A method for UWB NLOS identification and ranging error compensation based on feature similarity

Abstract

The application provides a UWB NLOS recognition and ranging error compensation method based on feature similarity, comprising the following steps: (1) acquiring distance observation values, channel propagation states, radio frequency channel features and CIR signals, including reference CIR information and sample CIR information, by using a UWB base station and a UWB tag; (2) calculating the channel similarity between the reference CIR and the sample CIR, and constructing a channel similarity feature vector; (3) constructing an NLOS recognition deep learning model based on the mode input of the radio frequency channel features and the channel similarity feature vector; and (4) adjusting the parameter structure and converting the NLOS recognition deep learning model into an error compensation model; the application calculates the feature similarity, enhances the feature representation capability, and thus improves the accuracy of UWB NLOS recognition; the application calculates the similarity of the CIR sequence, effectively reduces the input parameters, and thus reduces the calculation complexity of the model and improves the inference efficiency of the model; the application also adjusts the model parameters, simultaneously realizes the recognition and error compensation functions of the UWB NLOS, improves the utilization efficiency of the model, and reduces the size of the model.

Description

Technical Field

[0001] This invention belongs to the field of UWB ranging and positioning technology, specifically relating to a method for UWB NLOS identification and ranging error compensation based on feature similarity. Background Technology

[0002] While UWB ranging technology is renowned for its high precision, it is still susceptible to various factors in practical applications, including multipath effects, NLOS propagation, signal attenuation, clock synchronization errors, noise, hardware errors, temperature variations, bandwidth limitations, antenna position errors, and channel delay spread, all of which contribute to ranging errors. By understanding and analyzing these error sources, corresponding compensation and correction measures can be implemented to improve the accuracy and reliability of UWB ranging.

[0003] NLOS propagation refers to a situation where a wireless signal encounters obstacles during transmission between the transmitter and receiver, preventing the signal from reaching the receiver directly. This contrasts sharply with LOS propagation, where the signal travels directly from the transmitter to the receiver without any physical obstacles. NLOS propagation is particularly common in urban environments, indoors, or other complex environments, and has a significant impact on the performance and reliability of wireless networks. Because signals in NLOS conditions must bypass obstacles or propagate through reflection, diffraction, etc., they typically experience significant attenuation, and multipath effects are more pronounced under these conditions, which can lead to signal interference and affect the quality of the received signal.

[0004] In NLOS (Normally Unstable Oriented) wireless positioning, the uncertainty of signal propagation paths can reduce positioning accuracy. To improve communication quality and reliability under NLOS conditions, advanced signal processing techniques such as multipath diversity reception, spatial diversity, and frequency diversity can be employed. Facing the challenges posed by NLOS propagation, the design and optimization of wireless networks require a comprehensive consideration of various technologies and strategies to overcome the problems in the NLOS environment, thereby improving the overall reliability and efficiency of wireless communication. Therefore, the identification and error compensation of UWB NLOS paths are of great significance. Summary of the Invention

[0005] To address the aforementioned problems, this invention discloses a UWB NLOS identification and ranging error compensation method based on feature similarity. Through a dual-input attention neural network, it evaluates the input channel features and CIR similarity features, identifying NLOS first and then compensating for it. This significantly improves the ranging accuracy of UWB in NLOS environments and effectively suppresses the position estimation error of the target to be located. Furthermore, the model can switch between NLOS identification and ranging error compensation tasks, greatly improving the model's utilization efficiency.

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

[0007] A method for UWB NLOS identification and ranging error compensation based on feature similarity includes the following steps:

[0008] A: Channel Feature Acquisition. Ultra-wideband (UWB) falls under the radio frequency (RF) domain, allowing for the extraction of various channel characteristics from UWB sample data. The extracted UWB channel features include not only statistical information on observed distance values ​​such as distance, maximum noise level, and noise standard deviation, but also features related to the first path. The first path is the shortest propagation path from the transmitter to the receiver, typically representing the path with the shortest signal propagation time. In complex environments, the first path is usually the NLOS (Low-Noise-Side-Operate) path. Furthermore, the maximum noise level and noise standard deviation show significant differences in both LOS and NLOS environments. These channel features, along with the CIR (Channel Impulse Response) feature vector, are used as input parameters to achieve NLOS identification and ranging error compensation. CIR, representing the channel impulse response, measures the channel's response to impulse signals (i.e., instantaneous unit energy pulses). It is a crucial parameter characterizing channel transmission characteristics, reflecting signal changes during channel propagation, including NLOS, attenuation, delay, and multipath effects.

