A visible light positioning method and system based on time-space-angle feature fusion

Abstract

This invention proposes a visible light positioning method and system based on spatiotemporal and angular feature fusion, relating to the field of indoor positioning technology. The method includes: constructing a training dataset based on an indoor positioning scene, including visible light positioning pilot maps, coordinate labels, distance labels, and angle labels; inputting the training dataset into a positioning model for training, extracting preliminary features through a feature embedding module, and then capturing the temporal dependencies of the preliminary features and enhancing spatial perception through a feature fusion module to obtain fused features; inputting the fused features into position extraction branches and angle extraction branches at different depths to obtain position information and angle information respectively; calculating a loss function until the loss converges to obtain a trained positioning model; inputting the visible light positioning pilot map of the receiver to be positioned into the trained positioning model to obtain positioning information. This achieves high-precision and robust 3D positioning, effectively solving the problems of insufficient feature mining and poor scene adaptability in traditional VLP technology.

Description

Technical Field

[0001] This invention relates to the field of indoor positioning technology, and in particular to a visible light positioning method and system based on spatiotemporal-angle feature fusion. Background Technology

[0002] Visible light indoor positioning (VLP) utilizes LED lighting to achieve high-precision position calculations and is a mainstream positioning technology suitable for complex indoor scenarios. As indoor positioning scenarios become increasingly complex, deep learning modeling solutions have become a core trend driving the practical application of VLP technology, but existing methods still have shortcomings in engineering implementation.

[0003] Current deep learning-driven VLP positioning research and applications are limited to ideal environments where the positioning device and the light source plane are parallel and only two-dimensional translation is performed, without fully exploring the multi-dimensional data features of the positioning system. At the same time, existing models lack robust optimization design and are difficult to adapt to complex working conditions with multiple interferences and changing postures indoors, which restricts the practical application of visible light indoor positioning technology. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a visible light positioning method and system based on spatiotemporal-angle feature fusion. By integrating spatiotemporal-angle feature fusion with multi-task collaborative learning, it adapts to complex indoor scenes, achieving high-precision and robust 3D positioning. This effectively solves the problems of insufficient feature mining and poor scene adaptability in traditional VLP technology.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a visible light positioning method based on spatiotemporal-angular feature fusion, comprising: A training dataset was constructed based on indoor positioning scenarios, including visible light positioning pilot maps, coordinate labels, distance labels, and angle labels; The training dataset is input into the localization model for model training. Preliminary features are extracted through the feature embedding module, and then the temporal dependence of the preliminary features is captured and spatial awareness is enhanced through the feature fusion module to obtain fused features. The fused features are input into the position extraction branch and angle extraction branch at different depths to obtain position information and angle information respectively. The loss function is calculated and the parameters are updated until the loss function converges to obtain the trained localization model. The visible light positioning pilot diagram of the receiver to be located is input into the trained positioning model to obtain positioning information.

[0006] Secondly, the present invention provides a visible light positioning system based on spatiotemporal-angular feature fusion, comprising: The data acquisition module is configured to build a training dataset based on the indoor positioning scenario, including visible light positioning pilot maps, coordinate labels, distance labels, and angle labels; The model training module is configured to input the training dataset into the localization model for model training; wherein, preliminary features are extracted through the feature embedding module, and then the temporal dependency of the preliminary features is captured and spatial awareness is enhanced through the feature fusion module to obtain fused features; the fused features are input into the position extraction branch and angle extraction branch at different depths to obtain position information and angle information respectively; the loss function is calculated and the parameters are updated until the loss function converges to obtain the trained localization model; The positioning module is configured to input the visible light positioning pilot map of the receiver to be positioned into the trained positioning model to obtain positioning information.

[0007] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the visible light positioning method based on spatiotemporal-angle feature fusion as described in the first aspect.

[0008] Fourthly, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the visible light positioning method based on spatiotemporal-angular feature fusion as described in the first aspect.

[0009] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention acquires visible light signals from multiple sources and constructs a multidimensional dataset containing intensity, temporal sequence, and rotation angle. Through the synergistic effect of temporal, spatial, and angular feature encoding branches, it achieves deep fusion and shared representation of temporal, spatial, and angular features. Simultaneously, by combining a multi-task learning framework to synchronously estimate position coordinates, transmission distance, and rotation angle, it effectively overcomes the shortcomings of traditional methods, such as one-sided feature extraction and poor scene adaptability. This invention is adaptable to complex scenarios involving mixed LoS+NLoS channels and random device rotation, breaking through the limitations of 2D positioning. It not only significantly improves positioning accuracy and stability, achieving centimeter-level high-precision positioning, but also enhances the model's environmental adaptability and scalability, providing reliable positioning support for scenarios such as the Industrial Internet of Things.

[0010] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0011] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute a limitation thereof.

