Ultra-wideband indoor positioning method based on multi-mode diffusion model

Abstract

The invention discloses an ultra-wideband indoor positioning method based on a multi-mode diffusion model, which belongs to the field of indoor positioning and comprises the following steps: acquiring distance data and RSSI (Received Signal Strength Indicator) data between UWB anchor points and labels in an indoor environment based on a mobile multi-sensor platform; performing feature extraction and feature fusion on the multi-modal input data through a multi-modal feature extraction module; an improved denoising diffusion implicit model is adopted to perform accelerated reverse sampling, a U-Net architecture neural network is utilized to realize high-quality distance reconstruction, and a distance estimation value is output; and establishing a position estimation layer, performing time sequence modeling by adopting a long-short-term memory network, learning position estimation, and outputting a three-dimensional coordinate positioning value. According to the method, UWB distance data and RSSI data can be effectively fused, high-quality reconstruction of lost data is achieved, meanwhile, statistical characteristics and spatial-temporal correlation of the data are reserved, and the precision and robustness of indoor positioning are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of indoor positioning technology, and in particular to an ultra-wideband indoor positioning method based on a multi-modal diffusion model. Background Art

[0002] Ultra-WideBand (UWB) technology has broad application prospects in the field of indoor positioning due to its high resolution, high precision and strong anti-interference ability. The indoor measurement data of UWB sensors mainly include four types: distance measurement between tags and anchor points, time of arrival of tag signals at anchor points (TOA), received signal strength indicator (RSSI), and channel impulse responses (CIR) data between tags and anchor points. Among them, distance measurement is the basis of UWB positioning, and the distance between anchor points and tags is usually calculated based on TOA. RSSI reflects the signal strength of the receiver and is usually used to estimate the distance between devices.

[0003] RSSI measurement consists of two different components: the residual received signal strength indicator (RxRssi) and the first path received signal strength indicator (FpRssi). These two RSSI measurements provide complementary information about the signal propagation environment. RxRssi reflects the overall signal strength, while FpRssi indicates the strength of the first arrival signal path. The relationship between these two measurements is particularly valuable for determining line-of-sight (LOS) or non-line-of-sight (NLOS) conditions: in an ideal LOS environment, signal attenuation follows a relatively stable propagation law, and the RSSI value maintains a clear relationship with the distance between the tag and the anchor point. Path loss models or logarithmic distance models are usually used for accurate distance estimation.

[0004] However, in actual indoor environments, UWB signal propagation faces multiple challenges. 1. In an NLOS environment, when the signal encounters obstacles such as walls and furniture, reflection, refraction, and scattering will occur, resulting in multipath effects, which significantly reduce the ranging accuracy. 2. Environmental factors may further aggravate the signal propagation problem. When tags and base stations are blocked or subject to electromagnetic interference, continuous data loss may occur, which is manifested as multiple base stations being unable to obtain the distance and RSSI measurements of the tags at the same time. 3. Large-scale data loss not only affects the accuracy of a single measurement, but also leads to a complete loss of target tracking capabilities for a long time.

[0005] Existing methods have made important progress in solving the problem of accurate positioning in complex environments. For example, traditional TOA-based methods mainly rely on Kalman filtering (KF) and various optimization algorithms; deep learning-based methods such as convolutional neural networks (CNN), long-short-term memory networks (LSTM) and graph neural networks (GNN) can significantly alleviate the error of UWB positioning in NLOS environments by learning complex environmental patterns in UWB data, such as spatiotemporal features and location features.

[0006] However, existing methods have the following shortcomings in dealing with positioning problems in complex environments: 1. Traditional TOA-based methods mainly rely on Kalman filtering (KF) and various optimization algorithms. When dealing with data loss, simple preprocessing techniques such as forward filling or linear interpolation are usually used to maintain basic data continuity, but the basic statistical characteristics of UWB measurements cannot be preserved. 2. Deep learning-based methods improve performance by learning complex environmental patterns in UWB data, but still rely on basic interpolation techniques when dealing with missing data. Simple data completion strategies cannot preserve the statistical distribution of UWB measurements, resulting in distribution shifts and potential data contamination, which significantly reduces positioning accuracy during long periods of signal obstruction. 3. Most existing methods fail to fully utilize multimodal information. Although some methods combine distance and RSSI data, these modalities are usually treated as independent inputs or simple splicing is performed, failing to capture their complex interdependencies and ignoring the complementary information contained in the RSSI signal.

[0007] In recent years, diffusion models have recently demonstrated remarkable capabilities in generative modeling and data reconstruction tasks. These models learn to reverse the gradual noise process, enabling them to generate high-quality samples from complex data distributions and effectively reconstruct missing data while preserving statistical properties. The advantage of diffusion models lies in their ability to model complex data distributions through iterative denoising steps, making them particularly suitable for addressing the challenges of data loss in UWB positioning systems.

[0008] The limitations of existing technologies indicate that an innovative UWB indoor positioning method is needed that can effectively handle large-scale data loss, fully fuse multimodal information, and adapt to complex NLOS environments. Summary of the invention

[0009] The present invention provides an ultra-wideband indoor positioning method based on a multimodal diffusion model, aiming to solve the problem of reduced positioning accuracy caused by large-scale data loss and signal interference in an NLOS environment.

