Indoor environment composition method based on virtual anchor points under narrowband signal condition

Abstract

The invention provides a VA-based indoor environment composition method under a narrowband signal condition. In a single-base-station and single-antenna scene, high-precision environment map construction is realized through the RRoF of multipath signals. In order to solve the problems that the time delay resolution of a narrow-band signal is limited and absolute ToF measurement is easily influenced by clock phase deviation, an RToF is adopted as a supervision signal to suppress system errors. By designing DNN, end-to-end learning is carried out on nonlinear mapping from a user position to VA, so that manual modeling of a complex reflection model is avoided. In addition, an iterative optimization algorithm based on ToF geometric consistency check is introduced, the network is retrained by screening high-confidence samples, VA positions are self-supervised and corrected, and the result is ensured to be stable and convergent. The method gets rid of dependence on a multi-antenna array and special hardware, reduces the cost and deployment difficulty, and provides reliable composition technical support for indoor scenes such as underground parking lots and shopping malls.

Description

Technical Field

[0001] This invention relates to indoor environment mapping technology, specifically a method for indoor environment mapping based on virtual anchor points under narrowband signal conditions. Background Technology

[0002] Indoor environmental mapping is a crucial foundation for achieving precise location services. Traditional mapping methods typically rely on technologies such as LiDAR and Ultra-Wideband (UWB), which suffer from limitations such as high cost and complex deployment, restricting their large-scale application. Virtual Anchor (VA) technology offers a new approach to low-cost environmental mapping. This technology can transform the multipath reflection paths of wireless signals into environmental geometric structure information. Its core is to establish a mirror model of the base station (BS) relative to reflective surfaces such as walls and columns, enabling the VA to represent the spatial relationships of the environmental structure. Therefore, indoor mapping can be achieved without the need for additional dedicated hardware, significantly reducing costs and deployment complexity.

[0003] However, the application of VA technology in practical scenarios of single-antenna, narrowband communication still faces multiple challenges. Most existing methods rely on joint measurements of Time of Flight (ToF) and Angle of Arrival (AoA). Obtaining AoA requires a multi-antenna array, significantly increasing system complexity and hardware cost, making it unsuitable for single-antenna devices such as smartphones and IoT terminals. Furthermore, measurements based on absolute ToF are susceptible to clock phase deviations, resulting in systematic errors and decreased VA estimation accuracy. To overcome these limitations, Relative Time of Flight (RToF) is used as a monitoring signal. By calculating the time delay difference between the reflected path and the direct path, systematic errors can be effectively suppressed, creating conditions for single-base station mapping. On the other hand, indoor multipath propagation exhibits significant nonlinear characteristics, making it difficult for traditional geometric models to accurately characterize its reflection mapping relationship. Deep Neural Networks (DNNs), trained on large-scale data, can autonomously learn complex mappings from user location to VA location. They utilize Multilayer Perceptrons (MLPs) to capture complex spatial geometric constraints, significantly improving the model's environmental adaptability. However, their initial predictions are susceptible to multipath interference and measurement noise, necessitating the introduction of subsequent optimization mechanisms to enhance estimation robustness.

[0004] This invention proposes an indoor environment mapping method based on VA (Value at VA) under narrowband signal conditions, achieving environment reconstruction using only RToF (Real-Time ToF) under single-base station and single-antenna conditions. A DNN (Distributed Neural Network) is designed to learn the complex nonlinear mapping relationship from the user's location to the VA location, avoiding the manual design of complex reflection models and exhibiting good noise robustness. To further improve mapping accuracy, an iterative optimization algorithm based on geometric consistency testing is introduced. In each iteration, the consistency score between the theoretical ToF and the estimated ToF is calculated, and high-confidence samples are selected for retraining the DNN, forming a self-supervised optimization loop. Through multiple iterative optimizations, the system gradually converges to a stable VA estimation result. Summary of the Invention

[0005] The purpose of this invention is to provide an indoor environment mapping method based on VA under narrowband signal conditions. It can effectively utilize multipath RToF and deep learning technology to achieve accurate estimation of VA location, thereby generating an accurate indoor environment map.