[0012] Figure 1 The main flowchart of a visible light positioning method based on spatiotemporal-angular feature fusion provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of an indoor VLP scene provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the data received by the positioning device at different angles according to an embodiment of the present invention; Figure 4 This is a schematic diagram comparing the traditional dataset provided in this embodiment of the invention with the dataset of this embodiment; Figure 5 A schematic diagram of a model provided for an embodiment of the present invention; Figure 6 The Mamba enhancement module provided in this embodiment of the invention; Figure 7 The training losses for the three datasets provided in this embodiment of the invention under different models are shown; where (a) is the training loss of the LSTM-KAN model; and (b) is the training loss of the Attention-MLP model. Figure 8 The localization accuracy of the LSTM-KAN model trained using single signal-to-noise ratio datasets, mixed signal-to-noise ratio datasets, and mixed signal-to-noise ratio datasets with angular components is compared in this embodiment of the invention; where (a) is SNR = 10dB; (b) is SNR = 40dB; and (c) is the average loss under different signal-to-noise ratios. Figure 9 The localization accuracy of the Attention-MLP model trained using single signal-to-noise ratio datasets, mixed signal-to-noise ratio datasets, and mixed signal-to-noise ratio datasets with angular components is compared in this embodiment of the invention; where (a) is SNR = 10dB; (b) is SNR = 40dB; and (c) is the average loss under different signal-to-noise ratios. Figure 10 The present invention provides a comparison of the localization accuracy of four algorithm training models under different signal-to-noise ratios (SNRs) in embodiments of the invention; wherein, (a) is SNR = 10dB; (b) is SNR = 40dB; (c) is for multiple SNRs; ​​(d) is a box plot for SNR = 10; (e) is a box plot for SNR = 10; and (f) is a box plot for multiple SNRs. Figure 11 This is a schematic diagram comparing the performance of different signal-to-noise ratios provided in the embodiments of the present invention; wherein, (a) is the positioning accuracy of the model without fine-tuning and only with increased distance loss; (b) is the positioning accuracy of the model without fine-tuning and with increased distance loss and angle; and (c) is the positioning accuracy of the model with fine-tuning and with increased distance loss and angle. Figure 12Ablation experiments of the Mamba enhancement model provided in this embodiment of the invention on datasets of different sizes; (a) SNR = 10dB; (b) SNR = 40dB; (c) different signal-to-noise ratios; Figure 13 This is a schematic diagram comparing the training loss of the STAK-Ploc model and the Mambareinforcement model on datasets of different sizes provided in this embodiment of the invention; where (a) is a schematic diagram of training error; and (b) is a schematic diagram of validation loss. Detailed Implementation

[0013] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0014] Indoor positioning is a core supporting technology for the integration of Industrial Internet of Things (IIoT) and mobile computing. It enables the location tracking of equipment or personnel by capturing specific signals in the environment (such as Wi-Fi, Bluetooth, and visible light), playing an irreplaceable role in complex industrial scenarios such as production scheduling and resource management. Among these technologies, visible light positioning (VLP) has become the preferred solution for meeting high-precision positioning needs due to its advantages such as abundant spectrum resources, resistance to electromagnetic interference, high positioning accuracy, and controllable deployment costs. In particular, non-imaging VLP systems based on LED-PDs are more suitable for the actual application needs of industrial scenarios.

[0015] As industrial scenarios increasingly demand real-time and dynamic positioning services, existing VLP technologies are becoming increasingly inadequate for complex application scenarios: First, traditional positioning methods rely solely on Received Signal Strength (RSS) as the core criterion, resulting in an overly simplistic feature dimension. In indoor environments, light reflects off walls and equipment, creating multipath signals. This, combined with ambient light interference from natural light and stray lights, causes irregular fluctuations in RSS values. The algorithm cannot distinguish whether signal attenuation is due to location changes, external interference, or light path reflection. This signal distortion leads to inaccurate positioning calculations, resulting in weak anti-interference capabilities in complex industrial environments and compromising overall positioning robustness.

[0016] Secondly, traditional positioning methods mostly focus on 2D fixed-plane positioning, failing to consider complex variables in real-world scenarios such as device rotation and non-line-of-sight (NLoS) transmission. Since photodiodes (PDs) are not omnidirectional receivers and have a fixed field of view (FOV), their received light intensity is highly correlated with the incident angle of the LED light. When the device rotates, the PD sensor's orientation changes accordingly, and the originally perpendicularly incident LED light becomes obliquely incident. This results in significant attenuation and abrupt changes in the received RSS signal. In extreme cases, the corresponding LED light source may even move out of the PD's field of view, causing temporary loss of light signal and distortion of feature data.

[0017] Furthermore, even though some deep learning solutions achieve centimeter-level accuracy, they often directly extract static features from single-moment RSS signals, utilizing only instantaneous spatial sampling information for positioning modeling. On one hand, they fail to capture the temporal variation patterns of LED-PD signals over time, making it difficult to distinguish between instantaneous noise and true positioning features. On the other hand, they do not fully utilize the spatial distribution correlation information of multiple light sources, processing only individual LED signals in isolation. Simultaneously, they do not incorporate changes in device posture and receiving angle as effective feature inputs to the model, failing to compensate for signal distortion caused by the limited field of view of the PD. The models are only suitable for small-scale, fixed-posture, fixed-layout scenarios, and struggle to adapt to changes in light source layout, device posture deflection, and signal temporal disturbances, thus exhibiting insufficient generalization ability and adaptability to complex environments.

[0018] It is evident that the core shortcomings of the current technology are concentrated in the following aspects: First, feature extraction is one-sided, failing to effectively integrate key information such as signal temporal dependence, spatial distribution patterns, and device rotation angles, resulting in positioning accuracy being significantly affected by the environment and device status. Second, the model has insufficient adaptability, making it difficult to cope with complex scenarios such as mixed LoS+NLoS channels and random device rotation, limiting it to specific application conditions. In addition, the learning framework is singular, failing to improve the consistency of positioning-related parameter estimation through multi-task collaborative learning, affecting the stability and reliability of positioning results. Therefore, a more robust and comprehensive high-precision positioning solution is urgently needed.

[0019] Based on this, this invention proposes a visible light positioning method, system, medium, and device based on temporal-spatial-angle feature fusion. First, it studies the influence of each received signal component on positioning accuracy and proposes a strategy for training the positioning model using mixed signal-to-noise ratio data. Then, it analyzes the temporal, spatial, and angular features in the received signal strength and constructs a positioning framework based on a cascaded feature fusion module of LSTM-Attention to achieve progressive training of "encoding-fusion-decoupling". Furthermore, it can introduce the Mamba module according to actual needs to utilize its ability to dynamically capture long sequence features and achieve high-precision positioning.