[0010] The present invention adopts the following technical solution: an ultra-wideband indoor positioning method based on a multimodal diffusion model, comprising the following steps: S1. Data collection based on a mobile multi-sensor platform: The mobile multi-sensor platform is provided with a UWB base station, a laser radar and a computing unit, collects distance data and RSSI signal strength data between each UWB anchor point-tag in the indoor environment, obtains true value data through the laser radar, and obtains the true value of the motion trajectory of the mobile multi-sensor platform; S2, standardize the collected data, including: distance data processing, RSSI data processing, data missing processing, to obtain multi-modal input data; S3, performing feature extraction and feature fusion on multimodal input data through a multimodal feature extraction module; S4. Construct a multimodal diffusion model. Based on the forward Markov chain process of the diffusion model, use the improved denoising diffusion implicit model (DDIM) for reverse sampling, establish a U-Net architecture neural network, input the data after feature extraction and feature fusion, realize distance reconstruction through multi-step denoising iteration, and output the distance estimation value after training; S5, establish a position estimation layer, accept the distance estimation value input, use the long short-term memory network to perform time series modeling, learn the position estimation, and output the three-dimensional coordinate positioning value; S6. Conduct network training based on supervised learning and introduce a mixed loss function for iterative optimization to obtain a trained multimodal diffusion model, which is used to achieve high-precision indoor positioning in complex non-line-of-sight environments and effectively solve the problem of reduced positioning accuracy under conditions of signal interference and large-scale data loss.

[0011] Preferably, in step S1, the data collection based on the mobile multi-sensor platform includes the following sub-steps: S1.1. Several UWB anchor points are set up in the indoor environment, and each UWB anchor point is equipped with a UWB sensor; S1.2, fix the position of the UWB base station on the mobile multi-sensor platform and each UWB anchor point in the indoor environment, and determine the coordinates of the UWB base station and each UWB anchor point in a unified coordinate system through a calibration procedure; S1.3, control the mobile multi-sensor platform to move along a predetermined trajectory or randomly in the indoor environment, and establish data collection nodes through the ROS operating system; S1.4. Establish communication between the UWB base station on the mobile multi-sensor platform and each UWB anchor point in the indoor environment to obtain the distance value and RSSI signal strength value between the UWB anchor point and the tag at each timestamp to form different data sequences; obtain the reference position as the real-time true value through the laser radar on the mobile multi-sensor platform.

[0012] Preferably, in step S2, the standardization process includes the following sub-steps: S2.1, distance data processing: organize the distance data between multiple UWB anchor points and tags in parallel, aggregate the distance data of multiple consecutive timestamps to form a time window data block, and organize the aggregated data into a four-dimensional tensor structure ; S2.2, RSSI data processing: The RSSI signal strength data includes two RSSI signal strength values: received signal strength indicator data (RxRssi) and first path received signal strength indicator data (FpRssi), which respectively indicate the overall signal strength and the strength of the first arrival signal path; The two RSSI signal strength values ​​are organized through a tensor structure, and a channel is added to the first dimension to distinguish the two RSSI signal strength values ​​to obtain a tensor structure ; S2.3. Data missing processing: missing values ​​in the input data are filled using the forward filling method.

[0013] Preferably, in step S3, the multimodal feature extraction module adopts a parallel feature processing architecture to process RSSI signal strength data and distance data respectively, and performs feature extraction on the multimodal input data, and the method is as follows: S3.1.1, preprocessing and converting RSSI signal strength data through RSSI feature conversion module, mapping dual-channel features to a unified feature space; S3.1.2, performing multi-level feature extraction on the distance data and RSSI signal strength data along the distance path and RSSI path respectively, wherein the distance path and RSSI path use the same network structure but do not share parameters; during the feature extraction process, the interdependence between different modalities is learned through the convolution layer; The multi-level feature extraction includes a low-level feature layer, an intermediate feature layer and a high-level feature layer, forming a progressive feature extraction structure, and each layer performs feature transformation through 3D convolution; the low-level feature layer captures local spatial patterns, the intermediate feature layer integrates local features to form an abstract representation, and the high-level feature layer integrates spatial relationships to extract global semantic information.

[0014] Preferably, in step S3, feature fusion is performed on the multimodal input data in the following manner: S3.2.1. Set fusion points at multiple levels of feature extraction to combine the features of RSSI path and distance path. and Perform feature stitching; S3.2.2, perform feature transformation on the concatenated features through the convolution layer to learn the interdependence between different modalities; S3.2.3. Use 3×3×3 convolutional layers with channel numbers ranging from 32 to 16 to perform feature fusion and dimensionality reduction to form feature vectors at different scales.

[0015] Preferably, in step S4, a multimodal diffusion model is constructed, and an improved denoising diffusion implicit model is used to construct a Markov chain, and the method is as follows: S4.1. Based on the forward process of the diffusion model, a Markov chain is constructed, and the noise adding process is defined to gradually add noise to the original distance data. S4.2, adopt the linear noise scheduling strategy, use the reparameterization method to obtain any time step Noise data; S4.3, perform the reverse denoising process based on DDIM, use the improved denoising diffusion implicit model to convert the conditional probability into a deterministic form, and accelerate the sampling process of the diffusion model; S4.4, construct a time step embedding module to map discrete time steps into high-dimensional space; S4.5. Convert the temporal features through a two-layer fully connected network, map the 32-dimensional temporal embedding into a 64-dimensional feature vector, and inject it into each layer of the U-Net architecture neural network.

[0016] Preferably, in step S4, a U-Net architecture neural network is constructed as a backbone network of the diffusion model, including an encoder path and a decoder path; The encoder path extracts features step by step through multiple downsampling blocks, each of which contains two convolutional layers and a downsampling operation, while adding a projection of time-step embedding to adjust the noise prediction; The decoder path gradually restores the feature resolution through corresponding multiple upsampling operations, and fuses the features of the corresponding layers of the encoder through jump connections. Each upsampling module includes a deconvolution layer and two ordinary convolution layers.

[0017] Introduce skip connections between the corresponding encoder and decoder layers to solve the gradient vanishing problem: The U-Net architecture neural network outputs the predicted noise component , this output has the same size as the input data (1×T×A×N) and is used for deterministic distance reconstruction in the DDIM sampling process; Through the DDIM sampling strategy, the multimodal diffusion model is able to gradually recover high-quality distance measurements from noisy data.

