CN-121980213-A - Low-tag-dependence semi-supervised UWB positioning non-line-of-sight identification method

CN121980213ACN 121980213 ACN121980213 ACN 121980213ACN-121980213-A

Abstract

The invention discloses a semi-supervised UWB positioning non-line-of-sight identification method with low label dependence, which comprises the steps of firstly collecting CIR data of a UWB positioning system, constructing a data set containing CIR samples of line-of-sight LOS and non-line-of-sight NLOS, then carrying out non-supervision training by adopting a fully-connected automatic encoder based on a non-labeled subset containing only part of line-of-sight CIR samples, freezing encoder parameters after training is completed, inputting all remaining line-of-sight and non-line-of-sight samples into the frozen encoder, obtaining corresponding reconstructed samples, calculating error characteristics and similarity characteristics between original samples and reconstructed samples, constructing a comprehensive characteristic vector set for distinguishing line-of-sight/non-line-of-sight signals, finally training a multi-layer perceptron classifier based on the comprehensive characteristic vector set, carrying out semi-supervised learning by utilizing comprehensive characteristic training data, optimizing a classification model by minimizing a cross entropy LOSs function, and finally realizing accurate identification of line-of-sight and non-sight signal types.

Inventors

HU QINGSONG
GAO WENJIE
CHENG YUANXUN
WU DONG

Assignees

中国矿业大学

Dates

Publication Date: 20260505
Application Date: 20251222

Claims (6)

1. A semi-supervised UWB positioning non-line-of-sight identification method with low tag dependence is characterized by comprising the following steps: Step1, performing normalization preprocessing on the sight distance and non-sight distance channel impulse response data acquired in public UWB positioning, dividing a total data set into a non-tag sight distance sample data set L and a tagged mixed sight distance and non-sight distance sample data set U, and dividing a training set, a verification set and a test set respectively for the non-tag sight distance sample data set L and the tagged mixed sight distance and non-sight distance sample data set U; step2, performing unsupervised data reconstruction training on a training set of the unlabeled line-of-sight sample data set by using a full-connection automatic encoder, minimizing the mean square error loss, storing a full-connection automatic encoder model after training is completed, and freezing weight parameters; Step3, inputting all original samples in the labeled mixed line-of-sight and non-line-of-sight sample data set U into a fully-connected automatic encoder with freezing weight parameters to obtain corresponding reconstructed samples, and obtaining a sample pair for each original sample; Step4, for a sample pair formed by the labeled mixed line-of-sight and non-line-of-sight sample dataset U after full connection of an automatic encoder, calculating error features and similarity features of the sample pair to construct a comprehensive feature vector training set F train , a verification set F val and a test set F test for distinguishing line-of-sight and non-line-of-sight signals; step5, constructing and training a multi-layer perceptron classifier, performing classifier supervision training by using the comprehensive feature vector training set F train , and storing the multi-layer perceptron classifier after training is finished; Step6, for the channel impulse response sample to be tested, obtaining a sample pair of the original sample and the reconstructed sample, calculating a comprehensive feature vector, inputting the comprehensive feature vector into the trained multi-layer perceptron classifier to obtain a classification label, and realizing non-line-of-sight identification.
2. The low tag-dependent semi-supervised UWB positioning non-line-of-sight identification method of claim 1, wherein Step1, for the line-of-sight and non-line-of-sight channel impulse response CIR data of a positioning scene collected by the public dataset, the CIR is represented by a complex I/Q sample, and the I/Q sample array is converted into an RSSI array, expressed as Performing minimum-maximum normalization preprocessing on the RSSI; The unlabeled line-of-sight sample dataset L is represented as: Where l i is the data set ith sample, N l is the total sample data size; the labeled mixed line-of-sight and non-line-of-sight sample dataset U is represented as: Where u j is the j-th sample of the dataset, y j is its corresponding label, and N u is the total sample data size; Training, validation and test sets of the unlabeled line-of-sight sample dataset L are denoted L train 、L val and L test , respectively, and l=l train ∪L val ∪L test ; The training set, validation set, and test set for labeled mixed line of sight and non-line of sight sample dataset U are denoted U train 、U val and U test , respectively, and u=u train ∪U val ∪U test .
3. The low tag-dependent semi-supervised UWB positioning non-line-of-sight identification method of claim 2, wherein in Step2, the training set L train of the unlabeled line-of-sight sample dataset L is represented as: Where l m (train) is the mth sample, M is the total sample data size; While minimizing the mean square error loss, the training goal is to minimize the original input samples/ m (train) and reconstruct the output samples Reconstruction error, mean square error loss function between The expression is as follows: Wherein, l m (train) is the M-th sample, M is the total sample data volume; And reconstructing the output samples.
4. The method for semi-supervised UWB positioning non-line-of-sight identification with low tag dependency as defined in claim 3, wherein in Step3, all original samples U j in training set U train , verification set U val and test set U test of tagged mixed line-of-sight and non-line-of-sight sample dataset U are input into a fully connected automatic encoder with freezing weight parameters to obtain corresponding reconstructed samples Obtaining a sample pair for each original sample u j
5. The low tag-dependent semi-supervised UWB positioning non-line-of-sight identification method of claim 4, wherein Step4 is performed for all sample pairs obtained in Step3 Each of the CIR original and reconstructed samples has a length of n vector, denoted u j ＝[u j1 ,u j2 ,...,u jn , Computing the original samples u j and corresponding reconstructed samples The error features and the similarity features comprise a mean square error MSE, an average absolute error MAE, a cosine similarity CS and a pearson correlation coefficient PCC, and the calculation formula is specifically as follows: Where n is the total length of a single CIR sample vector, i is the index of the element in the CIR vector (i=1, 2,., n), u ji is the value of the i-th element of the j-th original CIR sample u j ; is the jth reconstructed CIR sample The value of the i-th element of (a); As the mean value of the original sample, In order to reconstruct the mean value of the samples,
6. The low-tag-dependency semi-supervised UWB positioning non-line-of-sight identification method of claim 5, wherein in Step5, when the classifier supervised training is performed using the integrated feature vector training set F train , the training objective is to minimize a class-two cross entropy loss function between the predicted tag and the real tag, the class-two cross entropy loss function The expression is as follows: Wherein, K is the total number of samples, y j is the real label of the jth sample, the real label is 0 for LOS data, and the real label is 1 for NLOS data; is the predicted probability of the jth sample belonging to NLOS data.