[0020] Example 1 like Figure 1 As shown, this embodiment discloses a visible light positioning method based on spatiotemporal-angular feature fusion, including the following steps: S1: Construct a training dataset based on indoor positioning scenarios, including visible light positioning pilot maps, coordinate labels, distance labels, and angle labels; S2: Input the training dataset into the localization model for model training; wherein, preliminary features are extracted through the feature embedding module, and then the temporal dependence of the preliminary features is captured and spatial awareness is enhanced through the feature fusion module to obtain fused features; the fused features are input into the position extraction branch and angle extraction branch at different depths to obtain position information and angle information respectively; the loss function is calculated and the parameters are updated until the loss function converges to obtain the trained localization model; S3: Input the visible light positioning pilot map of the receiver to be located into the trained positioning model to obtain positioning information.

[0021] Next, combined Figure 1 This embodiment provides a detailed description of a visible light positioning method based on spatiotemporal-angle feature fusion.

[0022] I. System Modeling For hybrid scenarios of line-of-sight (LoS) and non-line-of-sight (NLoS) positioning, an indoor visible light positioning system model is constructed.

[0023] (a) Indoor MIMO-VLP System This embodiment creates a... Figure 2 The indoor VLP scene shown has an internal size of [size missing]. There are a number of LEDs on the ceiling that serve both lighting and communication functions. A positioning device R is equipped with a signal receiving array composed of photodiodes. PD It moves randomly in space, and it possesses... Random rotation angles in three directions.

[0024] Assuming that the signal receiving array is uniformly distributed and its size is much smaller than the distance the device moves per unit time, the entire MIMO system can be approximated as a MIMO-VLP system during the modeling process.

[0025] (ii) Visible light propagation channel In a real-world indoor VLP environment, communication data emitted from the LED will travel through multiple channels to reach the receiving device, such as... Figure 2 As shown, the channel in these channels where the light source directly points to the PD receiver is called the line-of-sight (LOS) transmission channel. Its channel gain has a complex nonlinear relationship with the distance and angle between the two, as shown in Equation 1: (1) in, , The 'd' symbol represents the half-power angle of the LED, and the 'd' symbol represents the distance between the LED and the PD. It is the field of view of the PD. It is the receiving area of ​​the PD. and These are the radiation angle and the incident angle of the PD, respectively. It is the gain of the optical filter. It is the gain of the optical focuser.

[0026] The channel that is reflected by objects in the indoor space, such as walls, and points to the PD receiver is called a non-line-of-sight transmission channel, such as... Figure 2 As shown, assuming the number is The LED signal enters the receiving device after being reflected once from the walls in the four directions (north, south, west, and east) in the indoor environment. Its non-line-of-sight channel can be represented as... , with one of the sub-channels For example, it is expressed as follows: (2) in, Indicates the energy reflectivity of the wall surface. This indicates the area of ​​the reflective region on the wall surface, with subscripts 11 and 12 representing the transmitter and receiver, respectively.

[0027] Assuming the wall is perpendicular to the ground and has a flat surface, the inner wall acts as a mirror for the LED projection. The focal point of the line connecting the projection to the PD and the inner wall is the center coordinate of the reflecting surface, denoted as [missing information]. .

[0028] Because the receiving surface area of ​​the PD is small, the area of ​​the reflection zone on the wall is: .

[0029] (iii) The positioning device rotates. Unlike traditional wireless positioning devices, the signal radiation range of visible light communication devices is consistent with the direction of light radiation, rather than the full angle. In the positioning scenario of a portable mobile terminal with degrees of freedom, according to formulas (1) and (2), the channel gain is greater than 0 when the incident angle is less than or equal to the field of view angle of the PD.

[0030] exist Figure 3 middle Signals were received from two channels: one from direct light from the light source and the other from reflection off the wall. After rotation, the direct angle of the light source is greater than the field of view, so only reflected signals from the wall can be received; because the radiation space of the light source is cone-shaped, there is a certain radiation blind zone, for example... Since it cannot reflect light input, there is no NLoS channel for data transmission. Using a plane parallel to the ceiling as the reference plane for the rotation angle, the self-rotation angle of the device to be positioned is defined. .

[0031] In 3D indoor scenes, in order to better describe Represent it as follows Figure 2 The PDs in the middle are respectively based on the circumference of the PDs. The rotational angular components of the axis Formal representation, denoted as Similarly, the radiation angle Expressed in angular component form as .

[0032] When PD forms an angle with the reference plane, that is When the incident angle is not equal to 0, it is expressed in terms of angular components. Similarly .

[0033] II. Problem Analysis and Model Building This embodiment introduces the features of the model training data, the Mamba model, and the design of the loss function.

[0034] (I) Deep Model Training Data Analysis Traditional indoor VLP algorithms primarily rely on approximate modeling of the indoor environment or large-scale sampling (fingerprint database methods). When achieving high-precision positioning driven by received data, the time-domain expression of the received signal model is incorporated. (X represents the input signal,) (where N is the channel noise) It can be observed that even in a stable and ideal environment, the modeling process of the channel response is still quite complex; if the sampling data has a deviation and is affected by the continuously changing signal-to-noise ratio, the stability of the positioning process will be unpredictable.

[0035] Because deep learning-based localization algorithms have the advantage of only needing to extract hidden features from data through neural networks and do not rely on specific system modeling, they offer significant improvements in robustness and reliability compared to traditional localization methods.