[0018] Preferably, in step S5, the position estimation layer adopts a CNN-LSTM hybrid architecture to capture spatial relationships and time dependencies, and converts the reconstructed distance data into three-dimensional position coordinate outputs, including the following sub-steps: S5.1. Spatial feature extraction: Construct and design a convolutional neural network module to process multi-dimensional distance data through several 3D convolution blocks, extract spatial patterns in the anchor-label distance relationship, and preserve the temporal structure of the measurement; S5.2, Time Series Modeling: Use long short-term memory network to perform time series modeling on the processed spatial features, and use multiple continuous timestamp information to capture the time dependency in the target motion trajectory and enhance the time consistency of the positioning results; S5.3, Position coordinate regression: Construct a regression module, including a fully connected layer, to map temporal features to three-dimensional position coordinates; output the estimated position coordinates and perform end-to-end learning from distance measurement to position.

[0019] Preferably, in step S6, the hybrid loss function takes into account both the distance reconstruction error and the position estimation error, and balances the optimization target through weighted coefficients.

[0020] Preferably, in step S6, the network training based on supervised learning adopts an end-to-end training method to optimize parameters, including the following sub-steps: S6.1, loading the standardized distance data, RSSI signal strength data and the corresponding true value data into the training set and the test set; S6.2, sampling the real distance data and RSSI signal strength data, adding noise of a specified time step to generate training samples according to the forward process of the diffusion model, and performing forward calculation through the multimodal diffusion model to obtain the noise prediction value and the corresponding distance and position estimation; S6.3. Calculate the total loss, distance loss and position loss according to the mixed loss function; S6.4, updating the multimodal diffusion model parameters through the back propagation algorithm according to the gradient of the loss function; S6.5. Reverse sampling is performed through the denoising diffusion implicit model to generate reconstructed distance measurements from noisy data and predict the target position.

[0021] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects: 1. Effectively cope with non-line-of-sight environments and improve positioning accuracy: The indoor positioning method of the present invention significantly improves the accuracy of the UWB indoor positioning system through the application of multimodal data fusion and diffusion model. In complex non-line-of-sight (NLOS) environments, by fusing distance data and RSSI data, it effectively identifies and alleviates the measurement errors caused by multipath effects.

[0022] 2. Enhanced robustness in the event of data loss: The diffusion model of the present invention can effectively reconstruct lost measurement data while maintaining its statistical characteristics and temporal consistency. Even in the extreme case of large-scale data loss, it can still maintain acceptable positioning accuracy, and the increase in positioning error is much smaller than that of traditional methods.

[0023] 3. Improve computational efficiency: The present invention adopts the DDIM sampling strategy, which significantly reduces the number of sampling steps required for the diffusion model, improves the inference efficiency while maintaining the reconstruction quality, and can run in real time on resource-constrained embedded devices.

[0024] 4. Enhanced trajectory temporal consistency: The position estimation module based on the CNN-LSTM architecture of the present invention can effectively capture the temporal dependency in the motion trajectory, produce a smoother and more coherent positioning trajectory, and significantly reduce the trajectory jump phenomenon, especially in scenarios where the signal quality temporarily decreases.

[0025] 5. Strong adaptability: The ultra-wideband indoor positioning method of the present invention performs well in environments of different complexities, and as the complexity of the environment increases, its advantages over existing methods are more significant; from simple living room environments to complex laboratory environments, this method can maintain stable and reliable positioning performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 The figure is a frame diagram of the network structure of the ultra-wideband indoor positioning method of the present invention; Figure 2 This is a layout diagram of an experimental scenario of an embodiment of the present invention; Figure 3 This is a flowchart of the data processing phase of the ultra-wideband indoor positioning method of the present invention; Figure 4 It is a schematic diagram of multi-modal feature extraction and multi-scale feature fusion of the ultra-wideband indoor positioning method of the present invention. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the application is further elaborated in detail below in conjunction with the accompanying drawings. The described embodiments are only a part of the embodiments involved in the present invention. All non-innovative embodiments of other researchers in the field on this embodiment belong to the protection scope of the present invention. At the same time, for the step numbering in the embodiment of the present invention, it is only set for the convenience of explanation, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0028] The present invention proposes an ultra-wideband indoor positioning method based on a multimodal diffusion model, which is used to solve the data loss and signal interference problems faced by the UWB positioning system in a non-line-of-sight (NLOS) environment. An end-to-end system is formed by a data processing module, a multimodal feature extraction module, a U-Net architecture neural network based on a diffusion model, and a position estimation module. Through data input, feature fusion, loss calculation, and network training based on supervised learning, UWB distance data and RSSI signal strength data can be used for high-precision indoor positioning.

[0029] In one embodiment of the present invention, a complete experimental system is constructed to collect distance measurement data, RSSI signal strength data and distance true value data under multiple environments, and perform standardization processing.

[0030] like Figure 1 As shown, the standardized distance data, RSSI signal strength data and distance truth data are input into the multimodal denoising module, and the ultra-wideband indoor positioning method of the present invention is used to calculate the position estimation value and the distance estimation value, and perform indoor data estimation. Based on the network training of supervised learning, a hybrid loss function is introduced to iteratively optimize the position truth value and the distance truth value, and high-precision indoor positioning is achieved in a complex non-line-of-sight environment to verify the effectiveness of the method of the present invention.

[0031] The experimental system of this embodiment includes the following hardware devices: UWB sensors (using a configuration of 6 anchor points and 1 UWB base station, the anchor points are fixedly installed around the environment, and the UWB base station is installed on a mobile multi-sensor platform), a mobile multi-sensor platform (a robot platform with mobile capabilities, equipped with a UWB base station, a lidar, and an embedded computing unit).

[0032] Among them, the lidar is used to obtain a high-precision reference position as the true value, and the embedded computing unit is used to process the data and execute the positioning algorithm.

