and key channel features from the original CIR signals and all channel features collected by the UWB communication module to form the input dataset.

[0053] The pre-trained model and pre-trained weights are used to construct a complex implicit relationship between the input dataset obtained by feature selection and UWB line-of-sight / non-line-of-sight recognition through a neural network model; the pre-trained model is pre-trained using a supervised training method of the neural network to obtain the pre-trained weights.

[0054] The aforementioned freezing of the input layer and feature extraction block weights of the pre-trained model is used to retain the effective features extracted by the pre-trained model from the input dataset obtained by feature selection and the dimensionality-reduced representation.

[0055] The addition of a self-attention mechanism and retraining of the model are used to re-evaluate the effective features extracted from the input dataset obtained by feature selection and the dimensionality reduction representation of the pre-trained model and to learn effective classification representations. The model is retrained by freezing the input layer and feature extraction block weights of the pre-trained model and adding the attention mechanism through a supervised training method of neural networks, so that it has effective representations for UWB line-of-sight / non-line-of-sight recognition.

[0056] The aforementioned model compression, quantization, and TinyML deployment are used to perform volume compression, weight quantization, and TinyML deployment on models with effective representations for UWB line-of-sight / non-line-of-sight recognition, and deploy them on TinyML-enabled devices to achieve UWB line-of-sight / non-line-of-sight recognition.

[0057] Example 1:

[0058] The UWB line-of-sight / non-line-of-sight recognition method specifically includes the following steps:

[0059] Step 1, Feature Selection: From the raw CIR signal and all channel features acquired by the UWB device, 50 key CIR signal features starting from the FP_INDEX index of the FP_INDEX register in the UWB device, along with 7 key channel features, are selected to form a 57-dimensional input dataset, such as... Figure 3 As shown;

[0060] Step 2, build a pre-trained model and pre-train the weights: such as Figure 4 As shown, 1) a multilayer perceptron model is constructed as a pre-trained model, comprising an input layer consisting of a linear fully connected layer with 57 neurons, a feature extraction block consisting of a ReLU fully connected layer and a batch normalization layer with 16 neurons, a classifier module consisting of a ReLU fully connected layer and a batch normalization layer with 64 neurons, a ReLU fully connected layer and a batch normalization layer with 16 neurons, and an output layer consisting of a ReLU fully connected layer with 2 neurons; 2) the constructed multilayer perceptron model is pre-trained using the Adam (Adaptive Moment Estimation) optimizer and the cross-entropy loss function to obtain pre-trained weights; the training labels are the binary labels corresponding to the line-of-sight / non-line-of-sight states for each data point in the known input dataset.

[0061] Step 3: Freeze the input layer and feature extraction block weights of the pre-trained model, add a self-attention mechanism, and retrain the model: 1) Freeze the input layer and feature extraction block weights of the pre-trained model; 2) Add a self-attention mechanism between the feature extraction block and the classifier module of the pre-trained model; wherein, the Q, K, and V vectors in the self-attention mechanism are composed of three ReLU fully connected layers with 16 neurons each; 3) Retrain the model after freezing the input layer and feature extraction block weights of the pre-trained model and adding the attention mechanism using the Adam optimizer and cross-entropy loss function, so that it has an effective representation for UWB line-of-sight / non-line-of-sight recognition;

[0062] Step 4, Model compression, quantization, and TinyML deployment: (e.g.) Figure 5 As shown, 1) the original TensorFlow model is compressed into a TFLite file; 2) the parameters of the TFLite file of the compressed model are converted from "float32" format to "int8" format for full integer quantization; 3) the TFLite file after volume compression and weight quantization is converted into C / C++ source code and burned onto the Arduino Nano module to complete the deployment.

[0063] As can be seen from the above embodiments, UWB line-of-sight / non-line-of-sight recognition can be achieved through feature selection, building a pre-trained model and pre-training weights, freezing the input layer and feature extraction block weights of the pre-trained model, adding a self-attention mechanism, and retraining the model. UWB line-of-sight / non-line-of-sight recognition can be achieved on devices supporting TinyML through model compression, quantization, and TinyML deployment. The UWB communication module can collect complete raw CIR channel features and all channel features, and obtain relevant parameters from its internal registers. Feature selection is used to filter key CIR data and key channel features from the raw CIR data and all channel features collected by the UWB communication module to form the input dataset. Compared with using complete raw CIR information, the feature selection method significantly reduces the dimensionality of the input dataset without losing signal features closely related to UWB line-of-sight / non-line-of-sight recognition. The pre-trained model is used to construct the complex relationship between the input dataset and the UWB line-of-sight / non-line-of-sight recognition task. The constructed pre-trained model is pre-trained using a supervised neural network training method to obtain pre-trained weights. By freezing the input layer and feature extraction block weights in the pre-trained model, the effective features extracted from the input dataset obtained using feature selection and the dimensionality-reduced representations are retained. A self-attention mechanism is added to re-evaluate the effective features extracted from the input dataset and the dimensionality-reduced representations, and to learn effective classification representations. The model with frozen input layer and feature extraction block weights, along with the added attention mechanism, is retrained using a supervised neural network training method, resulting in effective representations for UWB line-of-sight / non-line-of-sight recognition. This enables high-performance UWB line-of-sight / non-line-of-sight recognition on PCs or workstations without computational resource constraints. Model compression, quantization, and TinyML deployment allow the UWB line-of-sight / non-line-of-sight recognition method to run efficiently, robustly, and effectively on computationally limited embedded devices with low power consumption.