[0036] However, the performance of deep learning localization models is highly dependent on training data, and high-quality datasets are a key factor in building reliable deep learning-based localization models. , Furthermore, there is a multiple T between the gains of the NLOS channel and the LOS channel: ; According to research, the area of ​​a typical photosensitive surface of a photodiode (PD) is no more than 1 cm². 2 Therefore, T=[0,0.0008]. Thus, the NLOS channel transmission signal can be approximated as a noise disturbance relative to the LOS channel transmission signal. Therefore, the time-domain expression can be transformed into... Previous research has demonstrated that datasets constructed by extracting pilot signals can effectively improve the training speed and localization accuracy of models in Loss scenarios. Therefore, pilot-based datasets are still used to train localization models in this scenario.

[0037] Furthermore, system noise, as a significant factor affecting positioning accuracy, is highly random and influenced by any object in the environment. Therefore, to further improve the robustness of the training model, traditional training methods typically collect data with the same communication environment and signal-to-noise ratio as samples. However, this embodiment uses mixed signal-to-noise ratio data to train the model. In a stable indoor environment, the amount of data at each signal-to-noise ratio should, as far as possible, follow a normal distribution according to the actual communication conditions. In unstable environments, a single-point continuous sampling method can be used to construct the dataset, achieving an optimal balance between training efficiency and positioning performance.

[0038] In addition, when the number, layout, and radiation angle of the LEDs used for positioning are determined, the probability of three types of valid data (pure line-of-sight data, non-line-of-sight-dominated data, and mixed line-of-sight data) appearing at the receiving end follows a fixed spatial distribution. To ensure that the neural network can correctly capture the spatial features in the data, sampling can be performed according to this spatial distribution when constructing the dataset. However, when the receiving device has degrees of freedom, its angle at a certain moment is an independent and random event, which causes the inherent spatial distribution of the system itself to fail and is difficult to function. To combat this reduction in positioning efficiency caused by data imbalance, this embodiment adds angle information to the data label and then samples data from multiple angles at the same location, enabling the model to effectively capture the influence of angle features on the data during training, thereby improving the positioning accuracy.

[0039] The core of the localization task is position estimation. Unlike conventional techniques that embed angle features into the model and regulate the localization error in the form of parameter constraints, this embodiment uses angle information as data labels. During the training process, the model can perceive the signal differences at different receiving angles at the same location, thereby effectively combating the decline in localization efficiency caused by data imbalance, enhancing model robustness, and accurately capturing the impact of angle features on the data.

[0040] Figure 4 This demonstrates the differences in format between the old and new datasets, among which Indicates the amount of data in the dataset. Indicates the number of sampled curves. and This indicates the number of sampling points on each sampling curve. This indicates the number of data points sampled at different rotation angles for each sampling point.

[0041] Based on the conclusions of previous studies, this embodiment still uses a dataset composed of signals extracted from pilot signals, and inputs the real and imaginary parts of the complex signal into the network after splitting them. That is, the input data includes the pilot complex baseband signal extracted from the received signal Y, the actual physical coordinates of the device, and the self-rotation angle.

[0042] (II) Multi-task learning localization method based on spatiotemporal-angle feature fusion 1. Positioning Model Design Previous research on LosS-VLP scenarios has demonstrated that temporal features are a core factor affecting positioning accuracy in deep learning-based indoor VLP methods. Therefore, this embodiment still uses the temporal feature module as the backbone for network design. Furthermore, the complementary relationship between spatial, temporal, and angular features is analyzed: spatial features reflect geometric positional relationships through signal strength distribution; temporal features capture motion continuity to suppress noise; and angular features directly model the impact of device rotation on channel gain. These three features work synergistically to improve positioning robustness in complex scenarios.

[0043] On the other hand, in any three-dimensional space, if the distances from the point to be measured to three coplanar known points are known, then the position of the point to be measured usually has two possible solutions that are symmetric about that plane. However, as... Figure 2 As shown, in this embodiment, since the plane is a spatial boundary (indoor ceiling), the point to be measured can only be located on one side of the plane, thus ensuring the uniqueness of the solution. Traditional RSS positioning methods primarily rely on algorithms to analyze the received signal strength and estimate the distance from the point to be measured to a known transmitting device. This means that the spatial characteristics between the transmitting and receiving devices can be indirectly represented as the position information of the receiving device in space, serving as a shallow feature for the positioning task. Since the propagation path of a direct signal is equal to the geometric straight-line distance between the transmitting and receiving ends, the propagation path of a multipath reflected signal must be greater than this straight-line distance. Therefore, accurate distance estimation can distinguish between direct signals and multipath reflected signals, improving robustness. Furthermore, using distance as a priori feature decouples the feature differences caused by distance, multipath, and position, unifying the signal distribution patterns across different distance intervals. This reduces the distribution offset of input data (such as scale variations in different scenarios), making it easier for the model to learn stable mapping relationships.

[0044] Therefore, distance estimation (such as...) is considered in the model design of this embodiment. Figure 5The distance labels (generated from coordinate labels) are used to improve positioning accuracy. Specifically, in the indoor positioning scenario of this embodiment, the positions of the LEDs are fixed and known, and the coordinates of the target to be located are also known during the training phase. Therefore, the distance from each LED to the target to be located is known. This is an implicit attribute, so the calculated distance is also included as a label in the model training task to enhance the model's ability to perceive spatial location features.

[0045] In addition, in the indoor positioning scenario described in this embodiment, by analyzing the system model and taking the dataset designed above as an example, the neural network learns the received data from different angles at the same location, effectively perceiving that various signals in the received signal follow a specific spatial distribution. In this embodiment, coordinates, transmission and reception distance, and device angle are used together as training labels, allowing the model to learn multi-dimensional spatial features and effectively improving the model's adaptability to complex indoor scenarios.