[0033] Before the experiment, the UWB coordinate system and the lidar coordinate system as well as the positions of the anchor points and UWB base stations in the coordinate system have been converted and aligned in advance to ensure that the measurements provided by all systems are running in a reference coordinate system.

[0034] In order to comprehensively evaluate the performance of this method in different scenarios, this embodiment designs three experimental environments of different complexity, such as Figure 2 As shown, it includes parameters such as different anchor point position coordinates, environment size, etc., as follows: Environment 1: Living room environment, such as Figure 2 As shown in (a), an indoor space of about 4m×8m with few obstacles and mainly line-of-sight (LOS) conditions is used as the benchmark test scenario; Environment 2: Apartment environment, such as Figure 2 As shown in (b), the multi-room residential environment of about 8m×12m contains obstacles such as walls and furniture, resulting in obvious non-line-of-sight conditions; Environment 3: Laboratory environment, such as Figure 2 As shown in (c), the large laboratory of about 30m×30m has the most complex spatial layout, severe non-line-of-sight environment and multipath effects, and is used to test the performance limits of positioning algorithms under extreme conditions.

[0035] In this embodiment, step S1 performs data collection based on a mobile multi-sensor platform to collect distance data and RSSI data between each UWB anchor point-tag in the indoor environment, obtains a high-precision reference position as the true value through a lidar, and obtains the true value of the motion trajectory of the mobile multi-sensor platform.

[0036] The specific method is as follows: In each experimental environment, the positions of the anchor points and UWB base stations are fixed, and their accurate coordinates in a unified coordinate system are determined through a pre-set calibration procedure; The mobile robot platform is controlled to move along a predefined or random trajectory in the test site, and a data acquisition node is established through the ROS operating system. The information of all sensors is recorded at the same time, including: UWB distance data, RSSI signal strength data, LiDAR data, and timestamps for synchronization of various data sources.

[0037] Among them, the RSSI signal strength data includes: the received signal strength indicator RxRssi and the first path received signal strength indicator FpRssi; the lidar data is used as the real positioning information value.

[0038] Repeat the experiment multiple times in different environments and scenarios to collect enough data samples to ensure the diversity and representativeness of the data. Finally, different data sequences are obtained, each with different duration, content, and trajectory.

[0039] In this embodiment, one sequence is selected from each sequence as a training set, and the rest are selected as validation sets, as shown in Table 1 below.

[0040] Table 1 The situation of each sequence in the experiment

[0041] In this embodiment, step S2 performs standardization processing on the collected data.

[0042] Data standardization is the first key step of the method of the present invention, including: distance data processing, RSSI data processing, data missing processing, to obtain multi-modal input data, such as Figure 3 As shown, the details are as follows:

[0043] First, considering the high sampling frequency (100 Hz) of UWB signals and the input requirements of the deep learning model, the distance data processing aggregates the distance measurement values ​​of 10 consecutive time stamps into one data block. This processing method helps to smooth instantaneous measurement fluctuations, reduce the amount of calculation, and retain spatial information while reducing the time granularity.

[0044] Organize the aggregated data into a four-dimensional tensor structure: ; Among them, 1 represents the initial channel dimension, is the time dimension (set to 16 in the experiment), is the number of anchor points (set to 6), is the combined sample size (set to 10). This data organization method effectively retains the temporal and spatial structure information of the data.

[0045] Then, for RSSI data, this embodiment adopts a processing method similar to that of distance data.

[0046] RSSI data contains two different components: RxRssi and FpRssi, which reflect the overall signal strength and the strength of the first arrival signal path, respectively.

[0047] The relationship between the two is particularly important for judging line-of-sight or non-line-of-sight conditions: When |RxRssi - FpRssi| < 6 dB, it is most likely a line-of-sight environment. When |RxRssi - FpRssi| > 10 dB, it is most likely a non-line-of-sight environment. When the difference is between 6-10 dB, the propagation condition is uncertain.

[0048] Organize RSSI data into a tensor structure : ; The superscript 2 in the first dimension represents two RSSI values.

[0049] Finally, data missing processing is performed. In an actual indoor environment, due to factors such as obstacle blocking, signal reflection and multipath effect, UWB distance data will often be missing during collection. This embodiment adopts a simple forward filling method to maintain data continuity.

[0050] In this embodiment, step S3 performs feature extraction and feature fusion on the multimodal input data through a multimodal feature extraction module.

[0051] Multimodal feature extraction is the second core component of the present invention, such as Figure 4As shown, a parallel feature processing architecture is designed to process UWB distance data and RSSI signal strength data separately.

[0052] First, the RSSI feature conversion module preprocesses and converts the RSSI data and maps the dual-channel features to a unified feature space, expressed as: ; in, It is a neural network module that includes convolutional layers, batch normalization, and ReLU activation functions. It is implemented as two consecutive 3D convolutional layers. The first layer maps 2-channel input to 8 channels, and the second layer maps 8 channels to 16 channels. Both use 3x3x3 convolution kernels and 1×1×1 strides. For passing HeadRSSI The output value after the layer.

[0053] Next, the UWB distance data and RSSI signal strength data are feature extracted along their respective processing paths. The two paths use the same network structure but do not share parameters, allowing each to independently learn the feature representation of a specific modality.

[0054] The convolution operation in the feature extraction process is designed as: ; in, represents a three-dimensional convolution operation, represents the convolution kernel size, represents the step length, , Represent the input and output of the convolution operation respectively.

[0055] The multiple levels include low-level feature layers, intermediate feature layers, and high-level feature layers, forming a progressive feature extraction structure. The low-level feature layer mainly captures local spatial patterns, the intermediate feature layer integrates local features to form a more abstract representation, and the high-level feature layer further integrates spatial relationships to extract global semantic information. Each level achieves feature transformation through 3D convolution operations while maintaining independence in the time dimension.