[0009] B: Calculation of CIR similarity feature vectors. The calculation of CIR similarity feature vectors involves several key steps, as follows:

[0010] B1: Reference Sampling Matching. Find the nearest reference sample by calculating the distance.

[0011] B2: Data Filtering. Data is cleaned by averaging and thresholding to remove outliers and noise, thereby improving data quality. The filtered data is then used to build a model for NLOS identification and compensation.

[0012] B3: Extract the CIR sequence near FP_Index from the measured channel characteristic data.

[0013] B4: Perform digital padding. The CIR data near FP_Index is too sparse to describe the similarity of CIR shapes. Therefore, it is necessary to pad the data at equal intervals between every two adjacent CIR data points. The slope between each pair of padded data points is used to describe the shape similarity. FP_Index is the index of the first path peak, referring to the position of the peak corresponding to the first signal path reaching the receiver in the received signal within the signal data array.

[0014] B5: Calculate a similarity vector representing the numerical values ​​of the CIR sequences using the CIR sequences before padding.

[0015] B6: Calculate the similarity vector that indirectly describes the shape using the padded CIR sequence, and then combine the error of the reference sampling and the ranging error of the sample to form a complete CIR similarity vector.

[0016] B7: Divide the processed dataset into training and testing sets. When dividing the dataset, ensure that the data is mixed evenly so that it can be used in the model training and evaluation process.

[0017] Similarity features in experimental data feature engineering mainly include cosine similarity, Spearman correlation coefficient, Pearson correlation coefficient, Canberra distance, standardized Euclidean distance, Jensen-Shannon divergence, and Hellinger distance. The type of similarity feature is selected based on the actual scenario, directly calculating the similarity vector between the CIR sequence near FP_Index and the reference sample. This serves as the numerical similarity information of the CIR vector. The shape difference between the prior reference CIR vector and the working sample CIR vector is also a useful reference for characterizing the LOS and NLOS channels. For a dimensionless one-dimensional vector like the CIR signal, its shape feature can be reflected using a slope vector. The specific calculation method is as follows:

[0018] (1) Cosine similarity, the cosine distance in n-dimensional space is shown in the following formula.

[0019]

[0020] Where x and y are two vectors to be compared, x i y i Let i be the i-th numerical component of the vector.

[0021] (2) Spearman correlation coefficient: The difference d is obtained by subtracting the corresponding elements of the two observed variables. The Spearman correlation coefficient can be calculated by the following formula.

[0022]

[0023] Where, d i Let n be the difference between corresponding elements of the two vectors to be compared, where n is the dimension of the vector.

[0024] (3) Pearson correlation coefficient. The formula for calculating the Pearson correlation coefficient in n-dimensional space is shown below.

[0025]

[0026] Where σ(x) and σ(y) are the variances of the two vectors to be compared, respectively, and cov(x, y) is the covariance between the two vectors.

[0027] (4) Canberra distance, the Canberra distance in n-dimensional space is shown in the following formula.

[0028]

[0029] Where, x i y i Let be the i-th numerical component of the vector to be compared.

[0030] (5) Standard Euclidean similarity. The formula for calculating standard Euclidean similarity can be expressed as follows.

[0031]

[0032] Where d(x, y) is the standard Euclidean distance between the two vectors.

[0033] (6) Jensen-Shannon divergence: Jensen-Shannon divergence can effectively measure the similarity between two probability distributions, overcoming the asymmetry problem in KL divergence and providing a standard way to quantify the difference between probability distributions. The following formula is the calculation formula for Jensen-Shannon divergence.

[0034]

[0035] Where KL is the KL divergence between the two directions.

[0036] (7) Hellinger distance is a method for measuring the difference between two probability distributions. It is used to measure the similarity between two probability distributions. For discrete probability distributions, the Hellinger distance is defined as shown in the following formula.

[0037]

[0038] Here, p and q are two probability distributions defined on the same event space, and p = {p1, p2, ..., pq} n}, q={q1,q2,...,q n}

[0039] C: Building an NLOS identification and error compensation model

[0040] Based on the design of CIR similarity features, a two-layer attention neural network is built. The two input layers are respectively input to the channel feature vector and the CIR similarity feature vector to construct an NLOS recognition deep learning model. In order to avoid the problem that the same attention mechanism is not good at evaluating different types of inputs, the attention mechanism is evaluated separately for the two sets of inputs. The purpose is to improve the model's efficiency in identifying and processing key information in the input data.