[0046] Based on the natural implicit coupling relationship between coordinates, distance, and angle, and taking into full account the complementarity between spatiotemporal angular features, this embodiment proposes a cascaded positioning framework of "encoding-fusion-decoupling" to improve the model's anti-interference ability and achieve high-precision positioning by utilizing auxiliary tasks.

[0047] like Figure 5 As shown, the core of the model consists of a feature encoder and a multi-task decoupling head. Firstly, it uses a kernel with a step size of... The convolutional blocks perform preliminary feature extraction on the input data, and then pass through a feature fusion module consisting of Bi-LSTM and 4head-Attention modules. The Bi-LSTM is specifically designed to capture temporal dependencies, the attention mechanism enhances spatial awareness, and angular features are extracted through a dedicated branch.

[0048] The extracted shared features are then input into the multi-task classification layer. Since coordinates and distance are highly correlated tasks, while angles are lowly correlated, different deep feature extraction modules are used in the multi-task classification layer to process the two types of tasks separately. Finally, the predicted coordinates, distances, and angles are output separately through linear layers based on a Kolmogorov-Arnold network (KAN) with complex mapping capabilities. Different spline functions and parameters are configured for different classification tasks using KAN.

[0049] Among them, the convolutional residual module undertakes spatial tasks, outputting coordinates and distances. The task features are intuitive and the modeling difficulty is relatively low, belonging to shallow tasks.

[0050] The angle inference module is a dedicated encoding branch for the model's angle features. It directly extracts signal features related to device rotation through convolutional layers, ensuring that angle information is fully modeled. Angle features are highly sensitive to relative relationships and subtle changes, requiring a deeper level of feature abstraction, making it a deep task. This design enables the model to simultaneously utilize spatial positioning references, temporal smoothing effects, and angle correction capabilities, achieving more robust 3D positioning performance.

[0051] In this embodiment, the above model is named STAK-PLoc (Spatio-Temporal-Angular KAN Progressive Localization).

[0052] In the picture Indicates the number of pilot signals extracted. M represents the number of transmit antennas, and M represents the number of subcarriers. These are the distance, coordinates, and angle vectors output by the network, respectively.

[0053] 2. Loss Function and Model Fine-Tuning Mechanism In this embodiment, there are two direct labels (coordinates and angle) and one indirect label (distance), such as Figure 5 As shown, before training begins, indirect labels must first be extracted, that is, a distance label set is constructed by generating distance labels from coordinate labels.

[0054] The model employs a progressive sharing strategy, drawing inspiration from transfer learning, to achieve efficient training through two stages: basic feature learning and fine-tuning. First, the network learns to establish basic feature representations, enabling it to perceive the influence of angular features on spatial features. Then, the general feature layer is frozen, and only high-level task-specific features are adjusted to prevent feature "forgetting."

[0055] To ensure consistent sensitivity of the model to training loss, this embodiment uses the same loss function for training the three tasks. In this embodiment, the loss function still uses mean squared error (MSE), which is very sensitive to outliers. At the same time, this embodiment notes that since the error of angle differs from the error of position and distance by magnitude, and angle features are only meaningful for multipath environment data, the main data affecting positioning accuracy comes from the Loss Channel. As can be seen from the above analysis, multipath data is negligible in the received signal. Therefore, to ensure the robustness of the positioning model, the angle task loss needs to be scaled as shown in formula (8) when designing the model loss.

[0056] .

[0057] in, For the loss of coordinate tasks, For the loss of distance from the mission, For the loss of angle task, This is the scaling factor, a constant that takes the value of 0 or 1.

[0058] During model training, as the training loss decreases, the impact of the angle loss on the overall loss will continuously increase. To prevent overfitting or gradient explosion, when the loss stops decreasing or even shows multiple consecutive increases, the shared layers of the model are frozen. =0, meaning that only coordinate loss and distance loss are used to fine-tune the classification task head. This model uses a cosine annealing learning rate scheduler. The detailed training process is shown in Table 1 below.

[0059] The two-dimensional tensor is obtained by processing the time-domain expression Y of the aforementioned received signal model. Specifically, the process involves: acquiring a time-domain received signal from an indoor environment, including line-of-sight propagation components, non-line-of-sight multipath propagation components, and noise interference; separating and extracting a pre-set complex pilot signal from the transmitter from the time-domain received signal; performing complex decomposition processing on the complex pilot signal to obtain the real and imaginary feature matrices corresponding to the complex pilot signal; and performing dimensional reconstruction processing on the real and imaginary feature matrices respectively to convert the one-dimensional sequence feature data into a two-dimensional tensor.

[0060] Table 1. Training process;

[0061] (III) Enhanced localization model based on structured state-space model like Figure 5 As shown, in the feature fusion backbone network of the STAK-PLoc model, bidirectional LSTM (Bi-LSTM) is used to capture the local bidirectional dependencies of the signal in the time domain. Simultaneously, the reconstructed network input data is treated as a dual-channel image and fed into the Multi-Head Attention module for processing. This module can extract similar features on each subcarrier, achieving global data interaction and thus enhancing the model's ability to perceive spatial and angular features.

[0062] While the model demonstrates good performance on small datasets, further improvements are needed to enhance its ability to handle long-sequence data and capture broader global dependencies within sequences, in order to build a more efficient and robust localization model. To this end, a long-sequence feature mining module based on Mamba-2 is introduced after the feature fusion module. This module dynamically adjusts its state mechanism based on the input, achieving a function similar to "dynamic focusing," maintaining high computational efficiency while avoiding the secondary computational complexity of traditional attention mechanisms (such as Transformer). The integration method of this module is as follows: Figure 6 As shown, the Mamba Block is implemented using a normalized residual network structure based on Mamba-2.

