CN-122028175-A - Wireless positioning method and system based on probability perception visual distance-non-visual distance dual-branch network

CN122028175ACN 122028175 ACN122028175 ACN 122028175ACN-122028175-A

Abstract

The invention belongs to the technical field of wireless communication and positioning, and particularly relates to a wireless positioning method and system based on a probability perception line-of-sight-non-line-of-sight dual-branch network. The method comprises the steps of obtaining multi-antenna CSI information of a receiving end, extracting geometric consistency characteristics of the CSI through a shared characteristic extractor, respectively inputting the extracted characteristics into an LOS expert branch and an NLOS expert branch which are arranged in parallel, respectively outputting position prediction values of the LOS and the NLOS, simultaneously calculating posterior probability weights of a current sample belonging to a sight distance or a non-sight distance state according to the CSI characteristics by using a probability routing network, and finally carrying out weighted fusion on the position prediction values of the two branches based on the posterior probability weights to obtain final user position estimation. The invention adopts a strategy of 'divide and conquer' and combines a task-oriented two-stage training mechanism, thereby effectively solving the problems of low positioning precision and poor hard decision robustness of a single model in a mixed propagation environment and remarkably improving the accuracy and stability of wireless positioning.

Inventors

ZHU YAPING
CHEN YILONG
WANG JUNYUAN
DENG HAO
HAN FENGXIA

Assignees

同济大学

Dates

Publication Date: 20260512
Application Date: 20260323

Claims (5)

1. A wireless positioning method based on a probability-aware line-of-sight-non-line-of-sight dual-branch network is characterized by comprising the following steps: step 1, data preprocessing, namely receiving a wireless signal sent by user equipment to be positioned by a receiving end, acquiring a Channel State Information (CSI) matrix, extracting amplitude information and phase information of the CSI matrix, stacking in a channel dimension to form a dimension of Is a real-valued tensor of (c), wherein, In order to receive the number of antennas, Is the number of subcarriers; step 2, shared feature extraction, namely constructing a shared feature extraction network, and carrying out feature extraction on input tensors to obtain potential feature vectors; Step 3, double-branch expert prediction, namely constructing a double-branch prediction network, namely, two parallel full-connection network branches which are respectively used as an LOS expert branch and an NLOS expert branch, inputting the potential feature vectors obtained in the sharing step 2, respectively learning the mapping relation between CSI and the position under the condition of visual range and non-visual range, and obtaining the position results of LOS expert and NLOS expert prediction; Step 4, probability route inference, namely constructing a probability route network, inputting the probability route network into the CSI tensor obtained in the step 1, and obtaining the vision distance propagation probability weight after processing And non-line-of-sight propagation probability weights ; Step 5, uncertainty perception fusion, namely weighting and summing the coordinates predicted by the LOS expert and the NLOS expert according to probability weights output by a probability routing network to obtain a final position, wherein a calculation formula is as follows Where z ε { LOS, NLOS }, As the CSI data, there is provided, And The position estimation values are output through LOS expert branches and NLOS expert branches respectively; Step 6, training in two stages: Firstly, performing supervision training by utilizing training data with a propagation state label, and simultaneously training a shared feature extraction network and a double-branch prediction network, so that regression loss under a state condition is minimized, and each branch is focused on feature mapping under a specific propagation scene; and step two, freezing parameters of the shared feature extraction network and the double-branch prediction network, training only the probability routing network, optimizing the parameters of the probability routing network by taking the position error minimization after final weighted fusion as a target, wherein a loss function is defined as follows: where z ε { LOS, NLOS } represents the propagation state, N is the number of samples, As the coordinates of the true position of the object, As a probability weight of the probability that the probability is high, Mapping functions that are LOS/NLOS expert branches, A mapping function of the network is extracted for the shared feature.
2. The method according to claim 1, wherein in step 2, the shared feature extraction network adopts a convolutional neural network CNN structure, and includes several two-dimensional convolutional layers, a batch normalization layer and an activation function layer, and compresses the feature map into a fixed-length potential feature vector through a global average pooling layer.
3. The method according to claim 1, wherein in step 4, the probabilistic routing network is a multi-layer fully connected network, the output layer uses a Softmax function with a temperature coefficient τ, and the calculation formula is as follows: Wherein, the As the CSI data, there is provided, For the output of the probabilistic routing network, Is a temperature coefficient.
4. The method according to claim 1, wherein in step 6, the phase one loss function is designed as: Wherein, the In order to indicate the function, , As the coordinates of the true position of the object, A mapping function for predicting network branches for a dual branch, The LOS data only updates the LOS expert branch and the shared feature extraction network, and the NLOS data only updates the NLOS expert branch and the shared feature extraction network.
5. A wireless positioning system based on a probabilistic-aware line-of-sight-non-line-of-sight dual-branch network, comprising a processor and a memory, the memory having stored therein a computer program, characterized in that when the processor executes the computer program, a method of wireless positioning based on a probabilistic-aware line-of-sight-non-line-of-sight dual-branch network as claimed in any one of claims 1 to 4 is performed.