[0041] Adjusting the parameter structure allows for the construction of an error compensation model. The adjusted parameters include the optimizer and loss function, enabling the switching between a classification model and a regression model. The difference lies in the output layer structure: the classification model's output layer has two neurons and two bias terms; the compensation model's output layer has only one neuron and one bias term.

[0042] Step D: By using NLOS to identify deep learning models and error compensation models, the identification of UWB NLOS paths and the compensation of ranging errors are achieved.

[0043] The beneficial effects of this invention are as follows:

[0044] (1) This invention uses a dual-input attention neural network to evaluate the input channel features and CIR similarity features, and first identifies and then compensates for NLOS, which greatly improves the ranging accuracy of UWB in NLOS environment and can significantly suppress the position estimation error of the target to be located.

[0045] (2) The present invention also improves the utilization efficiency of the model and reduces the size of the model by adjusting the model parameters to simultaneously realize the identification and error compensation functions of UWB NLOS. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the overall framework of the UWB NLOS identification and error compensation method proposed in this invention.

[0047] Figure 2 This is a schematic diagram of the overall process of the UWB NLOS recognition model proposed in this invention.

[0048] Figure 3 This is a schematic diagram of the overall process of the UWB NLOS error compensation model proposed in this invention.

[0049] Figure 4 This is a schematic diagram of the UWB NLOS identification and error compensation implementation method proposed in this invention patent. Detailed Implementation

[0050] 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.

[0051] like Figure 1 As shown, the NLOS identification and ranging error compensation algorithm based on UWB channel characteristics and CIR similarity characteristics described in this invention includes the following key steps:

[0052] (1) Use UWB sampling equipment to acquire distance observations, radio frequency channel characteristics and CIR signals (including reference CIR information and sampled CIR information).

[0053] (2) Calculate the channel similarity between the reference CIR and the sampled CIR, and construct the channel similarity feature vector.

[0054] (3) Based on the pattern input of radio frequency channel features and channel similarity feature vectors, a deep learning model for NLOS recognition is constructed.

[0055] (4) Adjust the parameter structure to transform the NLOS recognition model into an error compensation model.

[0056] Overall, the invention mainly includes two modules: acquiring input parameters and building a model.

[0057] The specific implementation steps include:

[0058] Step A: Experimental data feature engineering, which involves extracting channel features and calculating similarity feature vectors from the collected experimental data.

[0059] The specific steps of experimental data feature engineering are as follows:

[0060] Step A1: Reference sample matching. By calculating the distance, find the reference sample that is closest to the sample, and pair the channel characteristics and CIR sequence of the sample with the appropriate reference sample for comparison.

[0061] Step A2: Data filtering. Data is cleaned using average filtering and threshold filtering to remove outliers and noise, thereby improving data quality.

[0062] Step A3: Acquisition of channel characteristics, which can be directly obtained from the data collected by the UWB module.

[0063] Step A4: CIR sequence acquisition, extracting the CIR sequence near FP_Index from the measured channel feature data.

[0064] Step A5: Number padding. The CIR data near FP_Index is too sparse to describe the shape similarity of CIRs. It is necessary to pad some data at equal intervals between every two adjacent CIR data, and use the slope between each pair of padding data to describe the shape similarity.

[0065] Step A6: Similarity feature vector calculation. Using the CIR sequence before padding, calculate the similarity vector representing the CIR sequence values ​​according to the similarity feature calculation method mentioned above. Then, use the CIR sequence after padding to calculate the similarity vector indirectly describing the shape. Combine the error of the reference sampling and the ranging error of the sample to finally form the CIR similarity vector.

[0066] Step A7: Dataset splitting. The processed dataset is divided into training and testing sets. When splitting the dataset, it is important to ensure that the data is evenly mixed so that it can be used in the model training and evaluation process. In this experiment, the dataset splitting ratio is 1:1.

[0067] Step B: Construction of NLOS recognition and error compensation model. This two-layer attention fusion network consists of 7 modules, including input layer, linear combination and batch normalization, attention mechanism, feature merging, deep feature extraction, dropout layer and output layer. It contains two input layers, which take in channel feature vector and similarity feature vector respectively. In order to avoid the problem that the same attention mechanism has poor evaluation effect on different types of input, the attention mechanism is evaluated on the two sets of inputs separately.

[0068] Step C: After completing the construction of the NLOS identification and error compensation model, deploy the model to the edge to realize the identification of UWBNLOS paths and the compensation of ranging errors.