[0063] The core motivation for choosing Mamba-2 lies in its linear computational complexity, hardware-efficient design, and selective state mechanism, which dynamically focuses on key signal features in the input. Unlike the Transformer, which has quadratic complexity, Mamba-2's computational cost increases only linearly with sequence length (O(L)), making it particularly suitable for processing long-sequence RSS data involved in visible light localization (VLP). Furthermore, Mamba-2 has been mathematically proven to be equivalent to a structured attention mechanism, combining strong representational capabilities with the efficiency of a state-space model (SSM).

[0064] Compared to the Mamba model, Mamba-2 further enhances the parallelism of the training and inference processes. This feature is particularly suitable for localization tasks that require simultaneous processing of local details and global context: lower-level states can capture high-frequency details (such as precise location information), while higher-level states integrate global context (such as environmental topology), thereby achieving a natural fusion of multi-scale features.

[0065] III. Experimental Results and Analysis Multiple experiments were conducted to investigate the impact of the proposed dataset structure on the model's localization performance and to examine how the progressive spatiotemporal-angle localization network design strategy based on multi-task learning improves the reliability and robustness of indoor visible light localization. The experimental results were analyzed from multiple aspects, including model training efficiency, localization accuracy, and average error. Specific details are as follows. Unless otherwise specified, the basic experimental parameters in this embodiment are shown in Table 2.

[0066] Table 2. Basic parameters for experimental setup;

[0067] (a) Validation of the model training dataset In this embodiment, we use the two best-performing localization models LSTM-KAN and Attention-MLP proposed in previous research on LoS-VLP systems to verify the impact of adding a dataset with multiple signal-to-noise ratios on the accuracy of VLP systems containing NLoS signals, and the impact of adding angle estimation on the accuracy of receivers with spatial degrees of freedom in localization.

[0068] This embodiment trains LSTM-KAN and Attention-MLP using datasets with high signal-to-noise ratio (SNR), mixed SNR, and mixed SNR with angle labels, respectively. It is clear that... Figure 7 It was observed that the mixed signal-to-noise ratio dataset with angle labels was the most efficient and stable for training in both models, demonstrating that the angle labels and mixed signal-to-noise ratio conditions can enhance the model's ability to perceive the deep features of the system described in this embodiment, especially significantly improving the performance of the Attention-MLP model.

[0069] The trained model was validated using the same 500 test data points, and the accuracy performance of the trained model was evaluated using the experimental results of cumulative distribution and average precision. The experimental results are as follows: Figure 8 and 9 As shown.

[0070] In the two experiments, this embodiment was verified using test data with high, low and mixed signal-to-noise ratio conditions. First, it can be seen that as the signal-to-noise ratio increases, the positioning accuracy of the six models in both sets of experiments is enhanced, proving that the positioning model trained by constructing the dataset is effective.

[0071] Furthermore, consistent with the model training loss results, the mixed SNR dataset with angle labels exhibits higher localization accuracy during model training, followed by the model trained with only the mixed SNR dataset. This embodiment then analyzes the impact of mixed SNR on localization accuracy. In fact, in this embodiment, the amount of data for each SNR is random and non-uniform during mixed SNR training, and the SNR varies at different points within the same sampling. Therefore, this experiment still demonstrates strong robustness when mapped to real-world scenarios, proving its fairness. The experimental results also show that the accuracy of the model trained with a single SNR is significantly lower than that trained with mixed SNR, demonstrating that mixed accuracy training can enhance the robustness of the localization model when the data volume meets the training conditions, and the accuracy improvement is more pronounced under low SNR conditions.

[0072] Then, by combining the results of the two experiments, it was found that the Attention-MLP method, which was originally unusable in NLOS scenarios with receiver degrees of freedom due to huge errors, has accuracy that is very close to that of the LSTM-KAN method, and is almost identical under high signal-to-noise ratio conditions. Therefore, it can be concluded that in scenarios with receiver degrees of freedom, training the model by adding angle labels can effectively improve the model's localization accuracy again, and it can be asserted that if the dataset size is appropriately expanded, the localization accuracy will be improved even more.

[0073] Furthermore, to demonstrate the high reliability of STAK-PLoc in the scenario described in this embodiment, a set of comparative experiments was designed. LSTM-KAN, Attention-MLP, Bert-VLP, and an improved model with CRNN as its backbone, all performing well in similar scenarios, were used as comparative experimental members. The dataset with embedded mixed signal-to-noise ratio and angle labels, as described in the previous section, was used to train each model under the same epoch conditions. Experimental results are as follows: Figure 10 As shown.

[0074] Figure 10 Subfigures (a), (b), and (c) show the localization accuracy performance of the five models on three test datasets with 10dB, 40dB, and mixed signal-to-noise ratios, respectively. It is evident that the STAK-PLoc model proposed in this embodiment consistently exhibits the best CDF curve. To better analyze the advantages of the STAK-PLoc model on this dataset, box plots were further designed for comparison. Subfigures (e), (d), and (f) show that the STAK-PLoc model has a lower average error compared to the other three models, and the estimated error range is more concentrated, proving that this model not only has better reliability but also higher stability.