[0064] In summary, the UWB line-of-sight / non-line-of-sight identification method and device proposed in this invention can not only achieve high-performance UWB line-of-sight / non-line-of-sight identification on PCs or workstations, but also achieve fast, effective, and robust UWB line-of-sight / non-line-of-sight identification on embedded devices with limited computing resources. It is suitable for the application needs of many low-cost wireless positioning systems and helps to further reduce UWB ranging errors and improve positioning accuracy.

[0065] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.

Claims

1. A method for UWB line-of-sight / non-line-of-sight recognition, characterized in that: Includes the following steps: 1) Feature selection: Used to filter out key CIR signal features and key channel features from the raw CIR signals and all channel features acquired by the UWB communication module to form the input dataset; 2) Build a pre-trained model and pre-train the weights; 3) Freeze the input layer and feature extraction block weights of the pre-trained model, add a self-attention mechanism, and retrain the model to achieve UWB line-of-sight / non-line-of-sight recognition; Specifically as follows: The aforementioned freezing of the input layer and feature extraction block weights of the pre-trained model is used to retain the effective features extracted by the pre-trained model from the input dataset obtained by feature selection and the dimensionality-reduced representation. The addition of a self-attention mechanism and retraining of the model are used to re-evaluate the effective features extracted from the input dataset obtained by feature selection and the dimensionality reduction representation of the pre-trained model and to learn effective classification representations. The model is retrained by adding an attention mechanism after freezing the input layer and feature extraction block weights of the pre-trained model through a supervised training method of neural networks, so that it has effective representations for UWB line-of-sight / non-line-of-sight recognition. The process of adding a self-attention mechanism and retraining the model involves first freezing the input layer and feature extraction block weights of the pre-trained model; secondly, adding a self-attention mechanism between the feature extraction block and the classifier module of the pre-trained model, and retraining the model weights using a supervised neural network training method; wherein, the training labels are the binary labels corresponding to the line-of-sight or non-line-of-sight states for each data point in the known input dataset. The mathematical description of the self-attention mechanism is as follows: （1）； （2）； Among them, query (Q), key (K), and value (V) are the three key elements of the self-attention mechanism; Where X is the input data, and f1, f2, and f3 are the mapping rules corresponding to Q, K, and V, respectively; Where W1, W2, and W3 are respectively calculated linearly... The linear transformation matrices that transform a vector into a new vector have the same structure, are independent of each other, and are obtained through model training; Where B1, B2, and B3 are bias vectors that do not change the structure of Q, K, and V vectors and are obtained through model training. 4) Model compression, quantization, and TinyML deployment work together to achieve UWB line-of-sight / non-line-of-sight recognition on a TinyML-supporting computing module.

2. The UWB line-of-sight / non-line-of-sight recognition method according to claim 1, characterized in that: Step 2) describes the pre-trained model and pre-trained weights, which are used to construct a complex implicit relationship between the input dataset obtained by feature selection and UWB line-of-sight / non-line-of-sight recognition through a neural network model; the pre-trained model is pre-trained using a supervised training method of the neural network to obtain the pre-trained weights, as follows; The input of the pre-trained model is the input dataset obtained by the feature selection method, and the output is the UWB line-of-sight / non-line-of-sight recognition result; the model is a multilayer perceptron model including an input layer, a feature extraction module, a classifier module and an output layer; the training labels are the binary labels corresponding to the line-of-sight / non-line-of-sight states of each data point in the known input dataset; the pre-trained model uses a supervised training method of neural networks to train the weights.

3. The UWB line-of-sight / non-line-of-sight recognition method according to claim 1, characterized in that: Step 4) describes model compression, quantization, and TinyML deployment, which are used to perform volume compression, weight quantization, and TinyML deployment on models with effective representations of UWB line-of-sight / non-line-of-sight recognition, and deploy them on devices that support TinyML. Among them, model compression uses volume compression to compress the original neural network model into a TFLite file to reduce the size of the original model and make it easier to deploy in TinyML. Quantization is performed as weight quantization, which quantizes the TFLite file of the compressed model into all integers, making it more suitable for online inference on computing modules that support TinyML and further reducing the model size; Using TinyML for deployment, the TFLite file, which has been compressed in size and quantized in weight, is converted into C / C++ source code and burned onto a computing module that supports TinyML to complete the deployment.

4. A UWB line-of-sight / non-line-of-sight recognition device, used to perform the method according to any one of claims 1-3, characterized in that: Includes a UWB communication module and a computing module that supports TinyML; The UWB communication module collects complete original CIR channel features and all channel features, and obtains relevant parameters of its internal registers, requiring that the line-of-sight or non-line-of-sight state between the UWB tag and the UWB base station is known. The aforementioned TinyML-supporting computing module meets the minimum configuration requirements for deploying UWB line-of-sight / non-line-of-sight recognition algorithms.

Images