[0069] In step A, regarding the acquisition of channel characteristics, UWB falls under the radio frequency (RF) field. The channel characteristics can be extracted from UWB sample data, containing various channel features. The extracted UWB channel characteristics include not only statistical information about observed distance values ​​such as distance, maximum noise level, and noise standard deviation, but also relevant features of the first path. The first path is the shortest propagation path from the transmitter to the receiver, typically representing the path with the shortest signal propagation time. In complex environments, the first path is usually an NLOS (Low-Noise-Side-Operate) path. In fact, available first path features, such as the first path peak value and first path strength, are crucial for evaluating UWB ranging errors. Furthermore, channel characteristics such as maximum noise level and noise standard deviation also show significant differences in LOS / NLOS environments.

[0070] Similarity features in feature engineering mainly include cosine similarity, Spearman correlation coefficient, Pearson correlation coefficient, Canberra distance, normalized Euclidean distance, Jensen-Shannon divergence, and Hellinger distance. The similarity vector between the CIR sequence near the FP_Index and the reference sample can be directly calculated. The shape difference between the prior reference CIR vector and the working sample CIR vector is also a useful reference for characterizing LOS and NLOS channels. For a dimensionless one-dimensional vector like the CIR signal, its shape characteristics can be reflected using a slope vector, calculated as described above. These similarity features, to some extent, reflect the severity of NLOS interference.

[0071] In step B, the core of this model building lies in the attention mechanism, which aims to improve the model's efficiency in identifying and processing key information in the input data. Figure 2 and Figure 3The flowcharts for NLOS and ranging error compensation are shown separately. For each batch of data after linear transformation and batch normalization, a fully connected layer is used to calculate attention weights. In the classification task, the SoftMax activation function is used to ensure that the sum of the feature weights is 1, allowing it to be interpreted as a probability distribution. For the LOS / NLOS ranging error compensation task, a linear rectified unit activation function is used to generate non-negative attention weights. In this way, the calculated attention weights are multiplied by the corresponding linear transformation output, achieving feature weighting and effectively adjusting the influence of each feature on the model output. This not only increases the model's ability to handle complex data but also improves its interpretability, as the attention weights directly reflect the contribution of each feature to the prediction result.

[0072] After feature evaluation, feature fusion is required. The goal of feature fusion is to integrate information from two independent input sources, achieving a comprehensive understanding of the input data through effective feature fusion. This not only enhances the model's representation of the data but also lays a solid foundation for in-depth feature extraction and final decision-making. In this model, feature fusion is implemented through the following steps:

[0073] (1) Attention weights are calculated for channel features and CIR similarity features respectively, and these weights are applied to the corresponding features to highlight features severely affected by NLOS. This operation is implemented using multiply, allowing the model to automatically focus on the features most important for NLOS identification. The multiply function refers to the element-wise multiplication of the input tensor, which is often used to implement certain activation functions, weight sharing, or feature combination. By multiplying element-wise, different features or signals are effectively combined, thereby improving the model's expressive power.

[0074] (2) The concatenate operation is used to combine weighted channel features and similarity features to create a comprehensive feature representation. The concatenate function connects multiple tensors along a specified axis. This operation can merge different features or data into a larger tensor, thereby increasing the dimensionality of the information. This function is often used to merge features from different layers or to combine information from different sources when processing multimodal data. This operation helps the model better learn and express complex relationships and features.

[0075] In summary, the UWB NLOS recognition and ranging error compensation algorithm proposed in this invention fully utilizes the difference between the peak values ​​of CIR sequences in the NLOS environment and those in the LOS environment. Based on this, a method for calculating the CIR similarity vector is designed, which reduces the number of input parameters compared to directly inputting the CIR sequence. This model can also switch between NLOS recognition and ranging error compensation tasks, greatly improving its utilization efficiency. Simultaneously, by first recognizing NLOS and then compensating for it, the model significantly improves the ranging accuracy of UWB in the NLOS environment, providing a foundation for high-precision UWB positioning and effectively suppressing the position estimation error of the target to be located.

[0076] The NLOS identification and error compensation model, after the above steps, can be deployed to edge devices to perform specific tasks. The process is as follows: Figure 4 As shown.