[0075] In addition, this embodiment also found that the STAK-PLoc model exhibits the largest relative error reduction rate (RERR) as the signal-to-noise ratio (SNR) increases. This indicates that the improvement in model performance with increasing SNR is more significant, meaning the model's advantages are more pronounced in high SNR environments. The RERR calculation method is shown in Equation 10. Here, L2 represents the average loss under high SNR, and L1 represents the average loss under low SNR. Similarly, sensitivity, i.e., inference response speed, is also a comprehensive indicator of algorithm performance. Therefore, the four algorithms were tested under the same hardware conditions with the same 500 samples to be predicted. The results of the inference speed are shown in Table 3. It is known that the response speed of several algorithms is in the millisecond range. However, the STAK-PLoc model proposed in this embodiment reduces the inference speed by 1ms compared to LSTM-KAN, which has a larger accuracy gap, while improving the inference speed by nearly 4ms compared to CRNN, which has relatively close accuracy. The BERT model is a variant model based on the Transformer architecture, which inherits the high computational overhead of Transformer. Furthermore, in the regression task of this embodiment, its serial dependency means that each step of the calculation must wait for the previous step to complete, making parallelism impossible. This results in an inference speed that is much lower than the other models. Overall, the method proposed in this embodiment has high reliability.

[0076] Table 3. Comparison of the sensitivity of the five models;

[0077] On the other hand, to further verify the impact of distance loss, angle loss, and fine-tuning mechanisms on positioning accuracy, a set of comparative verification experiments was added to verify whether freezing the pre-trained model, adjusting the loss function, and then fine-tuning the multi-task head can indeed enhance the model's reliability. The experimental results are as follows: Figure 11 As shown, the model trained with both distance and angle losses has lower accuracy than the model trained with only distance loss, proving that angle error affects the distance gradient, causing the model to converge to a suboptimal solution. However, the STAK-Ploc model, obtained by first training the complete model (including angle) and then fine-tuning it (using only coordinate and distance losses), shows a significant improvement in localization accuracy compared to the other two models. This demonstrates the implicit regularization effect of angle information; the model is forced to learn more robust feature representations to fit both distance and angle simultaneously, thus avoiding the local optima that might occur when only optimizing the distance loss. It is proven that angle information enhances the model's potential through feature augmentation during the pre-training stage, and this potential is released through target decoupling during the fine-tuning stage, thereby enhancing the model's localization reliability.

[0078] (II) Ablation Experiment Based on Mamba-block Localization Enhancement Model In this embodiment, to verify the impact of the Mamba block on model performance, a set of ablation experiments were designed. In these experiments, two models were trained for 300 epochs using datasets of sizes 100x8x3 and 300x8x3, respectively, with training batch sizes of 10 and 30. The same three datasets were then used for testing. The experimental results are as follows: Figure 12 As shown.

[0079] Under different signal-to-noise ratio conditions, the localization enhancement model based on Mamba-block showed significant improvement in localization performance on the 30x8x3 dataset trained for 300 epochs, while its localization performance stability was relatively poor on the 100x8x3 dataset trained for 300 epochs. Therefore, the ablation experiment was further analyzed using Tables 4 and 5.

[0080] Table 4 Comparison of the average accuracy of STAK-PLOC and MAMBA enhancement models trained on 100 sets of signals for 300 epochs under different signal-to-noise ratios;

[0081] Table 5 compares the average accuracy of the STAK-PLOC and MAMBA enhancement models trained on 300 sets of signals over 300 epochs at different signal-to-noise ratios.

[0082] Firstly, with the increase in training data volume and the number of training epochs, the Mamba-based reinforcement model can achieve millimeter-level positioning accuracy overall. Secondly, as shown in Table 4, when SNR=10dB, the reliability of the Mamba reinforcement model is not entirely higher than that of the STAK-Ploc model, indicating its lower stability. However, with the increase in training data volume and the number of training epochs, its reliability and stability are greatly improved, fully verifying that the Mamba model performs better on datasets of a certain scale. Furthermore, in Table 4, when the SNR increases from 10dB to 40dB, the accuracy improvements for the two models are 36.2% and 34.9%, respectively. In Table 5, when the SNR increases from 10dB to 40dB, the accuracy improvements for the two models are 36.9% and 34%, respectively, demonstrating that the Mamba reinforcement model is more sensitive to changes in SNR. This indicates that the Mamba module has a strong ability to perceive noise features in the data and achieves the model's noise reduction function through dynamic mining of sequence features in the data. In conclusion, given sufficient data, a model based on Mamba enhancement is a better choice.

[0083] In the localization task, the accuracy is the Euclidean distance between the actual value and the estimated value, which shows the worst-case performance and reliability of the model. Stability is also an important indicator for evaluating model efficiency. Therefore, the mean absolute error (MAE) of the four models obtained under different data volumes and different training epochs was calculated. The calculation results are shown in Table 6. In addition, the number of parameters and prediction time of the four models are also listed in the table.

[0084] Table 6 Calculation results;

[0085] The table shows that, under the same training time, the Mamba reinforcement model exhibits high stability at different signal-to-noise ratios. However, it has more parameters than the baseline model. In the hardware environment of this experiment, the prediction time for 500 signals differed by more than three times between the two models. But considering scenarios with limited hardware capabilities such as computing power, model sensitivity is a crucial factor that needs to be considered more carefully. Furthermore, in... Figure 13The training loss curves of four models are listed. It was found that the Mamba reinforcement model experiences a faster decrease in validation loss, followed by a plateau. However, the model trained on 100 data points has the lowest validation loss, but also the lowest model accuracy. This demonstrates that the baseline model is more prone to getting trapped in local optima under small sample conditions. Mamba's structural design (selective state space) is indeed better at capturing long-range correlations and core dependencies in the data, while also exhibiting greater sensitivity to system noise. In conclusion, when real-time positioning requirements are high, the STAK-Ploc model should be prioritized; while when positioning accuracy and stability are more critical, the Mamba reinforcement model should be used.