[0006] The VA indoor environment mapping method based on narrowband signals described in this invention includes the following steps:

[0007] Step 1: Collect user location data and corresponding multipath signal parameters, calculate the RToF of each reflection path relative to the direct path, and convert it into relative distance to form a reference distance set;

[0008] Step 2: Design a DNN to learn a non-linear mapping from the user's location to the VA location;

[0009] Step 3: For each user and its predicted VA, calculate the predicted relative distance based on the BS location, user location, and predicted VA location, and form a predicted distance set.

[0010] Step 4: Use the Hungarian algorithm to establish an optimal one-to-one correspondence between the reference distance set and the predicted distance set, construct the cost matrix, and solve for the minimum total matching error;

[0011] Step 5: Calculate the difference in distance for each pair based on the matching results, and use the smooth L1 loss function as the training objective;

[0012] Step 6: Perform K-Means clustering on the VA positions obtained from multiple predictions. The cluster centers are the final VA positions corresponding to the main reflecting surfaces.

[0013] Step 7: Further optimize the VA position using a ToF-based iterative selection algorithm, specifically including:

[0014] Step 7 (1) Use the VA position predicted by the DNN as the initial value and set the iteration counter k = 1;

[0015] Step 7 (2): Calculate the theoretical Time of Flight (ToF) for each VA based on the current VA location estimate and the user location;

[0016] Step 7 (3): Obtain the estimated ToF set and construct the difference matrix to evaluate the consistency between the theoretical value and the estimated value;

[0017] Step 7 (4): Calculate the minimum difference score for each estimated ToF, and select the top Q samples with the smallest scores as the high confidence set;

[0018] Step 7 (5): Retrain the DNN using high-confidence samples as supervision signals to update the VA location estimate;

[0019] Step 7 (6): Calculate the average change of VA position between consecutive iterations. If the change is lower than the threshold β and continues for S iterations, terminate the optimization process. Otherwise, let k = k + 1 and return to step 7 (2).

[0020] Beneficial effects

[0021] This invention is a method for indoor environment mapping based on VA under narrowband signal conditions, which has the following advantages:

[0022] 1. High-precision estimation of VA and reliable environmental mapping can be achieved with a single base station and a single antenna;

[0023] 2. Deep learning is used to directly learn the complex nonlinear mapping from the user's location to the VA location, eliminating the need for manual feature design;

[0024] 3. Using RToF as the monitoring signal effectively avoids system deviations caused by phase errors;

[0025] 4. The Hungarian algorithm is used to solve the correspondence problem between the reference distance and the predicted distance, thus improving training stability;

[0026] 5. An iterative optimization algorithm is used to gradually improve the accuracy of VA estimation, and geometric consistency is used to select high-confidence samples to guide network training. Attached Figure Description

[0027] Figure 1 This is a flowchart of the present invention;

[0028] Figure 2 This is a diagram of the DNN architecture used to predict the location of the virtual avatar (VA). Detailed Implementation Plan

[0029] The present invention will now be described in further detail with reference to the accompanying drawings. Figure 1 The flowchart of this invention is as follows:

[0030] Step 1: Collect user location data and corresponding channel state information (CSI), and use a parameter estimation algorithm to obtain the Time-of-Flight (ToF) of the multipath signal. Assume τ1 represents the direct path, τ... l (l=2,…,L) represents the (L-1)th reflection path. RToF is defined as τ l -τ1, can be obtained through d l-1 =(τ l -τ1)·C is converted into a relative distance, where C is the speed of light. For a magnitude of N, ... b The training batches yield a reference distance set denoted as D = {d1, d2, ..., d...} J}, where J = (L-1) × N b .

[0031] Step 2: The network is implemented using MLP, and the proposed network architecture is as follows: Figure 2 As shown. Let the three-dimensional coordinates of the j-th user be... The input layer receives the user's 3D coordinates, and the output layer outputs the 3D coordinates of M VAs. The hidden layers use the Rectified Linear Unit (ReLU) as the activation function to facilitate effective training and mitigate the vanishing gradient problem. Gradient descent and backpropagation algorithms are employed to optimize the model parameters and minimize prediction error.