Description

Low-tag-dependence semi-supervised UWB positioning non-line-of-sight identification method Technical Field The invention relates to a method for identifying wireless signals transmitted by Non-Line Of Sight (NLOS) in Ultra-Wideband (UWB) positioning, in particular to a semi-supervised UWB positioning Non-Line Of Sight identification method with low tag dependence, belonging to the technical field Of Ultra-Wideband. Background High-precision positioning is one of core support technologies in the fields of industrial automation, intelligent infrastructure and the like. Ultra wideband technology is widely used due to its high time resolution and multipath interference resistance in indoor or dense multipath environments. However, the range error caused by non-line-of-sight propagation can significantly affect UWB positioning accuracy, and therefore, effective discrimination of NLOS states is critical to achieving high-accuracy UWB positioning. The existing non-line-of-sight recognition method mainly depends on a supervised deep learning model, such as 'a non-line-of-sight signal recognition method based on a wavelet glamer convolutional neural network' of China patent application No. 202211105627.4 and 'a non-line-of-sight signal recognition method based on deep learning' of China patent application No. 202110809540.4, which are required to collect a large amount of labeled data for training, and the implementation cost for realizing non-line-of-sight recognition in a UWB positioning system is increased. How to reduce the dependence on tagged data to reduce the implementation cost of implementing non-line-of-sight identification in UWB positioning systems is still a problem to be solved in the industry. Disclosure of Invention Aiming at the problems in the prior art, the invention provides a semi-supervised UWB positioning non-line-of-sight identification method with low tag dependence, which can effectively reduce the dependence on a large amount of tagged data on the premise of accurately identifying NLOS signals. In order to achieve the above purpose, the semi-supervised UWB positioning non-line-of-sight identification method with low tag dependence specifically comprises the following steps: Step1, performing normalization preprocessing on the sight distance and non-sight distance channel impulse response data acquired in public UWB positioning, dividing a total data set into a non-tag sight distance sample data set L and a tagged mixed sight distance and non-sight distance sample data set U, and dividing a training set, a verification set and a test set respectively for the non-tag sight distance sample data set L and the tagged mixed sight distance and non-sight distance sample data set U; step2, performing unsupervised data reconstruction training on a training set of the unlabeled line-of-sight sample data set by using a full-connection automatic encoder, minimizing the mean square error loss, storing a full-connection automatic encoder model after training is completed, and freezing weight parameters; Step3, inputting all original samples in the labeled mixed line-of-sight and non-line-of-sight sample data set U into a fully-connected automatic encoder with freezing weight parameters to obtain corresponding reconstructed samples, and obtaining a sample pair for each original sample; Step4, for a sample pair formed by the labeled mixed line-of-sight and non-line-of-sight sample dataset U after full connection of an automatic encoder, calculating error features and similarity features of the sample pair to construct a comprehensive feature vector training set F train, a verification set F val and a test set F test for distinguishing line-of-sight and non-line-of-sight signals; step5, constructing and training a multi-layer perceptron classifier, performing classifier supervision training by using the comprehensive feature vector training set F train, and storing the multi-layer perceptron classifier after training is finished; Step6, for the channel impulse response sample to be tested, obtaining a sample pair of the original sample and the reconstructed sample, calculating a comprehensive feature vector, inputting the comprehensive feature vector into the trained multi-layer perceptron classifier to obtain a classification label, and realizing non-line-of-sight identification. Further, in Step1, for the line-of-sight and non-line-of-sight Channel Impulse Response (CIR) data of a positioning scene acquired by the public dataset, the CIR is represented by a complex I/Q sample, and the I/Q sample array is converted into an RSSI array, which is represented asPerforming minimum-maximum normalization preprocessing on the RSSI; The unlabeled line-of-sight sample dataset L is represented as: Where l i is the data set ith sample, N l is the total sample data size; the labeled mixed line-of-sight and non-line-of-sight sample dataset U is represented as: Where u j is the j-th sample of the dataset, y j is its c