[0077] 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 NLOS recognition and ranging error compensation based on feature similarity, characterized in that: A dual-input attention neural network is used to evaluate the channel characteristics and CIR similarity features of the input, thereby realizing the identification of UWBNLOS paths and the compensation of ranging errors. When designing the CIR similarity feature vector, since there are obvious differences between CIR sequences in LOS and NLOS environments, similarity is calculated from the differences in CIR numerical values ​​and shapes. The designed CIR similarity vector comprehensively evaluates the degree of NLOS interference from both numerical and shape aspects, serving as an important basis for NLOS identification and error compensation. The CIR similarity feature vector includes similarity features such as a numerical similarity vector and a computable shape similarity vector between the sample CIR sequence and the reference sample CIR sequence; by directly calculating the partial similarity index between the CIR sequence near FP_Index and the reference sample CIR sequence, a similarity vector reflecting the numerical aspects of the sample and the reference sample CIR sequence is obtained. The computable shape similarity vector, which represents the shape difference features between the prior reference CIR vector and the working sampled CIR vector, uses shape features that are dimensionless one-dimensional vectors like CIR signals. The slope vector reflects the shape similarity of the CIRs by visualizing the original CIR sequence. When constructing the shape similarity vector, the CIR sequence is filled and the slope of the graphic is calculated to improve the shape restoration accuracy. The NLOS identification and error compensation neural network has two input layers that take in channel features and CIR similarity feature vectors, respectively. After evaluation by an attention mechanism, these features are fed into the neural network to achieve NLOS path identification.

2. The method for UWB NLOS recognition and ranging error compensation based on feature similarity according to claim 1, characterized in that: The position of FP_Index in CIR is directly related to the signal propagation time delay. The degree of NLOS interference is evaluated based on the peak arrival time and peak height of the CIR sequence.

3. The method for UWB NLOS recognition and ranging error compensation based on feature similarity according to claim 1, characterized in that: When calculating the CIR similarity feature vector, a reference sample close to the ideal LOS environment is selected. When calculating the CIR similarity features, the reference sample with the closest distance is selected for similarity vector comparison calculation.

4. The method for UWB NLOS recognition and ranging error compensation based on feature similarity according to claim 1, characterized in that: The CIR differences between LOS and NLOS environments are not always present. The differences are reflected in some CIR sequences near FP_Index. When calculating the similarity feature vector using CIR sequences, the original CIR sequences are truncated, and CIR sequences with larger differences are retained.

5. The method for UWB NLOS recognition and ranging error compensation based on feature similarity according to claim 1, characterized in that: The NLOS identification and error compensation steps include: Step A: Experimental data feature engineering, which involves extracting channel features and calculating CIR similarity feature vectors from the collected experimental data; The specific steps of experimental data feature engineering are as follows: Step A1: Reference sample matching. Find the reference sample that is closest to the sample by calculating the distance. Step A2: Data filtering, cleaning the data by using average filtering and threshold filtering to remove outliers and noise, thereby improving data quality; Step A3: Acquisition of channel characteristics, which can be directly obtained from the data collected by the UWB module; Step A4: CIR sequence acquisition. Extract the CIR sequence near FP_Index from the measured channel feature data. The differences in NLOSCIR sequences are mainly reflected through these CIR sequences. Step A5: Number padding. The CIR data near FP_Index is too sparse to describe the shape similarity of CIRs. It is necessary to pad some data at equal intervals between each pair of adjacent CIR data, and use the slope between the pairs of padding data to describe the shape similarity. Step A6: Similarity feature vector calculation. Calculate the similarity vector representing the CIR sequence values ​​using the CIR sequence before padding, and then calculate the similarity vector indirectly describing the shape using the CIR sequence after padding. Combine the error of the reference sampling and the ranging error of the sample to form the final input similarity vector. Step A7: Dataset partitioning. Divide the processed dataset into training and testing sets. When partitioning the dataset, ensure that the data is evenly mixed so that it can be used in the model training and evaluation process. In this training process, the dataset partitioning ratio is 1:

1. Step B: Based on the design of CIR similarity features, a UWB NLOS identification and ranging error compensation algorithm is constructed. This algorithm builds a two-layer attention neural network. The two input layers are respectively input to the channel feature vector and the CIR similarity feature vector. The two sets of inputs are evaluated by the attention mechanism, and feature extraction is performed through the deep neural network to complete the identification of NLOS. The channel feature vector and CIR similarity feature vector of the identified NLOS type data are used as input parameters to enter the ranging error compensation neural network to complete the compensation task.

6. The method for UWB NLOS recognition and ranging error compensation based on feature similarity according to claim 5, characterized in that: The NLOS recognition algorithm and the error compensation algorithm use the same neural network structure, and can switch flexibly between classification and compensation tasks by adjusting the parameters.

Images