[0032] Step 3: For each user and their predicted VA, calculate the predicted relative distance:

[0033]

[0034] Where P0 represents the location of BS. Each user generates M predicted distances, forming a predicted distance set I = {I1, I2, ..., I...} K}, where K = M × N b .

[0035] Step 4: Construct a cost matrix to quantify the error between each element in the reference distance set and the predicted distance set, and use the Hungarian algorithm to solve for the optimal matching relationship. The specific steps are as follows:

[0036] Step 4 (1) Construct the cost matrix Each element C jk =|d j -I k | represents the absolute error between the j-th reference distance and the k-th predicted distance;

[0037] Step 4 (2): Use the Hungarian algorithm to solve for the optimal allocation matrix A = {a jkThe optimization objective is to minimize the total matching cost.

[0038]

[0039] Among them, a jk ∈{0,1} represents d j Is it related to I? k Matching. Once the optimal match is determined, calculate the residual ΔΓ for each match. jk =I k -d j .

[0040] Step 5: Calculate the sum of the losses for all matching pairs:

[0041]

[0042] Wherein, ρ(·) is the smoothing L1 loss function, which combines the advantages of L2 loss and L1 loss, thus possessing both convergence stability and outlier robustness.

[0043] Step 6: Perform K-Means clustering on the VA positions obtained from multiple predictions. The cluster centers are the final VA positions corresponding to the main reflective surfaces.

[0044] Step 7: Iterative optimization to improve VA estimation accuracy. The specific steps are as follows:

[0045] Step 7 (1) Use the DNN trained in Step 2 to obtain an initial estimate of the VA position. Initialize the iteration counter k = 1;

[0046] Step 7 (2): Estimate VA based on the k-th iteration. and user location p u Calculate the theoretical flight time:

[0047]

[0048] Step 7 (3): Construct the difference matrix To assess the consistency between theoretical and estimated values, each element in the matrix is ​​defined as follows:

[0049]

[0050] in, A set of estimated flight times obtained from parameter estimation;

[0051] Step 7 (4): For each estimated flight time Calculate its minimum deviation from all theoretical flight times. Select the top Q samples with the lowest scores As a high-confidence sample set;

[0052] Step 7 (5): Use the high-confidence sample set as new training data to retrain the DNN and update it to obtain a new VA estimate;

[0053] Step 7 (6): Calculate the average position change of VA between successive iterations. If Δ (k) If <β satisfies the condition for S consecutive iterations, then terminate the optimization; otherwise, let k = k + 1 and return to step seven (2).

Claims

1. A method for indoor environment mapping based on virtual anchors (VA) under narrowband signal conditions, characterized in that: a) A deep neural network (DNN) is used to learn the non-linear mapping relationship from the user's location to the VA location; b) Use relative time of flight (RToF) as a monitoring signal to avoid the phase error of time of flight (ToF); c) Solve the optimal matching problem between the reference distance and the predicted distance using the Hungarian algorithm; d) The accuracy of VA estimation is gradually improved by adopting an iterative optimization algorithm based on ToF.

2. The indoor environment mapping method based on virtual anchor points under narrowband signal conditions according to claim 1, which utilizes a DNN to learn the nonlinear mapping relationship from the user's location to the VA location, is characterized in that... Establish a complex mapping relationship between user location and VA location. In indoor environments, due to the relatively fixed spatial distribution of major reflective surfaces such as walls and columns, a stable geometric constraint relationship exists between user location and VA location. DNN, through multi-layer nonlinear transformations, can effectively learn this complex spatial mapping pattern. Assume the 3D coordinates of the j-th user are... Train a DNN with parameter θ to learn a nonlinear function f. θ This function maps the user's location to the predicted VA location: in, Let θ represent the coordinates of the M VAs predicted by the j-th user. The network adopts a Multilayer Perceptron (MLP) structure, with the hidden layers using Rectified Linear Units (ReLU) as activation functions, and the parameter θ is optimized using gradient descent.