[0056] It is worth noting that although a three-dimensional convolution layer is used in this embodiment, the convolution kernel size of the depth dimension is set to 1. This is to ensure that the features of each time step are processed independently and to avoid implicit mixing of features between time steps.

[0057] Multimodal feature fusion is the third core component of the present invention. Fusion points are set at multiple levels of feature extraction. First, the features of the RSSI path and the distance path are spliced: ; in, represents the concatenation function, Represents the features after splicing.

[0058] The concatenated features are then transformed through a convolutional layer to learn the interdependencies between different modalities: ; in, represents the batch normalization operation, is a nonlinear activation function, Represents the fused features.

[0059] In the specific implementation, feature fusion and dimensionality reduction are performed using 3×3×3 convolutional layers with channels ranging from 16+16=32 to 16. This progressive feature fusion method not only ensures the independence of feature extraction, but also enhances the feature representation capability at multiple levels of abstraction, enabling the model to more effectively utilize multimodal information.

[0060] In this embodiment, step S4 constructs a multimodal diffusion model, based on the forward Markov chain process of the diffusion model, adopts an improved denoising diffusion implicit model (DDIM) for reverse sampling, establishes a U-Net architecture neural network, inputs data after feature extraction and feature fusion, realizes distance reconstruction through multi-step denoising iteration, and outputs a distance estimate after training.

[0061] The diffusion model is the core technology of the present invention, which is used for high-quality reconstruction of lost or heavily noisy distance measurement data. The forward process of the diffusion model is a process of gradually adding noise to the original data to form a Markov chain.

[0062] For any time step , this process is defined as: ; in, Represents the time step To time step The conditional probability distribution of adding noise, , Represents the time step , Noise data, I represents the identity matrix, represents a Gaussian normal distribution, is a predefined noise intensity scheduling parameter that controls the amount of added noise.

[0063] Using the reparameterization technique, we can directly obtain the noise data at any time step t: ; in, is the cumulative product, , is the original noise-free data, is random noise from a standard normal distribution.

[0064] In actual training, this embodiment adopts a linear noise scheduling strategy with 30 time steps. It increases linearly from a smaller value to a larger value, specifically from 0.0001 to 0.02.

[0065] In contrast to the forward process, the reverse process is the process of gradually recovering the original signal from the noisy data. However, the traditional diffusion model (DDPM) usually requires a large number of sampling steps, which reduces the efficiency of reasoning. To solve this problem, this embodiment uses an improved denoising diffusion implicit model (DDIM) to construct a Markov chain and convert the conditional probability into a deterministic form: ; in, represents the noise prediction network, implemented by the U-Net architecture; represents the control sampling variance.

[0066] When setting When , the sampling process becomes fully deterministic, significantly reducing the number of required sampling steps while maintaining the reconstruction quality.

[0067] To help the network effectively adapt to different noise levels and adjust the denoising strategy, this embodiment further designs a time step embedding module to map discrete time steps to a high-dimensional space: ; in, is the input dimension, is a predefined frequency factor, The time characteristic A quantity.

[0068] In particular, in this embodiment, Set to 32, Set to 10000.

[0069] Subsequently, these temporal features are further transformed through a two-layer fully connected network to map the 32-dimensional temporal embeddings into 64-dimensional feature vectors and injected into each layer of the U-Net.

[0070] The U-Net network structure in this embodiment includes an encoder path and a decoder path. The encoder path gradually extracts features through multiple downsampling modules, each of which includes two convolutional layers and a downsampling operation.

[0071] Taking the first downsampling module as an example, the number of input channels is 16×2=32 (the concatenation of distance features and multimodal features), the number of output channels is 32, the convolution kernel size is 3×3, the step size is 1×1, and the projection of the time step embedding is added to adjust the noise prediction.

[0072] The calculation process of the downsampling module is: ; in, represents the multimodal features from the auxiliary encoder, is the time step embedding information, Indicates Feature representation after layer downsampling, Represents a network module that contains feature extraction and downsampling operations.

[0073] The decoder path gradually restores the feature resolution through corresponding upsampling operations, while fusing the features of the corresponding layers of the encoder through skip connections.

[0074] The upsampling module consists of a deconvolution layer (transposed convolution) and two ordinary convolution layers, and the formula is ; in, Indicates The feature representation after upsampling of the layer, Represents a network module that contains feature extraction and upsampling operations.

[0075] For example, the first upsampling module reduces the number of feature channels from 32 to 16, uses a 2×2 deconvolution kernel with a stride of 2×2, and then connects two 3×3 convolutional layers for feature refinement.

[0076] To solve the gradient vanishing problem and promote multi-scale feature fusion, this embodiment further introduces a skip connection between the corresponding encoder and decoder layers: ; in, Represents the features after skip connection.

[0077] The network finally outputs the predicted noise component , with the same size as the input (1×16×6×10), is used for deterministic reconstruction during DDIM sampling. Through the DDIM sampling strategy, the model is able to gradually recover high-quality distance measurements from noisy data. This U-Net-based diffusion model architecture can effectively learn the noise distribution and thus recover high-quality distance measurement data from severely corrupted inputs.

[0078] In this embodiment, step S5 establishes a position estimation layer, accepts distance estimation value input, uses a long short-term memory network to perform time series modeling, learns position estimation, and outputs three-dimensional coordinate positioning values.

[0079] The position estimation module is the last core component of the present invention, which is responsible for converting the reconstructed distance data into the final three-dimensional position coordinates. This module adopts a CNN-LSTM hybrid architecture to effectively capture spatial relationships and temporal dependencies.

[0080] Spatial feature extraction is implemented by a series of 3D convolution operations to process multi-dimensional distance data and extract spatial patterns in the anchor-label distance relationship.