Description

Wireless positioning method and system based on probability perception visual distance-non-visual distance dual-branch network Technical Field The invention relates to the technical field of wireless communication and positioning, in particular to a wireless positioning method and system based on a probability perception line-of-sight-non-line-of-sight dual-branch network. Background With the development of 5G/6G networks and internet of things technologies, location Based Services (LBS) are increasingly demanded. Conventional Global Navigation Satellite Systems (GNSS) tend to fail in urban canyons in indoor or tall buildings. Wireless positioning technology based on cellular network, especially positioning by using physical layer Channel State Information (CSI), has great potential in realizing high-precision positioning because it contains multipath features (such as amplitude, phase, delay, etc.) richer than Received Signal Strength (RSSI). However, the wireless signal propagation environment is extremely complex. In practical applications, the received signal may arrive through a direct path (LOS) or may arrive through only a reflected, scattered path (NLOS) due to occlusion. There are substantial differences in the CSI statistics in these two propagation states and their geometrical laws mapped to physical locations. The existing positioning method based on deep learning generally has the following problems: 1) Many methods use a single neural network to fit all scenes, resulting in models that attempt to find "average" solutions among distinct distributions, producing "mean regression" effects, reducing positioning accuracy. 2) Hard classification is less robust, a partial approach attempts to classify the propagation state (LOS or NLOS) first, and then selects the corresponding model. In practice, however, the propagation state tends to be ambiguous (e.g., weak LOS paths), and erroneous classification decisions can result in the use of a completely erroneous positioning model, resulting in significant positioning errors. 3) The uncertainty modeling is lacking, namely the uncertainty of propagation state inference per se is often ignored by the existing method, and the trust degree of different models cannot be dynamically adjusted under a fuzzy scene. Therefore, a wireless positioning method capable of explicitly handling propagation heterogeneity and having high robustness is required. Disclosure of Invention The invention aims to provide a wireless positioning method and a system based on a probability perception line-of-sight-non-line-of-sight dual-branch network, which aim to solve the problems of low positioning precision and poor robustness in a mixed propagation environment in the prior art and realize high-precision wireless positioning through a decoupling feature learning and probability fusion mechanism. In order to achieve the above purpose, the invention adopts the following technical scheme: a wireless positioning method based on a probability-aware line-of-sight-non-line-of-sight dual-branch network comprises the following steps: step 1, data preprocessing, namely receiving a wireless signal sent by user equipment to be positioned by a receiving end, acquiring a Channel State Information (CSI) matrix, extracting amplitude information and phase information of the CSI matrix, stacking in a channel dimension to form a dimension of Is a real-valued tensor of (c), wherein,In order to receive the number of antennas,Is the number of subcarriers. And 2, shared feature extraction, namely constructing a shared feature extraction network, and carrying out feature extraction on the input tensor to obtain potential feature vectors. The shared feature extraction network adopts a Convolutional Neural Network (CNN) structure, comprises a plurality of two-dimensional convolutional layers, a batch normalization layer and an activation function layer, and compresses a feature map into potential feature vectors with fixed lengths through a global average pooling layer. And 3, constructing a double-branch prediction network, wherein the double-branch prediction network comprises two parallel fully-connected network branches which are respectively used as an LOS expert branch and an NLOS expert branch. And (3) inputting the potential feature vectors obtained in the sharing step (2), and respectively learning the mapping relation between the CSI and the position under the condition of the sight distance and the non-sight distance to obtain the position results predicted by the LOS expert and the NLOS expert. Step4, probability route inference, namely constructing a probability route network, inputting the probability route network as the CSI tensor obtained in the step 1, and obtaining the apparent distance propagation probability weight and the non-apparent distance propagation probability weight after processingAnd)。 The probability routing network is a multi-layer fully connected network, and the output l