3. The indoor environment mapping method based on virtual anchor points under narrowband signal conditions according to claim 1, which uses RToF as a monitoring signal, is characterized in that... To avoid system bias caused by phase errors. Traditional Time-of-Flight (ToF) measurements are susceptible to carrier phase ambiguity and multipath effects, leading to systematic errors. Real-Time-of-Flight (RToF), by calculating the time delay difference between the reflected path and the direct path, effectively eliminates errors caused by clock asynchrony between the transmitter and receiver. Specifically, it includes the following steps: a) Assume τ1 represents the direct path, τ l (l=2,…,L) represents the (L-1)th reflection path. RToF is defined as τ l -τ1, can be obtained through d l-1 =(τ l -τ1)·C converts this to a relative distance, where C is the speed of light. For a magnitude of N b The training batches yield a reference distance set denoted as D = {d1, d2, ..., d...} J }, where J = (L-1) × N b This set provides geometric supervision to ensure consistency between the predicted VA location and its mirror location relative to the base station (BS). b) For each user and their predicted VA, calculate the relative distance between the predictions: Where P0 represents the location of BS. Each user generates M predicted distances, forming a predicted distance set I = {I1, I2, ..., I...} K }, where K = M × N b .

4. The indoor environment mapping method based on virtual anchor points under narrowband signal conditions according to claim 1, which solves the optimal matching problem between reference distance and predicted distance using the Hungarian algorithm, is characterized in that... Establish optimal matching between distance sets with no prior correspondence. Since there is no pre-defined one-to-one correspondence between the reference distance set and the predicted distance set, direct matching by index will lead to significant matching errors. The Hungarian algorithm, through an optimal allocation strategy, can find the correspondence that minimizes the total matching error. Specifically, it includes the following steps: a) Construct a cost matrix Each element C jk =|d j -I k The error between the j-th reference distance and the k-th predicted distance is quantified. The optimal allocation matrix A = {a} is solved using the Hungarian algorithm. jk The optimization objective is to minimize the total matching cost. Among them, a jk ∈{0,1} represents d j Is it related to I? k Matching. Once the optimal assignment is determined, the residual for each matching pair is calculated as ΔΓ. jk =I k -d j Calculate the sum of the losses for all matching pairs: Here, ρ(·) is the smoothed L1 loss function (also known as Huber loss). This loss function combines the advantages of L2 and L1 losses: when the error is small, L2 loss is used, which has the characteristics of stable and continuously differentiable gradients; when the error is large, L1 loss is used, which is robust to outliers and maintains a constant gradient. In this way, this loss function not only improves the model's robustness to outliers but also ensures the stability of the model's convergence process. b) To reduce prediction noise and improve reliability, K-Means clustering is performed on the set of VA positions obtained by multiple predictions from DNN. The resulting cluster centers are the final VA positions corresponding to the main reflective surfaces.

5. The indoor environment mapping method based on virtual anchor points under narrowband signal conditions according to claim 1, characterized in that an iterative optimization algorithm is used to gradually improve the VA estimation accuracy. Self-supervised optimization is achieved through geometric consistency checks. Initial predictions in DNNs are affected by measurement noise and multipath interference, resulting in limited accuracy in VA estimation. Iterative optimization algorithms use geometric consistency checks between theoretical reflection paths and estimated data to progressively select high-confidence samples for network retraining. Specifically, the following steps are included: a) Let This represents the initial VA position predicted by the DNN, and the iteration counter k = 1 is set. b) Calculate the theoretical flight time in the k-th iteration: c) Construct the difference matrix Each element in the matrix is ​​defined as: in, A set of estimated flight times obtained from parameter estimation; d) Calculate the consistency score The top Q samples with the lowest scores are selected to form a high-confidence set. This is then used as a supervisory signal to retrain the DNN, thereby updating the VA estimate; e) Calculate the change in position When Δ (k) The optimization terminates when <β is satisfied for S consecutive iterations, and the final estimated VA position is output. Otherwise, let k = k + 1 and return to step b).

Images