[0081] In this embodiment, the position estimation module includes 4 consecutive 3D convolution blocks, and the specific structure of each block is as follows: The first convolution block: input channel 1, output channel 16, convolution kernel (1,1,5), step size 1, batch normalization, sigmoid activation function; The second convolution block: input channels 16, output channels 32, convolution kernel (1,1,3), step size 1, batch normalization, sigmoid activation function; The third convolution block: input channel 32, output channel 64, convolution kernel (1,1,3), step size 1, batch normalization, Sigmoid activation function; Fourth convolution block: input channel 64, output channel 128, convolution kernel (1,1,2), stride 1, batch normalization, Sigmoid activation function.

[0082] These convolutional blocks gradually extract features, reduce the spatial dimension and increase the number of channels. In particular, the last fully connected layer compresses the 128×6 feature dimension to 128 dimensions.

[0083] It should be noted that the size of the convolution kernel in the time dimension is set to 1. This design is to maintain the independence between time steps and avoid mixing the features of different time steps.

[0084] In order to make full use of the information of multiple consecutive timestamps and capture the temporal dependency in the target motion trajectory, this embodiment uses a long short-term memory (LSTM) network for time series modeling.

[0085] The LSTM network has the following parameter configuration: input dimension 128, hidden layer dimension 100, 3 layers of LSTM stacking, batch_first=True is used to ensure that the first dimension of the input tensor is the batch size. The LSTM network can effectively capture long-term dependencies in time series, which is particularly important for smoothing trajectory prediction and improving positioning stability.

[0086] The position coordinate regression module consists of fully connected layers, which map the temporal features of the LSTM output to the final three-dimensional position coordinates. The first fully connected layer maps the 100-dimensional LSTM output features to a 60-dimensional intermediate representation, using the ReLU activation function to increase nonlinearity; the second fully connected layer maps the 60-dimensional features to the final 3D output, corresponding to the three-dimensional spatial coordinates (x, y, z).

[0087] The hybrid architecture design of CNN-LSTM in this embodiment effectively combines the advantages of convolutional neural networks in spatial feature extraction and the capabilities of recurrent neural networks in time series processing, and can accurately infer the location coordinates of the target from the reconstructed distance data.

[0088] In order to optimize both the distance reconstruction performance and the position estimation accuracy, this embodiment also designs a joint loss function that takes both objectives into consideration.

[0089] The loss function consists of two parts: distance loss (using mean square error to measure the difference between the predicted distance and the true distance) and position loss (also using mean square error to calculate the difference between the predicted position and the true position). The specific formula is as follows: ; in, represents the mixed loss function, is the distance loss, specifically the mean square error between the true distance value and the predicted distance value; is the position loss, specifically the mean square error between the true positioning value and the predicted positioning value; and It is a weight coefficient for balancing the two loss terms. In the experiments of this embodiment, both are set to 1.0, giving equal importance to distance reconstruction and position estimation.

[0090] The model training in this embodiment adopts an end-to-end approach, optimizes all parameters of the system at the same time, and combines an early stopping strategy to prevent overfitting.

[0091] First, the preprocessed UWB distance data, RSSI data and the corresponding real location data are loaded into the training set and test set, using batch sizes of 256 and 32 respectively; then the real distance data is sampled, noise is added to generate training samples, and forward calculation is performed through the model to obtain the predicted distance and location; then the total loss, distance loss and location loss are calculated according to the joint loss function; finally, the model parameters are updated through the back propagation algorithm according to the gradient of the loss function.

[0092] During the training process, the model performance is regularly evaluated on the test set, and key indicators such as distance mean square error, position root mean square error (RMSE) are monitored. This embodiment uses the Adam optimizer for training, with an initial learning rate of 5e-5, no weight decay (weight_decay=0), and an early stopping strategy to prevent overfitting.

[0093] In the inference stage, this embodiment uses the DDIM sampling strategy to efficiently generate reconstructed distance measurements from noise data, and then predicts the target position. The specific process is: first, the features of UWB distance data and RSSI data are extracted through a multimodal encoder, and then the DDIM sampling cycle is used to start from random noise, and through an iterative denoising process of 30 time steps, the reconstructed distance measurement data is generated, and finally the reconstructed distance data is converted into three-dimensional position coordinates through the position estimation module. Compared with traditional DDPM sampling, DDIM sampling can significantly reduce the number of sampling steps (from thousands of steps to 30 steps), greatly improve the inference efficiency, and maintain the sampling quality, making this method suitable for real-time application scenarios.

[0094] In order to comprehensively evaluate the performance of the method of the present invention, this embodiment further designs a series of comparative experiments and tests its accuracy and robustness under different environments.

[0095] First, the root mean square error (RMSE) is used to measure the average distance error between the predicted position and the true position. Transformer-LSTM (a deep learning method based on Transformer and LSTM), STA-GNN-S (a positioning method based on spatiotemporal attention graph neural network), KF-LSTM (a hybrid method combining Kalman filter and LSTM) and CNN-LSTM (a traditional deep learning positioning method using CNN and LSTM) are selected as comparison methods. The experimental results show that the proposed method is superior to the comparison methods in all test scenarios.

[0096] More importantly, comparative experiments under different data missing rates (70%, 80%, 90%) show that when the data missing rate increases, the performance of all methods decreases, but the performance of the method of the present invention decreases the least. Even in the case of 90% data missing, the method of the present invention can still maintain a high positioning accuracy, while the performance of the comparison method drops sharply. This proves the excellent ability of the method of the present invention in dealing with large-scale data loss. Trajectory visualization analysis also shows that in the most challenging laboratory environment, the trajectory reconstructed by the method of the present invention is closest to the true trajectory, and can maintain a high tracking accuracy even in non-line-of-sight areas, while other methods often have large deviations or tracking loss when encountering non-line-of-sight areas.