[0086] This embodiment addresses the shortcomings of existing indoor visible light positioning methods in handling non-line-of-sight (LOS) signal transmission and indoor VLP scenarios where the target terminal is not parallel to the light source's emission plane. Aiming to construct a highly reliable and robust deep learning-based indoor VLP model, it proposes a method to enhance the stability of the positioning model by using RSS signals with mixed signal-to-noise ratios as the model training dataset. Simultaneously, a STAK-Ploc model based on spatiotemporal-angle feature fusion is designed and coupled with a Mamba module to further enhance its positioning accuracy, achieving highly reliable centimeter-level positioning. Experimental results show that the proposed method outperforms current advanced VLP models. This study concludes that there is still significant room for improvement in the model's positioning accuracy.

[0087] Example 2 This embodiment provides a visible light positioning system based on spatiotemporal-angular feature fusion, including: The data acquisition module is configured to build a training dataset based on the indoor positioning scenario, including visible light positioning pilot maps, coordinate labels, distance labels, and angle labels; The model training module is configured to input the training dataset into the localization model for model training; wherein, preliminary features are extracted through the feature embedding module, and then the temporal dependency of the preliminary features is captured and spatial awareness is enhanced through the feature fusion module to obtain fused features; the fused features are input into the position extraction branch and angle extraction branch at different depths to obtain position information and angle information respectively; the loss function is calculated and the parameters are updated until the loss function converges to obtain the trained localization model; The positioning module is configured to input the visible light positioning pilot map of the receiver to be positioned into the trained positioning model to obtain positioning information.

[0088] Example 3 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the visible light positioning method based on spatiotemporal-angle feature fusion as described in Embodiment 1 above.

[0089] Example 4 This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the visible light positioning method based on spatiotemporal-angle feature fusion as described in Embodiment 1 above.

[0090] The steps or modules involved in Embodiments 2 to 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0091] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A visible light positioning method based on spatiotemporal-angular feature fusion, characterized in that, include: A training dataset was constructed based on indoor positioning scenarios, including visible light positioning pilot maps, coordinate labels, distance labels, and angle labels; The training dataset is input into the localization model for model training. Preliminary features are extracted through the feature embedding module, and then the temporal dependence of the preliminary features is captured and spatial awareness is enhanced through the feature fusion module to obtain fused features. The fused features are input into the position extraction branch and angle extraction branch at different depths to obtain position information and angle information respectively. The loss function is calculated and the parameters are updated until the loss function converges to obtain the trained localization model. The visible light positioning pilot diagram of the receiver to be located is input into the trained positioning model to obtain positioning information.

2. The visible light positioning method based on spatiotemporal-angular feature fusion as described in claim 1, characterized in that, The indoor positioning scenario is a spatial environment with length, width, and height, with multiple LEDs with communication functions installed on the ceiling as transmitters; The device to be located, equipped with a signal receiving array composed of photodiodes, moves randomly in space, serving as the receiving end.

3. The visible light positioning method based on spatiotemporal-angular feature fusion as described in claim 1, characterized in that, The process of obtaining the visible light positioning pilot map is as follows: The receiver acquires time-domain received signals from an indoor space environment, including line-of-sight propagation components, non-line-of-sight multipath propagation components, and noise interference. The complex pilot signal preset by the transmitter is separated and extracted from the received signal in the time domain; The complex pilot signal is subjected to complex decomposition processing to obtain the real part feature matrix and the imaginary part feature matrix corresponding to the complex pilot signal; The real and imaginary feature matrices are reconstructed to convert the one-dimensional sequence feature data into a two-dimensional tensor, which serves as a visible light positioning pilot map.

4. The visible light positioning method based on spatiotemporal-angular feature fusion as described in claim 1, characterized in that, The coordinate labels represent the three-dimensional coordinates of the receiver in an indoor positioning scenario; The distance tag represents the straight-line distance between the transmitter and receiver in an indoor positioning scenario. The angle label represents the rotation angle of the receiving end.

5. The visible light positioning method based on spatiotemporal-angular feature fusion as described in claim 1, characterized in that, The loss function includes a coordinate loss function, a distance loss function, and an angle loss function; the angle loss function is adjusted by a scaling factor.

6. The visible light positioning method based on spatiotemporal-angular feature fusion as described in claim 5, characterized in that, The training process of the localization model includes two stages: when the loss function no longer decreases and shows multiple consecutive increases, the shared layer of the model is frozen, the angle loss function is discarded, and only the coordinate loss function and the distance loss function are retained, and the classification task is fine-tuned.

7. The visible light positioning method based on spatiotemporal-angular feature fusion as described in claim 1, characterized in that, Also includes: When the core requirement for positioning is real-time performance, the aforementioned positioning model is adopted; When the core requirements for positioning are accuracy and stability, a Mamba-2 module is added after the feature fusion module of the positioning model.

8. A visible light positioning system based on spatiotemporal-angular feature fusion, characterized in that, include: The data acquisition module is configured to build a training dataset based on the indoor positioning scenario, including visible light positioning pilot maps, coordinate labels, distance labels, and angle labels; The model training module is configured to input the training dataset into the localization model for model training; wherein, preliminary features are extracted through the feature embedding module, and then the temporal dependency of the preliminary features is captured and spatial awareness is enhanced through the feature fusion module to obtain fused features; the fused features are input into the position extraction branch and angle extraction branch at different depths to obtain position information and angle information respectively; the loss function is calculated and the parameters are updated until the loss function converges to obtain the trained localization model; The positioning module is configured to input the visible light positioning pilot map of the receiver to be positioned into the trained positioning model to obtain positioning information.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the visible light positioning method based on spatiotemporal-angle feature fusion as described in any one of claims 1-7.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the visible light positioning method based on spatiotemporal-angle feature fusion as described in any one of claims 1-7.

Images