[0097] In summary, the ultra-wideband indoor positioning method based on the multimodal diffusion model proposed in the present invention effectively solves the problems of large-scale data loss and signal interference in non-line-of-sight environments by combining multimodal feature extraction, diffusion model and position estimation technology, and significantly improves the performance and robustness of the UWB indoor positioning system. This method performs well in various complex environments, especially in the case of poor data quality, and provides strong support for high-precision indoor positioning in practical applications.

[0098] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. An ultra-wideband indoor positioning method based on a multimodal diffusion model, characterized in that: The steps include: S1. Data collection based on a mobile multi-sensor platform: The mobile multi-sensor platform is provided with a UWB base station, a laser radar and a computing unit, collects distance data and RSSI signal strength data between each UWB anchor point-tag in the indoor environment, obtains true value data through the laser radar, and obtains the true value of the motion trajectory of the mobile multi-sensor platform; S2, standardize the collected data, including: distance data processing, RSSI data processing, data missing processing, to obtain input data; S3, extracting and fusing features of input data through a multimodal feature extraction module; S4. Construct a multimodal diffusion model. Based on the forward Markov chain process of the diffusion model, use the improved denoising diffusion implicit model for reverse sampling, establish a U-Net architecture neural network, input the data after feature extraction and feature fusion, realize distance reconstruction through multi-step denoising iteration, and output the distance estimation value after training; S5, establish a position estimation layer, accept the distance estimation value input, use the long short-term memory network to perform time series modeling, learn the position estimation, and output the three-dimensional coordinate positioning value; S6. Conduct network training based on supervised learning, introduce a mixed loss function for iterative optimization, and obtain a trained multimodal diffusion model for high-precision indoor positioning in complex non-line-of-sight environments.

2. The ultra-wideband indoor positioning method based on a multimodal diffusion model according to claim 1, characterized in that: In step S1, data collection based on a mobile multi-sensor platform includes the following sub-steps: S1.

1. Several UWB anchor points are set up in the indoor environment, and each UWB anchor point is equipped with a UWB sensor; S1.2, fix the position of the UWB base station on the mobile multi-sensor platform and each UWB anchor point in the indoor environment, and determine the coordinates of the UWB base station and each UWB anchor point in a unified coordinate system through a calibration procedure; S1.3, control the mobile multi-sensor platform to move along a predetermined trajectory or randomly in the indoor environment, and establish data collection nodes through the ROS operating system; S1.

4. Establish communication between the UWB base station on the mobile multi-sensor platform and each UWB anchor point in the indoor environment to obtain the distance value and RSSI signal strength value between the UWB anchor point and the tag at each timestamp to form different data sequences; obtain the reference position as the real-time true value through the laser radar on the mobile multi-sensor platform.

3. The ultra-wideband indoor positioning method based on a multimodal diffusion model according to claim 1, characterized in that: In step S2, the standardization process includes the following sub-steps: S2.1, distance data processing: organize the distance data between multiple UWB anchor points and tags in parallel, aggregate the distance data of multiple consecutive timestamps to form a time window data block, and organize the aggregated data into a four-dimensional tensor structure , which is expressed as follows: ； Among them, 1 represents the initial channel dimension, T is the time dimension, A is the number of anchor points, and N is the combined sample size; S2.2, RSSI data processing: The RSSI signal strength data includes two RSSI signal strength values, namely: received signal strength indicator data and first path received signal strength indicator data, which respectively represent the overall signal strength and the strength of the first arrival signal path; The two RSSI signal strength values ​​are organized through a tensor structure, and a channel is added to the first dimension to distinguish the two RSSI signal strength values ​​to obtain a tensor structure , which is expressed as follows: ； Among them, the superscript 2 in the first dimension represents two RSSI signal strength values; S2.

3. Data missing processing: missing values ​​in the input data are filled using the forward filling method.

4. The ultra-wideband indoor positioning method based on a multimodal diffusion model according to claim 3, characterized in that: In step S3, the multimodal feature extraction module adopts a parallel feature processing architecture to process the distance data and RSSI signal strength data respectively, and extracts features from the input data in the following manner: S3.1.

1. Preprocess and convert the RSSI signal strength data through the RSSI feature conversion module, and map the dual-channel features to a unified feature space. The expression is: ； Among them, HeadRSSI is a neural network module containing convolutional layers, batch normalization and ReLU activation functions, including two consecutive 3D convolutional layers. The first 3D convolutional layer maps 2-channel input to 8 channels, and the second 3D convolutional layer maps 8 channels to 16 channels. For passing HeadRSSI Output value after the layer; S3.1.2, performing multi-level feature extraction on the distance data and RSSI signal strength data along the distance path and RSSI path respectively, wherein the distance path and RSSI path use the same network structure but do not share parameters; During feature extraction, the interdependencies between different modalities are learned through convolutional layers. The convolution operation is expressed as: ； in, represents a three-dimensional convolution operation, represents the convolution kernel size, represents the step length, , Represent the input and output of the convolution operation respectively; The multi-level feature extraction includes a low-level feature layer, an intermediate feature layer and a high-level feature layer, forming a progressive feature extraction structure, and each layer performs feature transformation through 3D convolution; the low-level feature layer captures local spatial patterns, the intermediate feature layer integrates local features to form an abstract representation, and the high-level feature layer integrates spatial relationships to extract global semantic information.

5. The ultra-wideband indoor positioning method based on a multimodal diffusion model according to claim 4, characterized in that: In step S3, feature fusion is performed on the multimodal input data in the following manner: S3.2.

1. Set fusion points at multiple levels of feature extraction to combine the features of RSSI path and distance path. and Perform feature stitching: ； in, represents the concatenation function, Represents the features after splicing; S3.2.2, the concatenated features are transformed through the convolution layer to learn the interdependence between different modalities: ； in, represents the batch normalization operation, is a nonlinear activation function, Represents the fused features; S3.2.

3. Use 3×3×3 convolutional layers with channel numbers ranging from 32 to 16 to perform feature fusion and dimensionality reduction to form feature vectors at different scales.

6. The ultra-wideband indoor positioning method based on a multimodal diffusion model according to claim 5, characterized in that: In step S4, the improved denoising diffusion implicit model is used to construct the Markov chain, and the method is as follows: S4.

1. Based on the forward process of the diffusion model, a Markov chain is constructed, and the noise adding process is defined to gradually add noise to the original distance data. , the forward process is defined as: ； in, Represents the time step To time step The conditional probability distribution of adding noise, , Represents the time step , Noise data, I represents the identity matrix, represents a Gaussian normal distribution, is a predefined noise intensity scheduling parameter used to control the amount of added noise; S4.2, adopt the linear noise scheduling strategy, use the reparameterization method to obtain any time step Noise data: ； in, is the cumulative product, , is the original noise-free data, is the random noise of standard normal distribution; S4.3, perform the reverse denoising process based on DDIM, use the improved denoising diffusion implicit model to convert the conditional probability into a deterministic form, and accelerate the sampling process of the diffusion model: ； in, represents the noise prediction network, implemented by the U-Net architecture; Indicates the control sampling variance. When set When , the sampling process becomes completely deterministic; S4.

4. Build a time step embedding module to map discrete time steps to high-dimensional space: ； in, is the input dimension, is a predefined frequency factor, The time characteristic Quantity; S4.

5. Convert the temporal features through a two-layer fully connected network, map the 32-dimensional temporal embedding into a 64-dimensional feature vector, and inject it into each layer of the U-Net architecture neural network.

7. The ultra-wideband indoor positioning method based on a multimodal diffusion model according to claim 6, characterized in that: In step S4, a U-Net architecture neural network is constructed as a backbone network of the diffusion model, including an encoder path and a decoder path; The encoder path extracts features step by step through multiple downsampling blocks. Each downsampling module contains two convolutional layers and a downsampling operation. At the same time, the projection of the time step embedding is added to adjust the noise prediction. The calculation process is: ； in, represents the multimodal features from the auxiliary encoder, is the time step embedding information, It represents the feature representation after downsampling of the i-th layer, and DownBlock represents the network module including feature extraction and downsampling operations; The decoder path gradually restores the feature resolution through corresponding multiple upsampling operations, and fuses the features of the corresponding layers of the encoder through jump connections. Each upsampling module contains a deconvolution layer and two ordinary convolution layers. The calculation process is: ； in, It represents the feature representation after upsampling of the i-th layer, and UpBlock represents the network module including feature extraction and upsampling operations; Introduce skip connections between the corresponding encoder and decoder layers to solve the gradient vanishing problem: ； in, Represents the features after skip connection; The U-Net architecture neural network outputs the predicted noise component , the noise component Same size as input data, used for deterministic distance reconstruction during DDIM sampling.

8. The ultra-wideband indoor positioning method based on a multimodal diffusion model according to claim 7, characterized in that: In step S5, the position estimation layer uses a CNN-LSTM hybrid architecture to capture spatial relationships and temporal dependencies, and converts the reconstructed distance data into three-dimensional position coordinate outputs, including the following sub-steps: S5.

1. Spatial feature extraction: Construct and design a convolutional neural network module to process multi-dimensional distance data through several 3D convolution blocks, extract spatial patterns in the anchor-label distance relationship, and preserve the temporal structure of the measurement; S5.2, Time Series Modeling: Use long short-term memory network to perform time series modeling on the processed spatial features, and use multiple continuous timestamp information to capture the time dependency in the target motion trajectory and enhance the time consistency of the positioning results; S5.3, Position coordinate regression: Construct a regression module, including a fully connected layer, to map temporal features to three-dimensional position coordinates; output the estimated position coordinates and perform end-to-end learning from distance measurement to position.

9. The ultra-wideband indoor positioning method based on a multimodal diffusion model according to claim 8, characterized in that: In step S5.1, the convolutional neural network module includes 4 consecutive 3D convolution blocks, which gradually extract and refine spatial features, reduce spatial dimensions and increase the number of channels; the structure of each 3D convolution block is as follows: The first convolution block: input channel 1, output channel 16, convolution kernel (1,1,5), step size 1, batch normalization, sigmoid activation function; The second convolution block: input channels 16, output channels 32, convolution kernel (1,1,3), step size 1, batch normalization, sigmoid activation function; The third convolution block: input channel 32, output channel 64, convolution kernel (1,1,3), step size 1, batch normalization, Sigmoid activation function; Fourth convolution block: input channel 64, output channel 128, convolution kernel (1,1,2), stride 1, batch normalization, Sigmoid activation function.

10. The ultra-wideband indoor positioning method based on a multimodal diffusion model according to claim 8, characterized in that: In step S6, the network training based on supervised learning uses an end-to-end training method to optimize parameters, including the following sub-steps: S6.1, loading the standardized distance data, RSSI signal strength data and the corresponding true value data into the training set and the test set; S6.2, sampling the real distance data and RSSI signal strength data, adding noise of a specified time step to generate training samples according to the forward process of the diffusion model, and performing forward calculation through the multimodal diffusion model to obtain the noise prediction value and the corresponding distance and position estimation; S6.

3. Calculate the total loss, distance loss and position loss according to the mixed loss function; The hybrid loss function takes into account both the distance reconstruction error and the position estimation error, and balances the optimization objectives through weighted coefficients. The formula is as follows: ； in, For the distance loss, the mean square error between the true distance value and the predicted distance value is used. is the position loss, using the mean square error between the true positioning value and the predicted positioning value, and is the weight coefficient that balances distance loss and position loss, represents the mixed loss function; S6.4, updating the multimodal diffusion model parameters through the back propagation algorithm according to the gradient of the loss function; S6.

5. Generate reconstructed distance measurements from noisy data and predict target locations by performing reverse sampling via a denoising diffusion implicit model.

Images